Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

How Machine Learning is Powering Neuroimaging to Improve Brain Health

How Machine Learning is Powering Neuroimaging to Improve Brain Health This report presents an overview of how machine learning is rapidly advancing clinical translational imaging in ways that will aid in the early detection, prediction, and treatment of diseases that threaten brain health. Towards this goal, we aresharing the information presented at a symposium, “Neuroimaging Indicators of Brain Structure and Function - Closing the Gap Between Research and Clinical Application”, co-hosted by the McCance Center for Brain Health at Mass General Hospital and the MIT HST Neuroimaging Training Program on February 12, 2021. The symposium focused on the potential for machine learning approaches, applied to increasingly large-scale neuroimaging datasets, to transform healthcare delivery and change the trajec- tory of brain health by addressing brain care earlier in the lifespan. While not exhaustive, this overview uniquely addresses many of the technical challenges from image formation, to analysis and visualization, to synthesis and incorporation into the clinical workflow. Some of the ethical challenges inherent to this work are also explored, as are some of the regulatory requirements for implementation. We seek to educate, motivate, and inspire graduate students, postdoctoral fellows, and early career investigators to contribute to a future where neuroimaging meaningfully contributes to the maintenance of brain health. Keywords Machine learning · Deep learning · Clinical translational neuroimaging · Brain health · MRI · PET · EEG · Transcranial magnetic stimulation * Randy L. Gollub Computer Science and Artificial Intelligence Laboratory, rgollub@partners.org Massachusetts Institute of Technology, Cambridge, MA 02139, USA Department of Psychiatry and Martinos Center Institute of Systems Neuroscience, Medical Faculty, Heinrich for Biomedical Imaging, Department of Radiology, Heine University Düsseldorf, Düsseldorf, Germany Massachusetts General Hospital, Boston, MA 02114, USA Institute of Neuroscience and Medicine, Brain & Behaviour Harvard-MIT Health Sciences and Technology, (INM-7) Research Centre Jülich, Jülich, Germany Massachusetts Institute of Technology, Cambridge, MA 02139, USA Department of Radiology, Brigham and Women’s Hospital and Harvard Medical School, MA 02115 Boston, USA University of Massachusetts Boston, Boston, MA 02125, USA Department of Psychology, Northeastern University, Boston 02115, USA Department of Radiology, Boston Children’s Hospital, Harvard Medical School, Boston, MA 02115, USA Martinos, Radiology, MGH, MIT, HMS & EECS, Cambridge 02114, USA Department of Neurology and McCance Center for Brain Health / Harvard Medical School, Massachusetts General Department of Psychiatry, Brigham and Women’s Hospital Hospital, Boston 02114, USA and Harvard Medical School, Boston 02115, USA 6 15 Centre for Medical Image Computing, University College Center for Brain Circuit Therapeutics, Department London, London, UK of Neurology, Psychiatry, and Radiology, Brigham and Women’s Hospital and Harvard Medical School, Martinos Center for Biomedical Imaging, Department 02115 Boston, USA of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston 02114, USA Vol.:(0123456789) 1 3 944 Neuroinformatics (2022) 20:943–964 set of important questions that we believe researchers should Introduction bear in mind when working in this area. Machine learning models vary in the amount of domain Machine learning is contributing to rapid advances in clinical knowledge they incorporate and how they do so. Some mod- translational imaging to enable early detection, prediction, els explicitly enforce that their outputs are consistent with and treatment of diseases that threaten brain health. Brain the physics of an imaging or measurement process. Other diseases, including cerebrovascular disease, depression, methods act upon features explicitly chosen because they migraine headaches, and dementia, are leading causes of are known to be relevant for the task at hand. These feature- global disability (Vos et al., 2020). Continued progress in based methods are often applied to established clinical use neuroimaging and machine learning, and the collection cases to reduce time, manual labor, and/or person-to-person of increasingly large-scale data sets, promise to transform variation (Gajawelli et al., 2019). "End-to-end" approaches healthcare by providing non-invasive, reliable indicators of abstract out explicit feature definition to go from raw data brain health, resilience, and vulnerability long before clinical all the way to interpretable quantitative metrics, for exam- manifestations of disease. But many technical challenges ple, of brain health. All traditional step-by-step processes, remain. On February 12th 2021, the MGH McCance Center such as artifact removal, registration, conversion between for Brain Health at Mass General Hospital, together with the temporal and spectral domains, and feature extraction could Harvard-MIT Health Sciences and Technology Neuroim- be encompassed in a single machine learning pipeline. aging Training Program, co-hosted a virtual symposium, In general, models with more explicitly encoded domain “Neuroimaging Indicators of Brain Structure and Function— knowledge are less flexible in adapting to cases where the Closing the Gap Between Research and Clinical Application,” measurement process may be inaccurately characterized. to highlight some of these remaining challenges and machine That said, these methods are able to incorporate known learning approaches to overcome them. Recorded videos of the relationships, which can guide the learning process and symposium presentations and discussions are available link prevent nonsensical results. The methods covered in this to YouTube videos https:// www. youtu be. com/ playl ist? list= report represent examples from each of these categories as PL0A- NKHLV rNF82 vdjey yaBRo iXg77 lCeW. well as intermediate cases which explicitly incorporate some In this symposium report, we explore a spectrum of domain knowledge but also allow the model significant flex- machine learning applications in neuroimaging and use sym- ibility in producing unconstrained final outputs. posium presentations to illustrate key points. We cover both It is important to acknowledge that the field is dynamic recent advances and outstanding challenges, beginning with with each area undergoing rapid transitions. In some cases, image acquisition, and ending with computation of quantita- machine learning techniques are already deployed or in tive metrics and initial clinical utilization (Fig. 1): advanced stages of testing for deployment into existing clinical workflows. In other cases, efforts are focused on Section I describes how machine learning improves volu- early research, with the goal of discovering scientific insight metric image acquisition and reconstruction. or extracting meaningful features from the images that can Section II  describes machine learning approaches to be fed into machine learning methods to generate clinically image processing, focusing on image harmonization, and meaningful biomarkers (FDA-NIH Biomarker Working methods to detect deviations from healthy brain structure Group, 2016; Mateos-Pérez et al., 2018). In all presenta- and function. tions and throughout this report, our goal is to educate, Section III describes machine learning advances in inter- motivate and inspire graduate students, post-doctoral fel- pretation and analysis of non-volumetric EEG data. lows, and early career investigators to contribute to a future Section IV describes multiple approaches to assess brain where imaging meaningfully contributes to the maintenance health using deviations from healthy aging. of brain health. Section V describes how machine learning techniques can be applied to individual patient imaging, and other diagnostics, to personalize medical treatments that Section I: Machine Learning for Improved improve brain health. Volumetric Image Acquisition and Reconstruction Finally, Section VI explores the implications of deploy- ing neuroimaging indicators of brain health into a clinical Machine learning techniques can be used to improve one of workflow. In particular, we focus on regulatory approval the earliest steps in the neuroimaging pipeline even before pathways of machine learning algorithms and the ethical an image is viewed by a clinician or researcher: image considerations involved in collecting, algorithmically ana- formation. Typically, scanner-acquired measurements lyzing, and acting upon the derived information. We raise a 1 3 Neuroinformatics (2022) 20:943–964 945 Fig. 1 Schematic illustration of the spectrum of machine learn- summaries to metrics used to quantify brain health, such as the vol- ing applications in clinical translational neuroimaging. A typical ume of a structure or the estimated age of a subject. Finally, these volumetric neuroimaging workflow is shown for MRI. A patient quantitative metrics, once comprehensively validated, can be used is scanned, creating a signal (i.e. k-space data) which is converted to inform patient care through early detection of subtle abnormali- to an image via a reconstruction algorithm in preparation for clini- ties and to guide treatments such as targeted brain stimulation. These cal review by a radiologist. In some cases, the reconstructed image steps are not just useful at the individual patient level but can also undergoes further computational processing to produce higher-level drive population level analyses that can lead to insights regarding summaries, such as segmentations or registrations to an atlas. Option- healthy and disordered brain structure and function ally, in the future, further computational processing can convert these represent an encoding of the patient anatomy under the Once all scanner data is acquired, machine learning is physics governing the imaging system. For example, the useful in reconstructing the image of interest itself. Model- measurements acquired from an MRI scanner represent based optimization techniques for MRI, CT, and PET imaging the Fourier transform of the image of interest (Nishimura, have typically provided iteratively refined solutions to under- 2010), while the measurements acquired from PET and CT determined inverse problems (Griswold et al., 2002; Lustig scanners represent the Radon transform of the image of et al., 2008; Pruessmann et al., 1999). Recently, deep-learning interest (Ramm & Katsevich, 1996). Recovering the under- methods quickly estimate solutions to the inverse imaging lying image from the acquired scanner data requires solv- problems, including network architectures that explicitly employ ing an inverse problem. Due to time, patient comfort and the physics of the imaging system (Hammernik et al., 2018; safety considerations, or monetary constraints, often only Putzky et al., 2019; Schlemper et al., 2018). Efforts to collect a limited number of scanner measurements are acquired, large-scale public datasets of raw imaging data have accelerated making this inverse problem highly under-determined. advances in image reconstruction by enabling rapid model Further, the acquired signals may be corrupted by imper- prototyping and by simplifying and standardizing evaluation fections in the imaging process, such as patient motion or of varying approaches. For example, the FastMRI dataset system noise. Techniques from machine learning, includ- provides publicly available k-space data for reconstructing over ing (1) model-based optimization methods (Griswold six thousand human brain MRIs (Zbontar et al., 2019). et al., 2002; Lustig et al., 2008; Pruessmann et al., 1999), Beyond accelerated reconstruction from limited measurements, (2) data-driven learning methods (Quan et al., 2018; Yang several machine learning approaches have been proposed et  al., 2018), and (3) combinations of these two strate- for correcting artifacts arising during image acquisition and gies (Hammernik et al., 2018; Schlemper et al., 2018), are reconstruction. For example, both optimization and learning- promising approaches for mitigating these image forma- based approaches have been proposed for MRI denoising tion issues, enabling faster, higher-quality image creation (Anand & Sahambi, 2010; Manjon & Coupe, 2019) and motion for downstream analysis. correction (Chun et al., 2012; Haskell et al., 2018; Pipe, 1999). One such application of machine learning involves deciding At this symposium, Ms. Nalini Singh presented two neural which exact scanner measurements to acquire. Given time or network layer structures which can be used to build networks financial imaging budget constraints, machine learning can be which correct each of these artifacts while also being used for used to identify which subset of measurements will be the most accelerated reconstruction (Singh et al., 2020). Unlike many other informative for reconstructing the final image. For example, reconstruction methods, these layers incorporate convolutions on several approaches have been proposed for learning the optimal both the frequency space and image space features. By operating k-space acquisition pattern for a specified class of MRI scans, in both spaces, these layers both correct artifacts native to the often identifying different patterns for different anatomies frequency space and manipulate image space representations (Bahadir et al., 2020; Wang et al., 2021; Weiss et al., 2021). to form coherent image structures. Figure 2 shows a detailed For neuroimaging applications in particular, sampling trajec- diagram of the layers, and Fig. 3 shows example reconstructions tories could be optimized for specific structures of interest for demonstrating the positive impact of this method on the quality the clinical question being asked. More recent work aims to of the reconstructed image in representing the true brain anatomy. optimize the acquisition pattern with even greater specificity Several deep-learning based approaches have also been for each individual patient (Zhang et al., 2019). proposed for metal artifact reduction in CT imaging of other 1 3 946 Neuroinformatics (2022) 20:943–964 Fig. 2 Two joint layer architectures combining frequency and image resents frequency space features at the nth  layer, and v represents space representations, embedded within full network architectures image space features at the nth  layer. At each layer, Batch Normali- for MRI reconstruction. Red squares represent frequency space quan- zation (BN), a convolution, and an activation function are applied to tities, while blue squares represent image space quantities. u rep-both u and v , summarized by `F-Conv' or `I-Conv', respectively n n n anatomies (Gjesteby et al., 2017; Hu et al., 2019); these tech- While machine learning-based approaches promise to niques could be extended to improve brain CT imaging for improve the speed, value, and quality of brain image acquisition, patients with deep brain stimulation (DBS) devices in situ. In several challenges must be solved before they are incorporated PET imaging, deep-learning based approaches have demon- into standard clinical worko fl ws. For example, many current strated improved correction of attenuation effects both with reconstruction methods require large datasets of thousands of (Ladefoged et al., 2018; Liu et al., 2018a) and without (Dong high-quality acquired signals from a particular imaging proto- et al., 2020; Liu et al., 2018b) concordant anatomical imaging, col. This requirement makes it difficult to apply these methods or to enable low-dose PET (Xu et al., 2017). Each of these inno- to new imaging protocols for which large datasets have not yet vations makes critical contributions to improving the safety, been collected. New techniques are being developed to adapt quality and/or value of clinically meaningful information about these learning-based approaches to either require fewer train- brain health which can be gleaned from the imaging study. ing examples or to transfer the information from previously 1 3 Neuroinformatics (2022) 20:943–964 947 Fig. 3 Example reconstructions from 4 × undersampled data (row 1), by comparing the final two columns. The Interleaved and Alternating zoomed-in image patches (row 2), difference patches between recon- architectures produce two slightly different reconstructions, both of structions and ground truth images (row 3), and frequency space which better eliminate blurring and 'ringing' artifacts, where multiple reconstructions (row 4) are shown here to visually communicate the copies of the image appear stamped on top of each other impact of this reconstruction approach. It is most easily appreciated collected datasets of one protocol to a new protocol of inter- can be incorporated in each workflow step, but similar prin- est (Han et al., 2018). And, for any reconstruction method, ciples extend to other volumetric imaging techniques such uncertainty quantification techniques will be needed to high- as PET and CT. light regions of reconstructions with a high likelihood of error (Edupuganti et al., 2021). These uncertainty quantifications will Quality Assurance and Harmonization enable radiologists to understand when more detail is needed to identify a particular feature of interest (Edupuganti et al., 2021), Expert human labellers typically perform image quality possibly requiring re-imaging of the patient. assessment (QA), but this process is labor-intensive and can suffer from low inter-rater reliability. Carefully designed machine learning techniques promise to enable fast, easily Section II: Machine Learning Applications accessible, consistent QA. Previously proposed approaches for Volumetric Image Processing use carefully curated quality metrics as input features to various types of classifiers which label images as usable or Once brain imaging data is collected and reconstructed, unusable (Esteban et al., 2017; Küstner et al., 2018; Pizarro there are several steps in the image analysis pipeline where et al., 2016). Crowdsourcing approaches have also been used machine learning can improve the extraction of meaningful, to improve the accuracy of these automatic QA tools. The quantitative features related to brain health. In this section use of many non-expert, human raters as inputs to a convo- we explore some of these advances, giving an overview of lutional neural net improves the accuracy of classification each step and examples of how machine learning models are over a single site data set (Keshavan et al., 2019). Further, used. We focus on MRI to demonstrate how ML techniques a web-based API acting as a quality metric repository has 1 3 948 Neuroinformatics (2022) 20:943–964 Fig. 4 Atlas construction (concept in panel A) can enable quantifica- construction versus a prospectively gathered longitudinal cohort) tion of brain development across ages (panel B- schematically indi- and can detect abnormalities as outliers to normal (panel C). ADC- cating the benefits of using a clinical cohort of individuals for atlas Apparent Diffusion Coefficient increased the volume of quality metric labeled, multi-site images for a machine learning prediction task to be invariant data available to be used to develop new, more generalizable to the scanners on which they were acquired (Dinsdale et al., QA tools (Esteban et al., 2019). 2021). In addition to ensuring the quality of individual scans, batch effects affecting images acquired at different locations Quantification of Brain Health and Detection or times must be eliminated to perform large-scale, multi- of Abnormality site studies. Machine learning approaches provide flexible methods to detect and remove the relevant site-specific Machine learning approaches can also be used to charac- effects. One approach is to directly convert data acquired terize healthy brain characteristics and identify deviations in one setting to the data that would have been acquired from the norm. During the symposium, Dr. Yangming Ou in a different setting. During the symposium, Dr. Cetin- described how to construct ‘normal’ atlases using group- Karayumak presented such a retrospective harmonization wise unbiased image registration. Brain MRI atlases sum- technique which represents diffusion MRI (dMRI) data as a marize healthy brain anatomy and typical signal intensity combination of spherical harmonic basis functions. Rotation- profiles at the voxel-, regional-, fiber-, and whole-brain invariant features are derived at each voxel from the computed levels (Guimond et al., 2000) (Fig. 4A). Brain atlases con- basis function coefficients for each image, and a mapping structed from imaging data can be used in multiple ways is computed between the features of target and reference to quantify brain health. One example is the quantification scanners in order to harmonize them (Cetin Karayumak of normal childhood development (Ou et al., 2017; Sotardi et al., 2019). A different approach is to learn intermediate et al., 2021). A series of constructed atlases from cohorts representations invariant to the scanner on which any image of healthy subjects clustered by age can enable longitu- was acquired. These intermediate representations can then dinal quantification of brain development from data sets be used to reconstruct images without site-specific effects where every subject was scanned once (Fig.  4B). This is (Moyer et al., 2020). Alternatively, instead of removing or not only cost effective when constructed from clinically transforming site-specific effects at the image level, a third acquired brain scans, but also has the potential to incorpo- strategy is to encourage downstream features derived from the rate a more comprehensive range of healthy variation than 1 3 Neuroinformatics (2022) 20:943–964 949 data acquired in a single or set of pooled research studies. archives by using deep learning to transform lower- Another use of quantitative brain atlases is to detect subtle quality images into higher-quality ones, thus enabling abnormalities due to a wide range of disorders (Pinto et al., use of advanced image segmentation tools. In particular, 2018) (Fig. 4C). Atlas-quantified voxel-wise deviation val- several such tools are built for MP-RAGE scans, which ues can be used as features in classical machine classifiers are popular due to their SNR efficiency and contrast. (O’Muircheartaigh et al., 2020) or deep convolutional neural At this symposium, Dr. Juan Eugenio Iglesias presented networks (Baur et al., 2021) to further improve the accuracy an approach to synthesize isotropic 1  mm MP-RAGE and generality of atlas-based detection of deviations from volumes from low-resolution scans of arbitrary contrast, brain health. This strategy has been used for structural MRI enabling their segmentation and analysis with standard (Baur et al., 2021; O’Muircheartaigh et al., 2020) and diffu- neuroimaging tools (Iglesias et al., 2020). An example sion MRI (Pinto et al., 2018). is shown in Fig.  5, where a 5  mm axial FLAIR scan is transformed into a 1 mm isotropic MP-RAGE scan, Segmentation and subsequently segmented with FreeSurfer, which requires 1  mm isotropic T1 data—and thus could not Automatic segmentation of brain images enables quantita- have processed the FLAIR scan directly, due to MR tive estimation of the volumes of brain structures that can contrast mismatch and insufficient resolution. lead to other indicators of brain health. These quantitative estimates enable population studies as well as longitudinal Visualization analysis within individual subjects. Most previous work on brain segmentation has focused on MRI, which provides Visualization frameworks can foster deeper understanding and detailed images with an ever increasing range of specifi- facilitate interpretation of high-dimensional clinical imaging cally tuned contrasts for visualizing different details of brain data. Further, targeted visualizations allow developers to anatomy and function (Akkus et al., 2017). design and optimize computational algorithms. State-of-the-art Freesurfer is one example of a widely used package visualization tools deal with challenges such as large amounts of for brain MRI analysis and includes a machine learning data such as in diffusion and functional MRI, and the inevitable approach to segment many brain structures (Fischl et al., variation of file formats across different institutions. Web-based 2002) as part of a larger image analysis pipeline. This tech- tools such as Fiberweb (Ledoux et al., 2017) or XTK (Haehn nique involves finding the maximum a posteriori estimate et al., 2014) have contributed to brain imaging visualizations of a segmentation of an anatomical brain region (e.g. hip- and 3D rendering of connectivity in recent years. Many other pocampus), given the image to be segmented and a linear visualization tools not limited to DTI data emerged recently, transform mapping it to an expertly curated segmentation such as Neurolines (Al-Awami et al., 2014) to visualize 3D brain atlas. This technique only employs approximately one hun- tissue in 2D, and comparative visualizations for fMRI brain dred labeled scans for a specific atlas, but the entire segmen- images (Jönsson et al., 2019). tation procedure takes several hours. There has therefore At the symposium, Ms. Loraine Franke presented her been recent interest in neural network-based segmentation work on developing web-based interactive visualization methods, which provide segmentations on the order of sec- tools for diffusion tractography imaging data (Franke & onds (Akkus et al., 2017; Despotović et al., 2015) to address Haehn, 2020; Franke et al., 2020). Her open-source tool, the requirement to perform expert level segmentation on FiberStars (Franke et al., 2020) (Fig. 6) enables researchers large scale image data sets. to create low-dimensional cluster representations of high In these approaches, a convolutional neural network typi- dimensional data, select, visualize, and compare multiple cally directly predicts human-labeled segmentations from clusters across multiple patients, and visualize individual patches or volumes and requires many labeled images to patient fiber tracts. By using different projection techniques train. Furthermore, these approaches are extremely sensitive for multidimensional scaling such as t-SNE (van der Maaten to shifts in input image intensity. To apply these methods to & Hinton, 2008), PivotMDS (Brandes & Pich, 2007) and scans of a different contrast or resolution, additional labels others, the FiberStars tool lets the user interactively explore must be collected and used to retrain or fine-tune the net- high dimensional data. For example, FiberStars enables works. Thus, recent research has focused on unsupervised users to answer research questions with comparative ensem- deep learning approaches for training brain segmentation ble visualizations, especially for evaluating and testing networks, (Dalca et al., 2019) or on adapting trained net- hypotheses, or to analyze factors combined with pathologi- works to new imaging analysis task scenarios (Kamnitsas cal findings. FiberStars addresses a large class of complex et al., 2016). visualization challenges for multidimensional data or data It is now possible to aggregate larger cohorts of composed of collections of patients. useful brain image data from clinical and/or research 1 3 950 Neuroinformatics (2022) 20:943–964 Fig. 5 Left column: coronal plane of an MP-RAGE scan (top, slice produced by Dr. Iglesias’s tool. Right column: automated segmenta- thickness: 1  mm)) and corresponding coronal plane of an axial tion of the original and synthetic MP-RAGE volumes produced by FLAIR scan (bottom, slice thickness: 5 mm) from the ADNI dataset FreeSurfer (Fischl et al., 2002) (adni-info.org). Middle column: synthetic 1  mm MP-RAGE volume Some of the most advanced work focuses on the diag- Section III: Machine Learning Advances nostic needs for patients with epilepsy, an area for which in Interpretation and Analysis EEG is already in active clinical use. In the current standard of Non‑volumetric EEG Data of care, making diagnoses and therapeutic decisions relies on painstaking manual annotation of many hours of EEG Another important, and emerging, area where machine recording by highly-trained expert epileptologists (Si, 2020). learning approaches are enhancing the understanding of brain Machine learning methods enable automatic detection of health is in clinical applications of electroencephalography, markers of epilepsy in interictal (non-seizure) data using or EEG data. EEG is multidimensional time series data, specific spectral, morphological, or network-based features. where multiple electrodes are placed on the scalp resulting While some feature-based approaches attempt to replicate in simultaneous channels of data being collected at a high the eye of the expert using features like those epileptologists time resolution. EEG is currently in clinical use for multiple, observe; other end-to-end deep learning and neural network specific applications, such as for diagnosing and monitoring models attempt to glean undiscovered signatures of epilepsy sleep disorders, epilepsy, disorders of consciousness, stroke, from the raw data itself. One example of this is classifying real-time electroconvulsive therapy (ECT) patient monitoring, routine EEGs into normal vs. abnormal, where abnormal and anesthesia (Roy et al., 2019a, b). EEG has the unique is, by definition, heterogeneous and context-dependent (van advantages of being non-invasive, relatively inexpensive, and Leeuwen et al., 2019). Machine learning based clinical deci- more adaptable to naturalistic or ambulatory settings compared sion support for epileptologists for diagnosis and localiza- to other imaging modalities. In some cases, even a few EEG tion of epileptic foci are highly promising as they reveal electrodes in a specific location can yield enough information interrelationships between brain regions and activity that for inference, without the need for EEG across all of the cortex. are difficult to discern by eye. Therefore, machine learning approaches can not only greatly In contrast to epilepsy, where EEG is already being used streamline existing clinical applications of EEG, but they can clinically, machine learning approaches are expanding the also open the door to new applications such as earlier, less potential for EEG-based diagnostic biomarkers for other expensive, or more accessible diagnostics (Michel & Murray, diseases, such as Alzheimer’s Disease (Escudero et al., 2006; 2012; Miranda et al., 2019). 1 3 Neuroinformatics (2022) 20:943–964 951 Fig. 6 Split screen showing 3D representations of fiber tract anatomy ent patients is displayed with additional two-dimensional representa- given by fibers of dMRI scans across different subjects. The menu tion radial plots at the bottom of each patient’s panel showing scalar bar at the left facilitates toggling on and off visualization of differ - values associated with each of the fiber tracts. For each anatomical ent subjects (top left), cluster (middle left, showing the Callosum tract, the 2D radial plots show mean and standard deviations of the Forceps Major), and coloring of the 3D tract by a selected scalar different scalars on each axis. Demographic information about each value (bottom left, showing a measurement of fractional anisotropy patient is shown above the 3D visualization, for example, age, gen- (FA2) from the DTI scan). High values of fractional anisotropy are der, height and weight. Each patient is anonymized by a number seen colored in red while lower values are colored in blue. Inter-hemi- in the labels next to the anatomical fiber tract name in purple. Other sphere crossing of the third patient shows no red colors and therefore relevant measurements for analysis are mean fiber length, number of no high fractional anisotropy values. Tractography from five differ - fibers or fiber similarity Gallego-Jutglà et al., 2015; Jelles et al., 1999; Lehmann et al., National Sleep Research Resource (Sleep Data—National 2007; Tzimourta et al., 2021; Woon et al., 2007). However, Sleep Research Resource, 2021  https:// sleep dat a. or g/), these methods are further from clinical deployment than those the PhysioNet Computing in Cardiology Challenge 2018 for epilepsy, mostly in feature discovery stages. Analogous (Ghassemi et al., 2018), and the TUH Abnormal EEG corpus strides are being made to discover novel, cost-ee ff ctive, and (Alhussein et  al., 2019; Gemein et  al., 2020; Roy et  al., ambulatory EEG-based biomarkers for diagnosing stroke, 2019a, b; Temple University EEG Corpus Downloads, 2021). schizophrenia, and attention deficit hyperactivity disorder (Ahmadlou & Adeli, 2011; Hosseini et  al., 2020; Phang et al., 2020; Sastra Kusuina Wijaya et al., 2015). While these Section IV: Brain Health as Assessed directions have great potential for impact if successful, since by Deviations from Healthy Aging they are new clinical applications of EEG, their success depends on connecting sound and robust machine learning Another machine learning approach to characterize brain health algorithm design to underlying physiology, which can prove is to summarize an image or biosignal into a single metric that elusive. Interpretability will likely also come into play, since reflects brain health, such as brain age estimation (Fig.  7d) clinicians must be convinced of the specific clinical utility of (Cole et al., 2019). The difference between estimated brain EEG for each new application. age and actual chronologic age, known variously as predicted A recurring theme in the development of machine age difference (PAD), Brain Age Index (BAI) or ΔBrainAGE, learning methods that is the same for EEG, as it is for any has identified accelerated aging in individuals with cognitive other imaging modality, is the availability of large, labeled impairment (Liem et al., 2017; Poddar et al., 2019), traumatic datasets. Three such sources for large EEG datasets are the brain injuries (Cole et al., 2015), schizophrenia (Cole et al., 1 3 952 Neuroinformatics (2022) 20:943–964 Fig. 7 Machine learning (ML) can estimate a patient’s brain age and at various ages; (c) cross validation to quantify the accuracy of the quantify abnormal (accelerated or delayed) aging. (a) training sam- ML model; and (d) when applied to target patients, the ML model ples consisting of normal brain MRIs from a large set of individu- can quantify deviations from normal brain aging als; (b) ML algorithm that learns how a normal brain MRI appears 2018), Alzheimer's disease (Bashyam et al., 2020), and diabetes Another symposium speaker, Dr. Haoqi Sun, presented (Franke et al., 2013). Deviations from expected brain age have his work on a feature-based machine learning model that also been reported for more subtle changes due to social and takes advantage of the fact that brain activity as recorded environmental influences, including a protective decrease in by EEG during sleep naturally varies with age (Leone brain aging for long-term meditation practice (Luders et al., et al., 2021; Paixao et al., 2020; Sun et al., 2019; Ye et al., 2016), music-making (Rogenmoser et al., 2018), and a higher 2020). Features from both time and frequency domains level of education (Steffener et al., 2016), as well as accelerated of each sleep stage are used to compute an overall brain aging associated with smoking and alcohol consumption age. Figure 8 shows the scatter plot of chronological age (Guggenmos et al., 2017; Ning et al., 2020). vs. sleep EEG-predicted brain age, and eight example sleep At the symposium, Dr. Ou presented his recent work (He EEGs from across the lifespan with their chronological et al., 2020, 2021) on a novel, deep convolutional neural net- age and calculated brain age shown. Dr. Sun showed that work brain age prediction model that uses both morphologi- across two large sleep EEG datasets, people with significant cal and contrast-based changes in brain MRI data to estimate neurological or psychiatric disease show a mean excess brain brain age. This work was enabled by collating 11 different age (compared to chronological age) of 4 years compared data sets and carefully curating a very large, harmonized to healthy controls on a population level, while those with dataset that included enough healthy subjects of all ages hypertension or diabetes show a mean excess brain age of to train, test and validate the method (Fig. 7a). By explic- 3.5 years compared to healthy controls (Sun et al., 2019). itly splitting the T1-weighted brain MRI into morphometry Sun and colleagues have validated the association of (spatial information) and contrast (tissue based signal infor- significant differences between sleep EEG based age and mation) channels, his attention-driven multi-channel fusion chronological age in patients with dementia and MCI (Ye network (Fig. 7b) improved the accuracy of age estimation et al., 2020), people diagnosed with HIV under antiretroviral as compared to each channel alone, or naive fusion of two therapy (Leone et al., 2021), and all cause mortality (Paixao channels without their proposed attention mechanisms, when et al., 2020). applied to 16,705 normal brain MRIs acquired over the lifes- As with sleep EEG, features of brain activity under gen- pan (0–97 years of age) (He et al., 2021). The team cross eral anesthesia have also been demonstrated to change with validated their work against multiple published brain age age, allowing the EEG patterns measured during adminis- estimation algorithms and using multiple independent test tration of general anesthesia to be evaluated as a marker of data sets (Fig. 7c). A critical advantage of this end-to-end brain age (Akeju et al., 2015; Lee et al., 2017; Purdon et al., method is that it has the potential to differentiate between 2015). Similar ideas about indicators of brain health under abnormal aging associated with contrast change (e.g., general anesthesia are motivating the development of EEG lesions) and those associated with morphometric changes machine learning methods to monitor and assess disorders (e.g., atrophy). This is an important contribution toward of consciousness, since no other behavioral markers can be increasing the specificity of brain age estimator biomark - used (Engemann et al., 2018). ers, a major issue for this line of research (Kaufmann et al., A fundamental challenge in using brain age estimation as 2019). an index of brain health and/or meaningful clinical indicator 1 3 Neuroinformatics (2022) 20:943–964 953 Fig. 8 Illustration of sleep EEG-based brain age. (Left) The scatter stages) (top in each subplot), where the top, middle, and bottom rows plot of chronological age vs. brain age where the diagonal dashed are patients with young, middle, and old chronological age (CA, in red line indicates where chronological age equals brain age. The years) respectively; while the left, middle, and right columns are sub- mean absolute deviation (MAD) is 7.8  years and Pearson’s correla- jects with young, middle, and old brain age (BA, in years). Compari- tion R = 0.82. (Right) The confusion matrix of example EEG spectro- son within each row reveals different sleep EEG microstructures for grams (bottom in each subplot) and hypnogram (trajectory of sleep different brain ages while at similar chronological age is that the rate of age-related changes in brain structure status at individual level is an active area of research (Al and function (e.g. sleep) vary across the lifespan such that Zoubi et al., 2018; Cole & Franke, 2017; Mohajer et al., early and late life changes are more readily detected, but 2020; Varikuti et al., 2018). For example, in the case of sleep are very subtle between 30 and 60 years of age. For both EEG-based brain age, the density of sleep spindles (count/ MRI- and EEG-based brain age prediction, the sensitivity is hour) one of the features used in the model, appears to be a lowest during this part of the lifespan. Not surprisingly, one heritable trait based on the expression of CACNA1l, a gene promising application of MRI-based brain age prediction that is associated with both schizophrenia and sleep spindle is early detection of future neuropsychiatric disorders in formation (Merikanto et al., 2019). children and/or adolescents (Chung et al., 2018). The relative Despite these limitations, it is intriguing, and poten- stability of structural MRI measures bound the temporal tially clinically advantageous, that lifestyle choices such resolution of brain age estimates using that modality (Cole as exercise and sleep can modify these quantitative metrics & Franke, 2017; Karch et al., 2019), while the significantly of brain age in directions that reflect known associations higher night-to-night variability of sleep EEG-based brain with brain health. Studies have shown that actively exercis- age estimates is both a challenge to overcome if looking ing leads to an orchestra of changes in energy metabolism, for stability, but also a potential additional source of oxidative stress, inflammation, tissue repair, growth factor meaningful signal to exploit in future work (Arnal et al., response, and regulatory pathways in the brain (Contrepois 2020; Arnardottir et al., 2021; Hogan et al., 2021). et al., 2020). Sleep has a bidirectional relationship with the There are important caveats to the use of a brain age as a immune system (Irwin, 2019), therefore there is evidence marker of brain health since deviations from chronological for and reason to expect that exercise can improve sleep and age could be due to multiple factors. Brain age estimation thereby improve brain health, which will be reflected in nor - studies remain population-level statistical tests because malized sleep-based brain age biomarkers in people with current approaches lack the sensitivity to accurately assess evidence of accelerated aging. clinically meaningful deviation at the individual patient level. Because of these limitations, brain age is currently viewed as a screening tool where large deviations call for Section V: Application of Machine Learning further investigation. More work is required to improve Techniques for Diagnostics, Prognostication, the specificity and clinical utility of brain age estimation. and Personalization of Medical Treatments Combining EEG- and MRI-based brain-age estimation techniques with and without additional features (e.g. Imaging plays a key role in the clinical evaluation of genomic markers, demographics, socioeconomics status, and pathological changes that can be readily distinguished environmental factors) to more accurately predict disease from a healthy brain. Neuroradiologists routinely use 1 3 954 Neuroinformatics (2022) 20:943–964 neuroimaging modalities such as CT, MRI, and PET for informed by both data from a specific patient and aggregated both qualitative and quantitative assessment of diseases information from larger patient datasets (Calhoun et al., 2021; from infectious, autoimmune, oncological, degenerative, Vieira et al., 2017; Zhang et al., 2020). and vascular etiologies. However, despite standardization Data-driven precision therapeutics are already being efforts, manual assessment is subject to inter- and intra- translated to the clinic using transcranial magnetic stimula- rater variability (Filippi et  al., 1995; Provenzale & tion (TMS) and other targeted brain stimulation approaches. Mancini, 2012; Provenzale et al., 2009; van Horn et al., At the symposium, Dr. Shan Siddiqi presented on this work, 2021). As such, there is intense interest in automating highlighting that TMS targets for any given symptom may radiological assessment with machine learning. A popular be identified based on the location of brain lesions that cause approach is radiomics (Beig et al., 2020), which focuses the same symptom (Cash et al., 2020; Davey & Riehl, 2005). on the extraction of pertinent quantitative imaging features Complementing Dr. Siddiqi and his team’s research is a often followed by incorporation of these features into a large body of work focused on machine learning-based target predictive machine learning algorithm. These imaging optimization of field distributions for transcranial magnetic features are computational imaging descriptors reflecting and/or electric stimulation that factor in the biophysical measures such as size, shape, intensity distribution, and properties of biological tissues or feedback from real-time intensity heterogeneity (Zhou et al., 2018). Indeed, these fMRI. “The Automatic Neuroscientist” framework uses feature-based radiomic approaches have found success for real-time fMRI in combination with Bayesian optimization early detection (Sørensen et al., 2016), diagnosis (Kniep “to automatically design the optimal experiment to evoke a et al., 2019; Regenhardt et al., 2021; Tanioka et al., 2020; desired target brain state.” (Lorenz et al., 2016). Machine Zhou et al., 2020), prognostication (Macyszyn et al., 2016; learning techniques have been applied towards rt-fMRI neu- Stefano et al., 2020; Tang et al., 2020), treatment response rofeedback studies, where a neurofeedback signal can be prediction/assessment (Cai et al., 2020; Chang et al., 2016; derived using supervised learning methods such as linear Hofmeister et al., 2020), and non-invasive determination models and support vector machines (LaConte et al., 2007). of molecular markers (Beig et  al., 2018; Pan et  al., In addition, both data driven and hypothesis driven analy- 2019) for a wide variety of diseases. More recently, deep ses of functional connectivity data have been used to predict learning approaches (Chang et al., 2018a, b; Rauschecker clinical outcomes including treatment response in patients et al., 2020; Titano et al., 2018) have gained traction for (Whitfield-Gabrieli et al., 2016) as well as to predict pedi- similar tasks due to these approaches foregoing the need atric vulnerability to psychiatric disorders including psy- to pre-engineer imaging features. Some approaches have chosis (Collin et al., 2019, 2020), depression (Chai et al., even shown the utility of combining radiomics with deep 2015), anxiety, and ADHD (Collin et al., 2020; Cui et al., learning (Lao et al., 2017; Xiao et al., 2019). While there 2020). At the symposium, Dr. Susan Whitfield-Gabrieli is great promise for these automated approaches, they do presented on these approaches, sharing evidence that con- not come without pitfalls. Radiomics, in particular, has nectivity between the medial prefrontal cortex (MPFC) and been challenged by variability stemming from differences the dorsolateral prefrontal cortex (DLFPC) can be used as in image acquisition, pre-processing, segmentation, and a biomarker to predict attentional problems in a normative feature implementation (Hoebel et  al., 2021; Kalpathy- pediatric population as assessed four years later, where Cramer et  al., 2016; Schwier et  al., 2019). Approaches greater baseline MPFC-DLPFC connectivity predicted to rectify these sources of variability and harmonize worsening of attentional issues (Whitfield-Gabrieli et al., radiomic features have been an active area of study (Carré 2020) while decreased baseline subgenual anterior cingulate et al., 2020; Marcadent et al., 2020; Orlhac et al., 2018; (sgACC)—DLPFC connectivity predicted worsening of anx- Parmar et al., 2014; Zwanenburg et al., 2020). Similarly, iety/depression. As psychiatric neuroimaging research has deep learning approaches also suffer from a lack of evolved from the description of patient cohorts using simple generalizability across different image acquisition settings group comparisons towards a focus on individual differences and patient populations (AlBadawy et al., 2018; Chang and “predictive” analytics, preliminary studies suggest that et al., 2020; Zech et al., 2018). These challenges will need intra-individual fluctuations of brain activity provide better to be addressed before these automated approaches can be prediction of symptoms than group-based studies. Machine effectively utilized. learning integrated with experience-sampling can be used to Beyond its neuroradiologic applications to promote brain produce novel brain-based predictive models of state fluctua- health with early detection, diagnostics, or prognostication tions (e.g., fluctuations of mind wandering) which general- related to neurologic disease, machine learning also has izes to both healthy and clinical populations (Kucyi et al., applications towards brain health as it relates to precision 2021). Dr. Whitfield-Gabrieli also highlighted the use of medicine—i.e. the development of personalized interventional mindfulness based rt-fMRI neurofeedback as a non-invasive, therapies for a broader range of neuropsychiatric disorders personalized circuit therapeutic to reduce symptom severity 1 3 Neuroinformatics (2022) 20:943–964 955 in psychotic patients as well as for teens with major depres- Machine intelligence in medical imaging is one of the sive disorder and/or anxiety. (Bauer et al., 2020; Stoeckel most vibrant fields within the application of machine learn- et al., 2014) These pioneering studies provide strong motiva- ing in healthcare, and one of its biggest subfields is quanti- tion to pursue imaging based treatments. tative imaging (QI). QI refers to extraction of quantifiable Overall, machine learning based methods have potential features from medical images that serve as biomarkers for to augment diagnostic and treatment workflows. As with all specific physiological conditions, such as features relating clinical interventions, the overarching goal is to improve to aspects of brain health which have been discussed above. patient outcomes, either within a specific decision point or A premarket submission for a QI function requires a func- longitudinally. While promising, more rigorous prospective tion description including the level of automation (manual, and external validation studies in diverse clinical scenarios semi-automatic or fully automatic), a brief description of and populations are needed before these methods can be the training algorithm, quantitative performance specifica- deployed for widespread use. tions, and instructions used for semi-automatic labeling of the training set. The biggest part of the premarket submis- sion is the technical performance assessment which should Section VI: Additional Considerations include a definition of the QI function, its relationship to for Clinical Deployment the measurand, and the use conditions. For example, this could be a “brain age” assessment from MRI data applica- ble to images of a specific resolution collected on a specific Regulatory Framework MRI system. It should also specify the performance metrics and characterize the performance of the QI function under To deploy any of the advances highlighted in the symposium the predefined conditions. In the mentioned example, per - that use machine learning algorithms in clinical practice, formance metrics could include accuracy as measured in proposals must first clear the regulatory process as set by deviation between actual age and estimated age in a nor- the Center for Devices and Radiological Health within the mative cohort as well as bias or precision as measured in FDA that handles medical devices. Most machine learn- reproducibility or repeatability. A priori acceptance criteria ing methods in healthcare are categorized as software as a regarding these performance metrics should also be set along medical device (SaMD) which is a subcategory under soft- with restrictions and limitations on usage, and the results of ware related to medical devices under the medical device a study presented where the outcomes are compared to the umbrella. The pathway to market depends on the risk asso- predefined acceptance criteria. ciated with the software, which in turn depends primarily on 1) significance of information provided by SaMD to a healthcare decision and 2) state of healthcare situation or Ethical Considerations During Machine Learning condition; a more critical situation yields a higher risk rat- Model Development ing. Traditionally, SaMD algorithms need to be locked, i.e., give the same output for the same input, after they are sub- As the machine learning applications in this report mature in mitted to the FDA for premarket approval. This is impracti- their development, there are a number of vital ethical issues cal for machine learning software in situations where it is to be taken into consideration. While not the primary focus often desirable to continuously update the machine learning of the symposium, both the organizer, Dr. Randy Gollub models based on user data (e.g. to accommodate updates in and keynote speaker, Dr. Simon Eickhoff emphasized the scanner hardware and software). The FDA has proposed a importance of these aspects, pointing out a few examples of new regulatory framework based on a total product life cycle how, where, and why they are relevant. Some of these ethical approach, wherein the initial premarket submission outlines considerations have established guidelines or technical best the modifications that might take place in the future. The practices that need to be more widely used; others are ongoing manufacturer can then continuously update their machine discussions for which there is not yet a clear-cut solution (see learning models based on new user data without having to for example the Fair ML for Health Workshop that was held go through a new premarket submission provided that the during the NeurIPS 2019 Workshop (Fair ML for Health— update is within the SaMD Pre-Specifications and algo- Accepted Papers, 2021,  https:// www . f air m lf or h ealt h. com/ rithm change protocol (Digital Health Center of Excellence, accep ted- papers). It is crucial that scientists and researchers 2021, https:// www. fda. gov/ medic al- devic es/ digit al- health- participate actively in these discussions at each stage of center- excel lence/ sof tw ar e- medic al- device- samd). These development of these methods. We note that this section of guidelines are under active discussion, development, and our report is by no means comprehensive; for more in-depth refinement in collaboration with industry, academic, and discussions of these issues, see (Beauvais et al., 2021) and clinical leaders. (Chen et al., 2020). 1 3 956 Neuroinformatics (2022) 20:943–964 Data Sharing research use. With the increasing use of large datasets for multi- ple studies and across long periods of time, it is difficult to track Data sharing across institutions may eventually become nec- all of the downstream uses of a single person’s data. After data essary to create large enough datasets to train sophisticated collection, development and validation of new methods, these machine learning algorithms. To mitigate the risks of breach methods may eventually be commercialized. However, inherent of privacy, security, and confidentiality, robust de-identification to any trained machine learning model or algorithm is the data algorithms that retain all necessary imaging data elements are that was used for such training. The data is inextricably tied to essential, and all modalities of data must be scrutinized to ensure any intellectual property or commercial potential that results that there are not additional unintended sources of protected from the development process. Is it fair to allow patenting of information amongst them (e.g. private DICOM metadata tags). trained models or algorithms on data collected from people who Secure cloud servers and backup protocols, expert curation and did not consent to its possible use for commercial prot fi ? Should maintenance, and strict guidelines and training for researchers those people be included in any such profit? These questions on how to securely access, store, and dispose of data are all addi- must be answered as machine learning models become inte- tional tools to minimize the chances of confidentiality breaches grated into clinical pipelines. or loss of data. Federated learning methods which allow data to be stored only at the location where it was collected while allow- Bias in Datasets ing for multisite analysis are another means to support robust, yet protected, data sharing (K. Chang, Balachandar, et al., 2018; In machine learning, the phrase “garbage in, garbage out” Chang, Grinband, et al., 2018). For all these approaches, fre- reflects the fact that bias, noise, or flaws in the underlying quent communication between all institutions involved will data used to train a model will undoubtedly affect the qual- also ensure that everyone is kept apprised of possible issues in ity, accuracy, and validity of the results. Therefore, ensuring a timely manner and that any changes are implemented in an high quality data that is highly representative of the popula- organized and efficient manner. tions under study is paramount to the ethical and effective development of these methods. One key component of this Informed Consent for Expanded or Later Use of Data for assessments of brain health is ensuring adequate repre- sentation of traditionally underrepresented subpopulations It is becoming increasingly common for large datasets to be used in research, including underrepresented minorities, women, and reused in multiple studies and towards die ff rent machine low and middle income nationals, transgender and gender learning algorithms once they have been collected. This is non-conforming individuals, undocumented immigrants, mainly due to the cost in time, money, and resources to amass and pregnant women, especially from an intersectional lens. an entirely new dataset for each research question. Repurposing This is especially important because of the specific men- existing datasets across many studies is overall a very efficient tal and behavioral health issues which impact brain health and effective option; however, the wishes of those from whom in many of these subpopulations. It also includes consid- the data is collected must be respected. Most current informed erations in the study design itself to ensure these popula- consent paradigms are based on data being collected for a sin- tions are not excluded inadvertently by data acquisition gle study and therefore obtain informed consent from a patient methods, for example by failing to include more than the for that single study alone. However, this system needs to be traditional binary options when documenting gender. Even modified to reflect that it is likely a patient’s data could be used if the intentions of researchers are to include all popula- for many studies even decades into the future, most of which tions, other aspects of the study design can inadvertently cannot even be fathomed at the time of data collection. Patients be biased towards certain populations. For example, studies should, at minimum, have the ability to ‘opt out’ of having their that require mobile phone downloads of certain apps or track data used in future studies without their explicit consent. Many social media use exclude populations who do not have access current guidelines state that if the data is de-identified, it can to smartphones or social media. be shared and used for new studies without re-obtaining con- Even once a study is underway, oversight and periodic sent, often after obtaining a waiver of consent from the local assessments of study recruitment practices should be done Institutional Review Board. However, current trends warrant re- to check for inadvertent exclusion of certain populations. For examination of these guidelines with a goal of securing patient example, the inclusion and exclusion criteria of many stud- consent for wider, protected data use at the time of enrollment. ies, especially randomized controlled trials, are often writ- ten with purely scientific or clinical considerations in mind Intellectual Property and Commercialization relating to the treatment or diagnostic in question. However, they can result in a study population that is too restricted Clearer intellectual property guidelines are needed regarding and not reflective of the actual population of interest. Stud- models or algorithms developed from patient data collected for ies that involve multiple study visits at different times may 1 3 Neuroinformatics (2022) 20:943–964 957 discourage participation of those who do not have the access are non-trivial. How should such findings be handled? Do or flexibility to come to the research site several times. Any clinicians have an obligation to inform patients? What about researcher using existing datasets should hold to these same for measures such as one of the indices of ‘brain age’ for standards when checking the subject profile of the already which the implications are still under study? Clear clinical collected data and include this information, including limita- practice guidelines for the handling of sensitive informa- tions of the dataset, in any resulting publications. tion relating to brain health must be developed prior to the Technical approaches have been suggested as tools for deployment of any such algorithm. handling disproportionate representation of certain sub- While many of these ethical considerations are gray areas groups within large datasets. For example, some approaches for which we can only postulate guidelines and not clear force neural networks to learn intermediate representations answers, ongoing discussion will hopefully lead to new which cannot be used to predict a protected attribute of inter- best practices that adhere to the highest ethical standards est, e.g. gender or race (Dinsdale et al., 2021). on each of the issues discussed, to safeguard patients and their anatomic and physiological datasets. For researchers, Quantifying/understanding Uncertainty it is important to keep in mind that the misinterpretation or generalization of one poorly designed high-profile study or Whenever possible, researchers and scientists should attempt one breach in confidentiality can be enough for an entire to quantify the uncertainty of their model predictions using field to lose credibility. statistical tools such as the confidence interval. Before such algorithms are implemented, clinicians should receive training on how to interpret the results given the limita- Conclusion tions of any model, including uncertainty. There is much attention being focused on these issues, including annual The MGH McCance Center for Brain Health and Harvard- workshops that have been held since 2019 at the MICCAI MIT Health Sciences and Technology Neuroimaging Training meetings- “UNSURE Uncertainty for Safe Utilization of Program co-hosted virtual symposium, “Neuroimaging Machine Learning in Medical Imaging” with presentations Indicators of Brain Structure and Function—Closing the Gap and awards for work in areas such as risk management of Between Research and Clinical Application,” explored the machine learning systems in clinical pipelines, measurement recent explosion of machine learning approaches augmenting errors, methods for modeling noise in data, validation of the clinical and scientific neuroimaging pipeline. Researchers uncertainty estimates, calibration of uncertainty measures presented cutting-edge techniques for acquiring more informative and more (Greenspan et al., 2019). imaging data, more effectively analyzing this acquired data, and more precisely acting on the insights from this analysis Ethical Considerations at the Stage of Deploying to guide and individualize treatment decisions. The work a Machine Learning Model presented at this symposium highlighted several open research directions which must be explored in order to implement these Privacy and Confidentiality of Information techniques in practice. For example, progress in this field will require techniques for robust generalization of machine learning As machine learning tools for brain health emerge, an techniques to more realistic, heterogeneous datasets as well important question to answer is: who should have access as methods for identifying the uncertainty present in machine to the outputs of these methods? Should the patient have learning-based predictions and presenting this information to unfettered access to their own ‘brain health’ information? end users within a clinical workflow. Toward these ends, our Should all clinicians who might interact with that patient field will need to ensure the availability of sufficiently large, have access? If it will be integrated into the medical record, curated data sets; the ability to share valuable data sets thus how do we prevent it from affecting billing or insurance engaging a diverse, committed scientific community (Eickhoff practices? How might the information bias someone’s inter- et al., 2016); and responsible stewardship of brain imaging data action with a patient, especially without accurate reporting to ensure appropriate protections for individual privacy as well (with uncertainty)? These are all considerations that cannot as intellectual property and proper handling of bias in these data. be taken lightly and will have to be addressed to develop With a firm commitment in these directions, machine learning clinical best practices. promises to dramatically improve the early detection, prediction, and treatment of diseases that threaten brain health. Incidental Findings Note: Interested readers may view recorded videos of all symposium presentations and discussions at this As with genetic information and testing, the probabilities of link https://www .y outube. com/ pla ylis t?lis t=PL0A -NKHL V incidental findings in large datasets such as neuroimaging rNF82 vdjey yaBRo iXg77 lCeW. 1 3 958 Neuroinformatics (2022) 20:943–964 Acknowledgements We would like to thank Dr. Ana Namburete for acquisition and sleep staging. Sleep, 43, zsaa097. https:// doi. org/ 10. her helpful review and input on sections of this manuscript. We also 1093/ sleep/ zsaa0 97 would like to extend our appreciation to the Neuroimaging Training Arnardottir, E. S., Islind, A. S., & Óskarsdóttir, M. (2021). The Future Program students and faculty members who participated in the Sympo- of Sleep Measurements: A Review and Perspective. Sleep Medi- sium report preparation and/or presentations- Bruce Rosen, Jayashree cine Clinics, 16, 447–464. https:// doi. org/ 10. 1016/j. jsmc. 2021. Kalpathy-Cramer, John Gustaf Wilhelm Samuelsson, and Katharina 05. 004 Viktoria Hoebel. Bahadir, C. D., Wang, A. Q., Dalca, A. V., & Sabuncu, M. R. (2020). Deep-learning-based Optimization of the Under-sampling Pat- tern in MRI EEE TCP. Transactions Computational Imaging, Funding Open Access funding provided by the MIT Libraries. 6, 1139–1152. Bashyam, V. M., Erus, G., Doshi, J., Habes, M., Nasrallah, I.M., Truelove- Open Access This article is licensed under a Creative Commons Attri- Hill, M., Srinivasan, D., Mamourian, L., Pomponio, R., Fan, Y., bution 4.0 International License, which permits use, sharing, adapta- Launer, L. J., Masters, C.L., Maruff, P., Zhuo, C., Völzke, H., tion, distribution and reproduction in any medium or format, as long Johnson, S. C., Fripp, J., Koutsouleris, N., Satterthwaite, T. D., as you give appropriate credit to the original author(s) and the source, Davatzikos, C., on behalf of the ISTAGING Consortium, the P. A. provide a link to the Creative Commons licence, and indicate if changes disease C., ADNI, and CARDIA studies. (2020). MRI signatures of were made. The images or other third party material in this article are brain age and disease over the lifespan based on a deep brain network included in the article's Creative Commons licence, unless indicated and 14 468 individuals worldwide. Brain, 143, 2312–2324. https:// otherwise in a credit line to the material. If material is not included in doi. org/ 10. 1093/ brain/ awaa1 60. the article's Creative Commons licence and your intended use is not Bauer, C. C. C., Rozenkrantz, L., Caballero, C., Nieto-Castanon, A., permitted by statutory regulation or exceeds the permitted use, you will Scherer, E., West, M. R., Mrazek, M., Phillips, D. T., Gabrieli, need to obtain permission directly from the copyright holder. To view a J. D. E., & Whitfield-Gabrieli, S. (2020). Mindfulness training copy of this licence, visit http://cr eativ ecommons. or g/licen ses/ b y/4.0/ . preserves sustained attention and resting state anticorrelation between default-mode network and dorsolateral prefrontal cor- tex: A randomized controlled trial. Human Brain Mapping, 41, 5356–5369. https:// doi. org/ 10. 1002/ hbm. 25197 Baur, C., Wiestler, B., Muehlau, M., Zimmer, C., Navab, N., & References Albarqouni, S. (2021). Modeling Healthy Anatomy with Artificial Intelligence for Unsupervised Anomaly Detection in Ahmadlou, M., & Adeli, H. (2011). Functional community analysis of Brain MRI. Radiology Artificial Intelligence, 3,. https:// doi. org/ brain: A new approach for EEG-based investigation of the brain 10. 1148/ ryai. 20211 190169 pathology. Neuro Image, 58, 401–408. https://doi. or g/10. 1016/j. Beauvais, M. J. S., Knoppers, B. M., & Illes, J. (2021). A marathon, not neuro image. 2011. 04. 070. a sprint – neuroimaging. Open Science and Ethics Neuroimage, Akeju, O., Pavone, K. J., Thum, J. A., Firth, P. G., Westover, M. B., 236. https:// doi. org/ 10. 1016/j. neuro image. 2021. 118041 Puglia, M., Shank, E. S., Brown, E. N., & Purdon, P. L. (2015). Beig, N., Bera, K., & Tiwari, P. (2020). Introduction to radiomics and Age-dependency of sevoflurane-induced electroencephalogram radiogenomics in neuro-oncology: implications and challenges. dynamics in children. British Journal of Anaesthesia, 115, i66– Neuro-Oncology Advance 2, iv3–iv14. https:// doi. org/ 10. 1093/ i76. https:// doi. org/ 10. 1093/ bja/ aev114 noajnl/ vdaa1 48 Akkus, Z., Galimzianova, A., Hoogi, A., Rubin, D. L., & Erickson, Beig, N., Patel, J., Prasanna, P., Hill, V., Gupta, A., Correa, R., Bera, K., B. J. (2017). Deep Learning for Brain MRI Segmentation: State Singh, S., Partovi, S., Varadan, V., Ahluwalia, M., Madabhushi, A., of the Art and Future Directions. Journal of Digital Imaging, & Tiwari, P. (2018). Radiogenomic analysis of hypoxia pathway is 30, 449–459. https:// doi. org/ 10. 1007/ s10278- 017- 9983-4 predictive of overall survival in Glioblastoma. Science and Reports, Al Zoubi, O., Ki Wong, C., Kuplicki, R. T., Yeh, H., Mayeli, A., Refai, 8, 7. https:// doi. org/ 10. 1038/ s41598- 017- 18310-0 H., et al. (2018). Predicting Age From Brain EEG Signals—A Brandes, U., & Pich, C. (2007). Eigensolver Methods for Progressive Machine Learning Approach. Frontiers Aging Neuroscience, 10, Multidimensional Scaling of Large Data, in: Kaufmann, M., 184. https:// doi. org/ 10. 3389/ fnagi. 2018. 00184. Wagner, D. (Eds.), Graph Drawing, Lecture Notes in Computer Al-Awami, A. K., Beyer, J., Strobelt, H., Kasthuri, N., Lichtman, J. Science. Springer, Berlin, Heidelberg, pp. 42–53. https://doi. or g/ W., Pfister, H., & Hadwiger, M. (2014). NeuroLines: A Subway 10. 1007/ 978-3- 540- 70904-6_6 Map Metaphor for Visualizing Nanoscale Neuronal Connectivity. Cai, J., Zheng, J., Shen, J., Yuan, Z., Xie, M., Gao, M., Tan, H., Liang, IEEE Transactions on Visualization and Computer Graphics, 20, Z., Rong, X., Li, Y., Li, H., Jiang, J., Zhao, H., Argyriou, A. 2369–2378. https:// doi. org/ 10. 1109/ TVCG. 2014. 23463 12 A., Chua, M. L. K., & Tang, Y. (2020). A Radiomics Model for AlBadawy, E. A., Saha, A., & Mazurowski, M. A. (2018). Deep learning Predicting the Response to Bevacizumab in Brain Necrosis after for segmentation of brain tumors: Impact of cross-institutional train- Radiotherapy. Clinical Cancer Research, 26, 5438–5447. https:// ing and testing. Medical Physics, 45, 1150–1158. https:// doi. org/ 10. doi. org/ 10. 1158/ 1078- 0432. CCR- 20- 1264 1002/ mp. 12752 Calhoun, V. D., Pearlson, G. D., & Sui, J. (2021). Data-driven Alhussein, M., Muhammad, G., & Hossain, M. S. (2019). EEG Pathol- approaches to neuroimaging biomarkers for neurological and ogy Detection Based on Deep Learning. IEEE Access, 7, 27781– psychiatric disorders: Emerging approaches and examples. Cur- 27788. https:// doi. org/ 10. 1109/ ACCESS. 2019. 29016 72 rent Opinion in Neurology, 34, 469–479. https://doi. or g/10. 1097/ Anand, C. S., & Sahambi, J. S. (2010). Wavelet domain non-linear WCO. 00000 00000 000967 filtering for MRI denoising. Magnetic Resonance Imaging, 28, Carré, A., Klausner, G., Edjlali, M., Lerousseau, M., Briend-Diop, J., 842–861. https:// doi. org/ 10. 1016/j. mri. 2010. 03. 013 Sun, R., Ammari, S., Reuzé, S., Alvarez Andres, E., Estienne, Arnal, P. J., Thorey, V., Debellemaniere, E., Ballard, M. E., Bou Hernandez, T., Niyoteka, S., Battistella, E., Vakalopoulou, M., Dhermain, F., A., Guillot, A., Jourde, H., Harris, M., Guillard, M., Van Beers, Paragios, N., Deutsch, E., Oppenheim, C., Pallud, J., & Robert, C. P., Chennaoui, M., & Sauvet, F. (2020). The Dreem Headband (2020). Standardization of brain MR images across machines and compared to polysomnography for electroencephalographic signal 1 3 Neuroinformatics (2022) 20:943–964 959 protocols: Bridging the gap for MRI-based radiomics. Science and Cole, J. H., Ritchie, S. J., Bastin, M. E., Valdés Hernández, M. C., Reports, 10, 12340. https:// doi. org/ 10. 1038/ s41598- 020- 69298-z Muñoz Maniega, S., Royle, N., Corley, J., Pattie, A., Harris, S. Cash, R. F. H., Weigand, A., Zalesky, A., Siddiqi, S. H., Downar, J., Fitzger- E., Zhang, Q., Wray, N. R., Redmond, P., Marioni, R. E., Starr, ald, P .B., & Fox, M D. (2020). Using Brain Imaging to Improve J. M., Cox, S. R., Wardlaw, J. M., Sharp, D. J., & Deary, I. J. Spatial Targeting of Transcranial Magnetic Stimulation for Depres- (2018). Brain age predicts mortality. Molecular Psychiatry, 23, sion Biological Psychiatry S0006322320316681. https:// doi. org/ 10. 1385–1392. https:// doi. org/ 10. 1038/ mp. 2017. 62 1016/j. biops ych. 2020. 05. 033 Collin, G., Nieto-Castanon, A., Shenton, M. E., Pasternak, O., Kelly, S., Cetin Karayumak, S., Bouix, S., Ning, L., James, A., Crow, T., Shenton, Keshavan, M. S., Seidman, L. J., McCarley, R. W., Niznikiewicz, M., et al. (2019). Retrospective harmonization of multi-site diffu- M. A., Li, H., Zhang, T., Tang, Y., Stone, W. S., Wang, J., & sion MRI data acquired with different acquisition parameters. Neuro Whitfield-Gabrieli, S. (2019). Brain functional connectivity data Image, 184, 180–200. https://d oi.o rg/1 0.1 016/j.n euroi mage.2 018.0 8. enhance prediction of clinical outcome in youth at risk for psy- 073. chosis. NeuroImage Clin., 26, 102108. https:// doi. org/ 10. 1016/j. Chai, X. J., Hirshfeld-Becker, D., Biederman, J., Uchida, M., Doehrmann, nicl. 2019. 102108 O., Leonard, J. A., et al. (2015). Functional and structural brain cor- Collin, G., Seidman, L. J., Keshavan, M. S., Stone, W. S., Qi, Z., Zhang, T., relates of risk for major depression in children with familial depres- Tang, Y., Li, H., Anteraper, S. A., Niznikiewicz, M. A., McCarley, sion. Neuro Image Clinical, 8, 398–407. https:// doi. org/ 10. 1016/j. R. W., Shenton, M. E., Wang, J., & Whitfield-Gabrieli, S. (2020). nicl. 2015. 05. 004. Functional connectome organization predicts conversion to psychosis Chang, K., Balachandar, N., Lam, C., Yi, D., Brown, J., Beers, A., in clinical high-risk youth from the SHARP program. Molecular Psy- Rosen, B., Rubin, D. L., & Kalpathy-Cramer, J. (2018). Dis- chiatry, 25, 2431–2440. https://doi. or g/10. 1038/ s41380- 018- 0288-x tributed deep learning networks among institutions for medical Contrepois, K., Wu, S., Moneghetti, K. J., Hornburg, D., Ahadi, S., Tsai, M.-S., imaging. Journal of the American Medical Informatics Associa- Metwally, A. A., Wei, E., Lee-McMullen, B., Quijada, J. V., Chen, S., tion, 25, 945–954. https:// doi. org/ 10. 1093/ jamia/ ocy017 Christle, J. W., Ellenberger, M., Balliu, B., Taylor, S., Durrant, M. G., Chang, K., Beers, A. L., Brink, L., Patel, J. B., Singh, P., Arun, N. T., Hoe- Knowles, D. A., Choudhry, H., Ashland, M., & Snyder, M. P. (2020). bel, K. V., Gaw, N., Shah, M., Pisano, E. D., Tilkin, M., Coombs, Molecular Choreography of Acute Exercise. Cell, 181, 1112-1130.e16. L. P., Dreyer, K. J., Allen, B., Agarwal, S., & Kalpathy-Cramer, J. https:// doi. org/ 10. 1016/j. cell. 2020. 04. 043 (2020). Multi-Institutional Assessment and Crowdsourcing Evalua- Cui, H., Giuliano, A. J., Zhang, T., Xu, L., Wei, Y., Tang, Y., Qian, tion of Deep Learning for Automated Classification of Breast Den- Z., Stone, L. M., Li, H., Whitfield-Gabrieli, S., Niznikiewicz, sity. Journal of the American College of Radiology, 17, 1653–1662. M., Keshavan, M. S., Shenton, M. E., Wang, J., & Stone, W. https:// doi. org/ 10. 1016/j. jacr. 2020. 05. 015 S. (2020). Cognitive dysfunction in a psychotropic medication- Chang, K., Zhang, B., Guo, X., Zong, M., Rahman, R., Sanchez, D., naïve, clinical high-risk sample from the ShangHai-At-Risk-for- et al. (2016). Multimodal imaging patterns predict survival in Psychosis (SHARP) study: Associations with clinical outcomes. recurrent glioblastoma patients treated with bevacizumab. Neuro- Schizophr. Res. Biomarkers in the Attenuated Psychosis Syn- Oncology, 18, 1680–1687. https:// doi. or g/ 10. 1093/ neuonc/ drome, 226, 138–146. https:// doi. org/ 10. 1016/j. schres. 2020. 06. now086. 018 Chang, P., Grinband, J., Weinberg, B. D., Bardis, M., Khy, M., Cadena, Dalca, A. V., Yu, E., Golland, P., Fischl, B., Sabuncu, M. R., & Iglesias, G., Su, M.-Y., Cha, S., Filippi, C. G., Bota, D., Baldi, P., Poisson, J. E. (2019). Unsupervised Deep Learning for Bayesian Brain L. M., Jain, R., & Chow, D. (2018). Deep-Learning Convolu- MRI Segmentation. ArXiv190411319v2. tional Neural Networks Accurately Classify Genetic Mutations in Davey, K., & Riehl, M. (2005). Designing transcranial magnetic stimu- Gliomas. American Journal of Neuroradiology, 39, 1201–1207. lation systems. IEEE Transactions on Magnetics, 41, 1142–1148. https:// doi. org/ 10. 3174/ ajnr. A5667https:// doi. org/ 10. 1109/ TMAG. 2004. 843326 Chen, I.Y., Pierson, E., Rose, S., Joshi, S., Ferryman, K., & Ghassemi, M. Despotović, I., Goossens, B., & Philips, W. (2015). MRI segmenta- (2020). Ethical Machine Learning in Health Care. ArXiv200910576 Cs. tion of the human brain: Challenges, methods, and applications. Chun, S. Y., Reese, T. G., Ouyang, J., Guerin, B., Catana, C., Zhu, X., Computational and Mathematical Methods in Medicine, 2015, et al. (2012). MRI-based nonrigid motion correction in simul- 450341. https:// doi. org/ 10. 1155/ 2015/ 450341 taneous PET/MRI. Journal Nuclear Medicine, 53, 1284–1291. Digital Health Center of Excellence Software as a Medical Device https:// doi. org/ 10. 2967/ jnumed. 111. 092353. (SaMD). FDA. (2021). https:// www. fda. gov/ medic al- devic es/ Chung, Y., Addington, J., Bearden, C. E., Cadenhead, K., Cornblatt, digit al- health- center- excel lence/ softw are- medic al- device- samd B., Mathalon, D. H., McGlashan, T., Perkins, D., Seidman, L.J., (Accessed 15 June 2021). Tsuang, M., Walker, E., Woods, S.W., McEwen, S., van Erp, T. Dinsdale, N. K., Jenkinson, M., & Namburete, A. I. L. (2021). Deep G. M., & Cannon, T. D. (2018). North American Prodrome Lon- learning-based unlearning of dataset bias for MRI harmonisation gitudinal Study (NAPLS) Consortium and the Pediatric Imaging, and confound removal. Neuro Image, 228,117689. https:// doi. Neurocognition, and Genetics (PING) Study Consortium. Use of org/ 10. 1016/j. neuro image. 2020. 117689 Machine Learning to Determine Deviance in Neuroanatomical Dong, X., Lei, Y., Wang, T., Higgins, K., Liu, T., Curran, W. J., Mao, H., Maturity Associated With Future Psychosis in Youths at Clini- Nye, J. A., & Yang, X. (2020). Deep learning-based attenuation cally High Risk. JAMA Psychiatry, 75, 960–968. https://doi. or g/ correction in the absence of structural information for whole-body 10. 1001/ jamap sychi atry. 2018. 1543. positron emission tomography imaging. Physics in Medicine & Cole, J. H., & Franke, K. (2017). Predicting Age Using Neuroimaging: Biology, 65, 055011. https:// doi. org/ 10. 1088/ 1361- 6560/ ab652c Innovative Brain Ageing Biomarkers. Trends in Neurosciences, Edupuganti, V., Mardani, M., Vasanawala, S., & Pauly, J. (2021). Uncer- 40, 681–690. https:// doi. org/ 10. 1016/j. tins. 2017. 10. 001 tainty Quantification in Deep MRI Reconstruction. IEEE Transac- Cole, J. H., Leech, R., & Sharp, D. J. (2015). Prediction of brain age tions on Medical Imaging, 40, 239–250. https:// doi. org/ 10. 1109/ suggests accelerated atrophy after traumatic brain injury. Annals TMI. 2020. 30250 65 of Neurology, 77, 571–581. https:// doi. org/ 10. 1002/ ana. 24367 Eickhoff, S., Nichols, T. E., Van Horn, J. D., & Turner, J. A. (2016). Cole, J. H., Marioni, R. E., Harris, S. E., & Deary, I. J. (2019). Brain Sharing the wealth: Neuroimaging data repositories. Neuro age and other bodily “ages”: Implications for neuropsychiatry. Image, 124, 1065–1068. https:// doi. org/ 10. 1016/j. neuro image. 2015. 10. 079. Molecular Psychiatry, 24, 266–281. https:// doi. or g/ 10. 1038/ s41380- 018- 0098-1 1 3 960 Neuroinformatics (2022) 20:943–964 Engemann, D. A., Raimondo, F., King, J. -R., Rohaut, B., Louppe, image-domain metal artifact reduction, in: Developments in G., Faugeras, F., et al. (2018). Robust EEG-based cross-site and X-Ray Tomography XI. Presented at the Developments in X-Ray cross-protocol classification of states of consciousness. Brain, Tomography XI, International Society for Optics and Photonics, 141, 3179–3192. https:// doi. org/ 10. 1093/ brain/ awy251. 103910W. https:// doi. org/ 10. 1117/ 12. 22744 27 Escudero, J., Abásolo, D., Hornero, R., Espino, P., & López, M. (2006). Greenspan, H., Tanno, R., Erdt, M., Arbel, T., Baumgartner, C., Dalca, Analysis of electroencephalograms in Alzheimer’s disease patients A., Sudre, C. H., Wells, W. M., Drechsler, K., & Linguraru, M. with multiscale entropy. Physiological Measurement, 27, 1091– G. (2019). Uncertainty for Safe Utilization of Machine Learn- 1106. https:// doi. org/ 10. 1088/ 0967- 3334/ 27/ 11/ 004 ing in Medical Imaging and Clinical Image-Based Procedures: Esteban, O., Birman, D., Schaer, M., Koyejo, O. O., Poldrack, R. A., First International Workshop, UNSURE 2019, and 8th Interna- & Gorgolewski, K. J. (2017). MRIQC: Advancing the automatic tional Workshop, CLIP 2019, Held in Conjunction with MICCAI prediction of image quality in MRI from unseen sites. PloS One, 2019, Shenzhen, China, October 17, 2019, Proceedings Springer 12,. Nature. Esteban, O., Blair, R. W., Nielson, D. M., Varada, J. C., Marrett, Griswold, M. A., Jakob, P. M., Heidemann, R. M., Nittka, M., Jellus, S., Thomas, A. G., et al. (2019). Crowdsourced MRI quality V., Wang, J., Kiefer, B., & Haase, A. (2002). Generalized auto- metrics and expert quality annotations for training of humans calibrating partially parallel acquisitions (GRAPPA). Magnetic and machines. Science Data, 6, 1–7. https:// doi. org/ 10. 1038/ Resonance in Medicine, 47, 1202–1210. https://doi. or g/10. 1002/ s41597- 019- 0035-4mrm. 10171 Fair ML for Health - Accepted Papers. (2021). https://www .f airmlf orhealt h. Guggenmos, M., Schmack, K., Sekutowicz, M., Garbusow, M., com/ accep ted- papers (Accessed 28 July 2021). Sebold, M., Sommer, C., et al. (2017). Quantitative neurobio- FDA-NIH Biomarker Working Group. (2016). BEST (Biomarkers, logical evidence for accelerated brain aging in alcohol depend- EndpointS, and other Tools) Resource. Food and Drug Admin- ence. Translational Psychiatry, 7, 1–7. https:// doi. org/ 10. 1038/ istration (US), Silver Spring (MD).s41398- 017- 0037-y. Filippi, M., Horsfield, M.A., Bressi, S., Martinelli, V., Baratti, C., Reganati, Guimond, A., Meunier, J., & Thirion, J. -P. (2000). Average Brain Models: P., Campi, A., Miller, D.H., & Comi, G. (1995). Intra- and inter- A Convergence Study. Computer Vision and Image Understanding, observer agreement of brain MRI lesion volume measurements in 77, 192–210. https:// doi. org/ 10. 1006/ cviu. 1999. 0815. multiple sclerosis. A comparison of techniques. Brain Journal Neu- Haehn, D., Rannou, D., Ahtam, B., Grant, P., & Pienaar, R. (2014). rology, 118(Pt 6), 1593–1600. https:// doi. org/ 10. 1093/ brain/ 118.6. Neuroimaging in the Browser using the X Toolkit Front. Neuro- 1593 informatics 8. https://doi. or g/10. 3389/ conf. fninf. 2014. 08. 00101 Fischl, B., Salat, D. H., Busa, E., Albert, M., Dieterich, M., Haselgrove, Hammernik, K., Klatzer, T., Kobler, E., Recht, M. P., Sodickson, D. C., van der Kouwe, A., Killiany, R., Kennedy, D., Klaveness, K., Pock, T., & Knoll, F. (2018). Learning a variational network S., Montillo, A., Makris, N., Rosen, B., & Dale, A. M. (2002). for reconstruction of accelerated MRI data. Magnetic Resonance Whole brain segmentation: Automated labeling of neuroanatomi- in Medicine, 79, 3055–3071. https://doi. or g/10. 1002/ mr m.26977 cal structures in the human brain. Neuron, 33, 341–355. https:// Han, Y. S., Yoo, J., Ye, J. C. (2018). Deep Learning with Domain doi. org/ 10. 1016/ s0896- 6273(02) 00569-x Adaptation for Accelerated Projection-Reconstruction MR. Franke, K., Gaser, C., Manor, B., & Novak, V. (2013). Advanced ArXiv170301135 Cs. http:// arxiv. org/ abs/ 1703. 01135 BrainAGE in older adults with type 2 diabetes mellitus. Fron- Haskell, M. W., Cauley, S. F., & Wald, L. L. (2018). Targeted Motion tiers Aging Neuroscience 5 https:// doi. org/ 10. 3389/ fnagi. 2013. Estimation and Reduction (TAMER): Data Consistency Based 00090 Motion Mitigation for MRI using a Reduced Model Joint Optimi- Franke, L., & Haehn, D. (2020). Modern Scientific Visualizations on the zation. IEEE Transactions on Medical Imaging, 37, 1253–1265. Web. Informatics, 7, 37. https://doi. or g/10. 3390/ inf ormatic s7040 037 https:// doi. org/ 10. 1109/ TMI. 2018. 27914 82 Franke, L., Weidele, D. K. I., Zhang, F., Cetin-Karayumak, S., Pieper, He, S., Gollub, R. L., Murphy, S. N., Perez, J. D., Prabhu, S., Pienaar, S., O’Donnell, L. J., Rathi, Y., & Haehn, D. (2020). FiberStars: R., et al. (2020). Brain Age Estimation Using LSTM on Children’s Visual Comparison of Diffusion Tractography Data between Brain MRI. Proceeding IEEE International Symposium Biomedical Multiple Subjects. ArXiv200508090 Cs. Imaging, 2020, 420–423. https:// doi. org/ 10. 1109/ isbi4 5749. 2020. Gajawelli, N., Tsao, S., Kromnick, M., Nelson, M., & Leporé, N. 90983 56. (2019). Image Postprocessing Adoption Trends in Clinical Medi- He, S., Pereira, D., David Perez, J., Gollub, R. L., Murphy, S. N., Prabhu, S., cal Imaging. Journal of the American College of Radiology, 16, Pienaar, R., Robertson, R. L., Ellen, Grant, P., & Ou Y. (2021). Multi- 945–951. https:// doi. org/ 10. 1016/j. jacr. 2019. 01. 005 channel Attention-Fusion Neural Network for Brain Age Estimation: Gallego-Jutglà, E., Solé-Casals, J., Vialatte, F. -B., Elgendi, M., Accuracy, Generality, and Interpretation with 16,705 Healthy MRIs Cichocki, A., & Dauwels, J. (2015). A hybrid feature selection across. Lifespan Medical Image Analysis 102091 https:// doi. org/ 10. approach for the early diagnosis of Alzheimer’s disease. Journal 1016/j. media. 2021. 102091 of Neural Engineering, 12,016018. https://doi. or g/10. 1088/ 1741- Hoebel, K. V., Patel, J. B., Beers, A. L., Chang, K., Singh, P., Brown, J. M., 2560/ 12/1/ 016018 et al. (2021). Radiomics Repeatability Pitfalls in a Scan-Rescan MRI Gemein, L. A. W., Schirrmeister, R. T., Chrabąszcz, P., Wilson, D., Study of Glioblastoma. Radiology Artificial Intelligence, 3,. https:// Boedecker, J., Schulze-Bonhage, A., Hutter, F., & Ball, T. doi. org/ 10. 1148/ ryai. 20201 90199 (2020). Machine-learning-based diagnostics of EEG pathology. Hofmeister, J., Bernava, G., Rosi, A., Vargas, M. I., Carrera, E., Montet, NeuroImage, 220, 117021. https://doi. or g/10. 1016/j. neur oimag e. X., Buergermeister, S., Poletti, P. -A., Platon, A., Lovblad, K -O., 2020. 117021 & Machi, P. (2020). Clot-Based Radiomics Predict a Mechanical Ghassemi, M. M., Moody, B. E., Lehman, L. -W. H., Song, C., Li, Q., Thrombectomy Strategy for Successful Recanalization in Acute Sun, H., Mark, R. G., Westover, M. B., & Clifford, G. D. (2018). Ischemic Stroke. Stroke, 51, 2488–2494. https:// doi. org/ 10. 1161/ You Snooze, You Win: the PhysioNet/Computing in Cardiology STROK EAHA. 120. 030334. Challenge 2018, in: 2018 Computing in Cardiology Conference Hogan, J., Sun, H., Paixao, L., Westmeijer, M., Sikka, P., Jin, J., Tesh, R., (CinC). Presented at the 2018 Computing in Cardiology Confer- Cardoso, M., Cash, S. S., Akeju, O., Thomas, R., & Westover, M. B. ence (CinC), pp. 1–4. https:// doi. org/ 10. 22489/ CinC. 2018. 049 (2021). Night-to-night variability of sleep electroencephalography- based brain age measurements. Clinical Neurophysiology, 132, 1–12. Gjesteby, L., Yang, Q., Xi, Y., Shan, H., Claus, B., Jin, Y., Man, https:// doi. org/ 10. 1016/j. clinph. 2020. 09. 029 B. D., & Wang, G. (2017). Deep learning methods for CT 1 3 Neuroinformatics (2022) 20:943–964 961 Hosseini, M.-P., Hemingway, C., Madamba, J., McKee, A., Ploof, N., Küstner, T., Gatidis, S., Liebgott, A., Schwartz, M., Mauch, L., Martirosian, Schuman, J., & Voss, E. (2020). Review of Machine Learning Algo- P., Schmidt, H., Schwenzer, N. F., Nikolaou, K., Bamberg, F., Yang, rithms for Brain Stroke Diagnosis and Prognosis by EEG Analysis. B., & Schick, F. (2018). A machine-learning framework for auto- ArXiv200808118 Cs Eess. http:// arxiv. org/ abs/ 2008. 08118 matic reference-free quality assessment in MRI. Magnetic Resonance Hu, Z., Jiang, C., Sun, F., Zhang, Q., Ge, Y., Yang, Y., Liu, X., Zheng, Imaging, 53, 134–147. https:// doi. org/ 10. 1016/j. mri. 2018. 07. 003 H., & Liang, D. (2019). Artifact correction in low-dose dental CT LaConte, S. M., Peltier, S. J., & Hu, X. P. (2007). Real-time fMRI imaging using Wasserstein generative adversarial networks. Medi- using brain-state classification. Human Brain Mapping, 28, cal Physics, 46, 1686–1696. https:// doi. org/ 10. 1002/ mp. 13415 1033–1044. https:// doi. org/ 10. 1002/ hbm. 20326 Iglesias, J. E., Billot, B., Balbastre, Y., Tabari, A., Conklin, J., Alexander, Ladefoged, C. N., Marner, L., Hindsholm, A., Law, I., Højgaard, L., & D. C., Golland, P., Edlow, B. L., & Fischl, B. (2020). Joint super- Andersen, F. L. (2018). Deep Learning Based Attenuation Cor- resolution and synthesis of 1 mm isotropic MP-RAGE volumes from rection of PET/MRI in Pediatric Brain Tumor Patients: Evalua- clinical MRI exams with scans of different orientation, resolution and tion in a Clinical Setting. Frontiers in Neuroscience, 12, 1005. contrast. ArXiv201213340 Cs Eess.https:// doi. org/ 10. 3389/ fnins. 2018. 01005 Irwin, M. R. (2019). Sleep and inflammation: Partners in sickness and Lao, J., Chen, Y., Li, Z.-C., Li, Q., Zhang, J., Liu, J., & Zhai, G. (2017). in health. Nature Reviews Immunology, 19, 702–715. https://d oi. A Deep Learning-Based Radiomics Model for Prediction of org/ 10. 1038/ s41577- 019- 0190-z Survival in Glioblastoma Multiforme. Science and Reports, 7, Jelles, B., van Birgelen, J. H., Slaets, J. P. J., Hekster, R. E. M., Jonkman, 10353. https:// doi. org/ 10. 1038/ s41598- 017- 10649-8 E. J., & Stam, C. J. (1999). Decrease of non-linear structure in the Ledoux, L -P., Morency, F. C., Cousineau, M., Houde, J-C., Whittingstall, EEG of Alzheimer patients compared to healthy controls. Clinical K., & Descoteaux, M. (2017). Fiberweb. Diffusion Visualization and Neurophysiology, 110, 1159–1167. https:// doi. org/ 10. 1016/ S1388- Processing in the Browser Frontiers Neuroinformatics, 11. https:// 2457(99) 00013-9doi. org/ 10. 3389/ fninf. 2017. 00054 Jönsson, D., Bergström, A., Forsell, C., Simon, R., Engström, M., Lee, J. M., Akeju, O., Terzakis, K., Pavone, K. J., Deng, H., Houle, Ynnerman, A., & Hotz, I. (2019). A Visual Environment for T. T., Firth, P. G., Shank, E. S., Brown, E. N., & Purdon, P. Hypothesis Formation and Reasoning in Studies with fMRI L. (2017). A Prospective Study of Age-dependent Changes in and Multivariate Clinical Data. The Eurographics Association. Propofol-induced Electroencephalogram Oscillations in Chil- https:// doi. org/ 10. 2312/ vcbm. 20191 232 dren. Anesthesiology, 127, 293–306. https:// doi. org/ 10. 1097/ Kalpathy-Cramer, J., Mamomov, A., Zhao, B., Lu, L., Cherezov, D., ALN. 00000 00000 001717 Napel, S., Echegaray, S., Rubin, D., McNitt-Gray, M., Lo, P., Lehmann, C., Koenig, T., Jelic, V., Prichep, L., John, R. E., Wahlund, L. Sieren, J.C., Uthoff, J., Dilger, S.K.N., Driscoll, B., Yeung, -O., et al. (2007). Application and comparison of classification algo- I., Hadjiiski, L., Cha, K., Balagurunathan, Y., Gillies, R., & rithms for recognition of Alzheimer’s disease in electrical brain activ- Goldgof, D. O (2016). Radiomics of Lung Nodules: A Multi- ity (EEG). Journal of Neuroscience Methods, 161, 342–350. https:// Institutional Study of Robustness and Agreement of Quantita-doi. org/ 10. 1016/j. jneum eth. 2006. 10. 023. tive Imaging Features. Tomogrography Ann Arbor Michigan, Leone, M. J., Sun, H., Boutros, C. L., Liu, L., Ye, E., Sullivan, L., 2(430–437). https:// doi. org/ 10. 18383/j. tom. 2016. 00235 Thomas, R. J., Robbins, G. K., Mukerji, S. S., & Westover, Kamnitsas, K., Baumgartner, C., Ledig, C., Newcombe, V. F. J., Simpson, M. B. (2021). HIV Increases Sleep-based Brain Age Despite J. P., Kane, A. D., Menon, D. K., Nori, A., Criminisi, A., Rueckert, Antiretroviral Therapy. Sleep zsab058. https://d oi.o rg/1 0.1 093/ D., & Glocker, B. (2016). Unsupervised domain adaptation in brain sleep/ zsab0 58 lesion segmentation with adversarial networks. ArXiv161208894 Cs. Liem, F., Varoquaux, G., Kynast, J., Beyer, F., Kharabian Masouleh, http:// arxiv. org/ abs/ 1612. 08894 S., Huntenburg, J. M., Lampe, L., Rahim, M., Abraham, A., Karch, J. D., Filevich, E., Wenger, E., Lisofsky, N., Becker, M., Butler, Craddock, R. C., Riedel-Heller, S., Luck, T., Loeffler, M., O., Mårtensson, J., Lindenberger, U., Brandmaier, A. M., & Kühn, Schroeter, M. L., Witte, A. V., Villringer, A., & Margulies, D. S. (2019). Identifying predictors of within-person variance in MRI- S. (2017). Predicting brain-age from multimodal imaging data based brain volume estimates. NeuroImage, 200, 575–589. https:// captures cognitive impairment. NeuroImage, 148, 179–188. doi. org/ 10. 1016/j. neuro image. 2019. 05. 030https:// doi. org/ 10. 1016/j. neuro image. 2016. 11. 005 Kaufmann, T., van der Meer, D., Doan, N. T., Schwarz, E., Lund, M. J., Liu, F., Jang, H., Kijowski, R., Bradshaw, T., & McMillan, A. B. Agartz, I., Alnæs, D., Barch, D. M., Baur-Streubel, R., Bertolino, (2018a). Deep Learning MR Imaging-based Attenuation A., Bettella, F., Beyer, M. K., Bøen, E., Borgwardt, S., Brandt, C. Correction for PET/MR Imaging. Radiology, 286, 676–684. L., Buitelaar, J., Celius, E. G., Cervenka, S., Conzelmann, A., & https:// doi. org/ 10. 1148/ radiol. 20171 70700 Westlye, L. T. (2019). Common brain disorders are associated with Liu, F., Jang, H., Kijowski, R., Zhao, G., Bradshaw, T., & McMillan, heritable patterns of apparent aging of the brain. Nature Neurosci- A. B. (2018b). A deep learning approach for 18F-FDG PET ence, 22, 1617–1623. https:// doi. org/ 10. 1038/ s41593- 019- 0471-7 attenuation correction. EJNMMI Physics, 5, 24. https://d oi.o rg/ Keshavan, A., Yeatman, J. D., & Rokem, A. (2019). Combining Citizen 10. 1186/ s40658- 018- 0225-8. Science and Deep Learning to Amplify Expertise in Neuroimag- Lorenz, R., Monti, R. P., Violante, I. R., Anagnostopoulos, C., Faisal, ing. Frontiers Neuroinformatics, 13, 29. https://doi. or g/10. 3389/ A. A., Montana, G., & Leech, R. (2016). The Automatic Neu- fninf. 2019. 00029. roscientist: A framework for optimizing experimental design Kniep, H. C., Madesta, F., Schneider, T., Hanning, U., Schönfeld, M. with closed-loop real-time fMRI. NeuroImage, 129, 320–334. H., Schön, G., Fiehler, J., Gauer, T., Werner, R., & Gellissen, S. https:// doi. org/ 10. 1016/j. neuro image. 2016. 01. 032 (2019). Radiomics of Brain MRI: Utility in Prediction of Meta- Luders, E., Cherbuin, N., & Gaser, C. (2016). Estimating brain age static Tumor Type. Radiology, 290, 479–487. https://doi. or g/10. using high-resolution pattern recognition: Younger brains in 1148/ radiol. 20181 80946 long-term meditation practitioners. NeuroImage, 134, 508– Kucyi, A., Esterman, M., Capella, J., Green, A., Uchida, M., Bieder- 513. https:// doi. org/ 10. 1016/j. neuro image. 2016. 04. 007 man, J., Gabrieli, J. D. E., Valera, E. M., & Whitfield-Gabrieli, Lustig, M., Donoho, D. L., Santos, J. M., & Pauly, J. M. (2008). S. (2021). Prediction of stimulus-independent and task-unrelated Compressed Sensing MRI. IEEE Signal Processing Magazine, thought from functional brain networks. Nature Communications, 25, 72–82. https:// doi. org/ 10. 1109/ MSP. 2007. 914728 12, 1793. https:// doi. org/ 10. 1038/ s41467- 021- 22027-0 1 3 962 Neuroinformatics (2022) 20:943–964 Macyszyn, L., Akbari, H., Pisapia, J. M., Da, X., Attiah, M., Pigrish, V., of Aging, 88, 150–155. https://doi. or g/10. 1016/j. neur obiola ging. et al. (2016). Imaging patterns predict patient survival and molecular 2019. 12. 015 subtype in glioblastoma via machine learning techniques. Neuro- Pan, C.-C., Liu, J., Tang, J., Chen, X., Chen, F., Wu, Y.-L., et al. (2019). Oncology, 18, 417–425. https:// doi. org/ 10. 1093/ neuonc/ nov127. A machine learning-based prediction model of H3K27M muta- Manjon, J. V., & Coupe, P. (2019). MRI denoising using Deep Learn- tions in brainstem gliomas using conventional MRI and clinical ing and Non-local averaging. ArXiv191104798 Math. features. Radiotherapy Oncology, 130, 172–179. https://doi. or g/ Marcadent, S., Hofmeister, J., Preti, M. G., Martin, S. P., Van De 10. 1016/j. radonc. 2018. 07. 011. Ville, D., & Montet, X. (2020). Generative Adversarial Net- Parmar, C., Rios Velazquez, E., Leijenaar, R., Jermoumi, M., Carvalho, works Improve the Reproducibility and Discriminative Power S., Mak, R. H., Mitra, S., Shankar, B. U., Kikinis, R., Haibe-Kains, of Radiomic Features. Radiology Artificial Intelligence, 2, B., Lambin, P., & Aerts, H. J. W. L. (2014). Robust Radiomics e190035. https:// doi. org/ 10. 1148/ ryai. 20201 90035 feature quantification using semiautomatic volumetric segmenta- Mateos-Pérez, J. M., Dadar, M., Lacalle-Aurioles, M., Iturria- tion. PloS One, 9, e102107. https:// doi. org/ 10. 1371/ journ al. pone. Medina, Y., Zeighami, Y., & Evans, A. C. (2018). Structural 01021 07 neuroimaging as clinical predictor: A review of machine learn- Phang, C. -R., Noman, F., Hussain, H., Ting, C. -M., & Ombao, H. (2020). ing applications. NeuroImage Clinical, 20, 506–522. https:// A Multi-Domain Connectome Convolutional Neural Network for doi. org/ 10. 1016/j. nicl. 2018. 08. 019 Identifying Schizophrenia From EEG Connectivity Patterns. IEEE Merikanto, I., Utge, S., Lahti, J., Kuula, L., Makkonen, T., Lahti‐ Journal of Biomedical and Health Informatics, 24, 1333–1343. Pulkkinen, M., & Pesonen, A. K. (2019). Genetic risk fac-https:// doi. org/ 10. 1109/ JBHI. 2019. 29412 22. tors for schizophrenia associate with sleep spindle activity in Pinto, A. L. R., Ou, Y., Sahin, M., & Grant, P. E. (2018). Quantitative healthy adolescents. Journal of Sleep Research, 28 https://doi. Apparent Diffusion Coefficient Mapping May Predict Seizure org/ 10. 1111/ jsr. 12762 Onset in Children With Sturge-Weber Syndrome. Pediatric Michel, C. M., & Murray, M. M. (2012). Towards the utilization of EEG Neurology, 84, 32–38. https:// doi. org/ 10. 1016/j. pedia trneu rol. as a brain imaging tool. NeuroImage, 61, 371–385. https:// doi. org/ 2018. 04. 004 10. 1016/j. neuro image. 2011. 12. 039 Pipe, J. G. (1999). Motion correction with PROPELLER MRI: Applica- Miranda, P., D Cox, C., Alexander, M., Danev, S., & R. T., Lakey, tion to head motion and free-breathing cardiac imaging. Magnetic J., (2019). Overview of current diagnostic, prognostic, and Resonance in Medicine, 42, 963–969. https://doi. or g/10. 1002/ (sici) therapeutic use of EEG and EEG-based markers of cognition, 1522- 2594(199911) 42:5% 3c963:: aid- mrm17% 3e3.0. co;2-l mental, and brain health. Integrative Molecular. Medicine, 6. Pizarro, R. A., Cheng, X., Barnett, A., Lemaitre, H., Verchinski, B. https:// doi. org/ 10. 15761/ IMM. 10003 78 A., Goldman, A. L., Xiao, E., Luo, Q., Berman, K. F., Callicott, Mohajer, B., Abbasi, N., Mohammadi, E., Khazaie, H., Osorio, R. S., J. H., Weinberger, D. R., & Mattay, V. S. (2016). Automated Rosenzweig, I., Eickhoff, C. R., Zarei, M., Tahmasian, M., & Quality Assessment of Structural Magnetic Resonance Brain Eickhoff, S. B. (2020). Gray matter volume and estimated brain Images Based on a Supervised Machine Learning Algorithm. age gap are not linked with sleep-disordered breathing. Human Front. Neuroinformatics, 10, 52. https:// doi. org/ 10. 3389/ fninf. Brain Mapping, 41, 3034–3044. https:// doi. org/ 10. 1002/ hbm. 2016. 00052 24995 Poddar, J., Pradhan, M., Ganguly, G., & Chakrabarti, S. (2019). Bio- Moyer, D., Ver Steeg, G., Tax, C. M. W., & Thompson, P. M. (2020). chemical deficits and cognitive decline in brain aging: Interven- Scanner invariant representations for diffusion MRI harmoniza- tion by dietary supplements. Journal of Chemical Neuroanatomy, tion. Magnetic Resonance in Medicine, 84, 2174–2189. https:// 95, 70–80. https:// doi. org/ 10. 1016/j. jchem neu. 2018. 04. 002 doi. org/ 10. 1002/ mrm. 28243 Provenzale, J. M., Ison, C., & Delong, D. (2009). Bidimensional meas- Ning, K., Zhao, L., Matloff, W., Sun, F., & Toga, A. W. (2020). Asso- urements in brain tumors: Assessment of interobserver variabil- ciation of relative brain age with tobacco smoking, alcohol con- ity. AJR. American Journal of Roentgenology, 193, W515-522. sumption, and genetic variants. Science and Reports, 10, 10. https:// doi. org/ 10. 2214/ AJR. 09. 2615 https:// doi. org/ 10. 1038/ s41598- 019- 56089-4 Provenzale, J. M., & Mancini, M. C. (2012). Assessment of intra- Nishimura, D. G. (2010). Principles of magnetic resonance imaging. observer variability in measurement of high-grade brain tumors. Self-Published. Journal of Neuro-Oncology, 108, 477–483. https:// doi. org/ 10. O’Muircheartaigh, J., Robinson, E. C., Pietsch, M., Wolfers, T., Aljabar, 1007/ s11060- 012- 0843-2 P., Grande, L. C., Teixeira, R. P. A. G., Bozek, J., Schuh, A., Pruessmann, K. P., Weiger, M., Scheidegger, M. B., & Boesiger, P. Makropoulos, A., Batalle, D., Hutter, J., Vecchiato, K., Steinweg, J. (1999). SENSE: Sensitivity encoding for fast MRI. Magnetic K., Fitzgibbon, S., Hughes, E., Price, A. N., Marquand, A., Reuckert, Resonance in Medicine, 42, 952–962. D., & Edwards, A. D. (2020). Modelling brain development to detect Purdon, P. L., Pavone, K. J., Akeju, O., Smith, A. C., Sampson, A. L., white matter injury in term and preterm born neonates. Brain, 143, Lee, J., Zhou, D. W., Solt, K., & Brown, E. N. (2015). The Ageing 467–479. https:// doi. org/ 10. 1093/ brain/ awz412 Brain: Age-dependent changes in the electroencephalogram dur- Orlhac, F., Boughdad, S., Philippe, C., Stalla-Bourdillon, H., Nioche, ing propofol and sevoflurane general anaesthesia. British Journal C., Champion, L., Soussan, M., Frouin, F., Frouin, V., & Buvat, I. of Anaesthesia, 115, i46–i57. https:// doi. org/ 10. 1093/ bja/ aev213 (2018). A Postreconstruction Harmonization Method for Multicenter Putzky, P., Karkalousos, D., Teuwen, J., Miriakov, N., Bakker, B., Radiomic Studies in PET. Journal of Nuclear Medicine, 59, 1321– Caan, M., & Welling, M. (2019). i-RIM applied to the fastMRI 1328. https:// doi. org/ 10. 2967/ jnumed. 117. 199935 challenge. ArXiv191008952. Ou, Y., Zöllei, L., Retzepi, K., Castro, V., Bates, S. V., Pieper, S., Quan, T. M., Nguyen-Duc, T., & Jeong, W.-K. (2018). Compressed Andriole, K. P., Murphy, S. N., Gollub, R. L., & Grant, P. E. Sensing MRI Reconstruction Using a Generative Adversarial Net- (2017). Using clinically acquired MRI to construct age-specific work With a Cyclic Loss. IEEE Transactions on Medical Imaging, ADC atlases: Quantifying spatiotemporal ADC changes from 37, 1488–1497. https:// doi. org/ 10. 1109/ TMI. 2018. 28201 20 birth to 6-year old. Human Brain Mapping, 38, 3052–3068. Ramm, A. G., & Katsevich, A. I. (1996). The Radon Transform and https:// doi. org/ 10. 1002/ hbm. 23573 Local Tomography. CRC Press. Paixao, L., Sikka, P., Sun, H., Jain, A., Hogan, J., Thomas, R., & Rauschecker, A. M., Rudie, J. D., Xie, L., Wang, J., Duong, M. T., Botzolakis, E. J., Kovalovich, A. M., Egan, J., Cook, T. C., Bryan, Westover, M. B. (2020). Excess brain age in the sleep electro- R. N., Nasrallah, I. M., Mohan, S., & Gee, J. C. (2020). Artificial encephalogram predicts reduced life expectancy. Neurobiology 1 3 Neuroinformatics (2022) 20:943–964 963 Intelligence System Approaching Neuroradiologist-level Differen- Sun, H., Paixao, L., Oliva, J. T., Goparaju, B., Carvalho, D. Z., van tial Diagnosis Accuracy at Brain MRI. Radiology, 295, 626–637. Leeuwen, K. G., Akeju, O., Thomas, R. J., Cash, S. S., Bianchi, https:// doi. org/ 10. 1148/ radiol. 20201 90283 M. T., & Westover, M. B. (2019). Brain age from the electro- Regenhardt, R. W., Bretzner, M., Zanon Zotin, M. C., Bonkhoff, A. K., encephalogram of sleep. Neurobiology of Aging, 74, 112–120. Etherton, M. R., Hong, S., Das, A. S., Alotaibi, N. M., Vranic, J. https:// doi. org/ 10. 1016/j. neuro biola ging. 2018. 10. 016 E., Dmytriw, A. A., Stapleton, C. J., Patel, A. B., Kuchcinski, G., Tang, T., Jiao, Y., Cui, Y., Zhao, D., Zhang, Y., Wang, Z., Meng, Rost, N. S., & Leslie-Mazwi, T. M. (2021). Radiomic signature of X., Yin, X.-D., Yang, Y.-J., Teng, G., & Ju, S. (2020). Penum- DWI-FLAIR mismatch in large vessel occlusion stroke. Journal of bra-based radiomics signature as prognostic biomarkers for Neuroimaging. https:// doi. org/ 10. 1111/ jon. 12928 thrombolysis of acute ischemic stroke patients: A multicenter Rogenmoser, L., Kernbach, J., Schlaug, G., & Gaser, C. (2018). Keep- cohort study. Journal of Neurology, 267, 1454–1463. https:// ing brains young with making music. Brain Structure & Func-doi. org/ 10. 1007/ s00415- 020- 09713-7 tion, 223, 297–305. https:// doi. org/ 10. 1007/ s00429- 017- 1491-2 Tanioka, S., Ishida, F., Yamamoto, A., Shimizu, S., Sakaida, H., Toyoda, Roy, S., Kiral-Kornek, I., & Harrer, S. (2019a). ChronoNet: A Deep M., et al. (2020). Machine Learning Classification of Cerebral Recurrent Neural Network for Abnormal EEG Identification, in: Aneurysm Rupture Status with Morphologic Variables and Hemo- Riaño, D., Wilk, S., ten Teije, A. (Eds.), Artificial Intelligence in dynamic Parameters. Radiology Artificial Intelligence, 2, e190077. Medicine, Lecture Notes in Computer Science. Springer Inter-https:// doi. org/ 10. 1148/ ryai. 20191 90077 national Publishing, Cham, pp. 47–56. https:// doi. org/ 10. 1007/ Temple University EEG Corpus (2021). https://www .isip. picon epr ess.com/ 978-3- 030- 21642-9_8proje cts/ tuh_ eeg/ html/ downl oads. shtml (Accessed 29 Apr 2021). Roy, Y., Banville, H., Albuquerque, I., Gramfort, A., Falk, T. H., & Titano, J. J., Badgeley, M., Schefflein, J., Pain, M., Su, A., Cai, M., Swin- Faubert, J. (2019b). Deep learning-based electroencephalog- burne, N., Zech, J., Kim, J., Bederson, J., Mocco, J., Drayer, B., raphy analysis: A systematic review. Journal of Neural Engi- Lehar, J., Cho, S., Costa, A., & Oermann, E. K. (2018). Automated neering, 16, 051001. https://doi. or g/10. 1088/ 1741- 2552/ ab260c deep-neural-network surveillance of cranial images for acute neu- Schlemper, J., Caballero, J., Hajnal, J.V., Price, A., & Rueckert, D. rologic events. Nature Medicine, 24, 1337–1341. https:// doi. org/ 10. (2018). A Deep Cascade of Convolutional Neural Networks 1038/ s41591- 018- 0147-y for Dynamic MR Image Reconstruction. IEEE Translational Tzimourta, K. D., Christou, V., Tzallas, A. T., Giannakeas, N., Astrakas, Medicine Imaging 491–503. L. G., Angelidis, P., & Tsipouras, M. G. (2021). Machine Learn- Schwier, M., van Griethuysen, J., Vangel, M. G., Pieper, S., Peled, S., ing Algorithms and Statistical Approaches for Alzheimer’s Disease Tempany, C., Aerts, H. J. W. L., Kikinis, R., Fennessy, F. M., & Analysis Based on Resting-State EEG Recordings. A Systematic Fedorov, A. (2019). Repeatability of Multiparametric Prostate Review International Journal of Neural Systems, 2130002. https:// MRI Radiomics Features. Science and Reports, 9, 9441. https:// doi. org/ 10. 1142/ S0129 06572 13000 23 doi. org/ 10. 1038/ s41598- 019- 45766-z van Horn, N., Kniep, H., Broocks, G., Meyer, L., Flottmann, F., Bechstein, Si, Y. (2020). Machine learning applications for electroencephalograph M., Götz, J., Thomalla, G., Bendszus, M., Bonekamp, S., Pfaff, J. signals in epilepsy: A quick review. Acta Epileptologica, 2, 5. A. R., Dellani, P. R., Fiehler, J., & Hanning, U. (2021). ASPECTS https:// doi. org/ 10. 1186/ s42494- 020- 00014-0. Interobserver Agreement of 100 Investigators from the TEN- Singh, N. M., Iglesias, J. E., Adalsteinsson, E., Dalca, A. V., & Golland, SION. Study Clinical Neuroradiology. https:// doi. org/ 10. 1007/ P. (2020). Joint Frequency and Image Space Learning for Fourier s00062- 020- 00988-x Imaging. ArXiv200701441 Cs Eess. van der Maaten, L., & Hinton, G. (2008). Visualizing Data using t-SNE. Sleep Data - National Sleep Research Resource – NSRR (2021). https:// Journal of Machine Learning Research, 9, 2579–2605. sleep data. org/ (Accessed Apr 29 2021). van Leeuwen, K. G., Sun, H., Tabaeizadeh, M., Struck, A. F., van Putten, Sørensen, L., Igel, C., Liv Hansen, N., Osler, M., Lauritzen, M., M. J. A. M., & Westover, M. B. (2019). Detecting abnormal electro- Rostrup, E., Nielsen, M., Alzheimer’s Disease Neuroimaging encephalograms using deep convolutional networks. Clinical Neuro- Initiative and the Australian Imaging Biomarkers and Lifestyle physiology, 130, 77–84. https:// doi. org/ 10. 1016/j. clinph. 2018. 10. 012 Flagship Study of Ageing. (2016). Early detection of Alzhei- Varikuti, D. P., Genon, S., Sotiras, A., Schwender, H., Hoffstaedter, F., Patil, mer’s disease using MRI hippocampal texture. Human Brain K. R., Jockwitz, C., Caspers, S., Moebus, S., Amunts, K., Davatzikos, Mapping, 37, 1148–1161. https:// doi. org/ 10. 1002/ hbm. 23091 C., & Eickhoff, S. B. (2018). Evaluation of non-negative matrix fac- Sotardi, S., Gollub, R. L., Bates, S. V., Weiss, R., Murphy, S. N., Grant, torization of grey matter in age prediction. NeuroImage, 173, 394– P. E., & Ou, Y. (2021). Voxelwise and Regional Brain Apparent 410. https:// doi. org/ 10. 1016/j. neuro image. 2018. 03. 007 Diffusion Coefficient Changes on MRI from Birth to 6 Years of Vieira, S., Pinaya, W. H. L., & Mechelli, A. (2017). Using deep learning Age. Radiology, 298, 415–424. https:// doi. org/ 10. 1148/ r adiol. to investigate the neuroimaging correlates of psychiatric and neu- 20202 02279 rological disorders: Methods and applications. Neuroscience and Stefano, A., Comelli, A., Bravatà, V., Barone, S., Daskalovski, I., Savoca, Biobehavioral Reviews, 74, 58–75. https:// doi. org/ 10. 1016/j. neubi G., Sabini, M. G., Ippolito, M., & Russo, G. (2020). A preliminary orev. 2017. 01. 002 PET radiomics study of brain metastases using a fully automatic seg- Vos, T., Lim, S. S., Abbafati, C., Abbas, K. M., Abbasi, M., Abbasi- mentation method. BMC Bioinformatics, 21, 325. https://d oi.o rg/1 0. fard, M., Abbasi-Kangevari, M., Abbastabar, H., Abd-Allah, F., 1186/ s12859- 020- 03647-7 Abdelalim, A., Abdollahi, M., Abdollahpour, I., Abolhassani, H., Steffener, J., Habeck, C., O’Shea, D., Razlighi, Q., Bherer, L., & Stern, Y. Aboyans, V., Abrams, E. M., Abreu, L. G., Abrigo, M. R. M., (2016). Differences between chronological and brain age are related Abu-Raddad, L. J., Abushouk, A. I., & Murray, C. J. L. (2020). to education and self-reported physical activity. Neurobiology of Global burden of 369 diseases and injuries in 204 countries and Aging, 40, 138–144. https://doi. or g/10. 1016/j. neur obiola ging. 2016. territories, 1990–2019: A systematic analysis for the Global Bur- 01. 014 den of Disease Study 2019. The Lancet, 396, 1204–1222. https:// Stoeckel, L. E., Garrison, K. A., Ghosh, S. S., Wighton, P., Hanlon, doi. org/ 10. 1016/ S0140- 6736(20) 30925-9 C. A., Gilman, J. M., et al. (2014). Optimizing real time fMRI Wang, G., Luo, T., Nielsen, J.-F., Noll, D. C., & Fessler, J. A. (2021). neurofeedback for therapeutic discovery and development. Neu- B-spline Parameterized Joint Optimization of Reconstruction roImage Clinical, 5, 245–255. https:// doi. or g/ 10. 1016/j. nicl. and K-space Trajectories (BJORK) for Accelerated 2D MRI. 2014. 07. 002. ArXiv210111369. 1 3 964 Neuroinformatics (2022) 20:943–964 Weiss, T., Senouf, O., Vedula, S., Michailovich, O., Zibulevsky, M., Network Open, 3, e2017357–e2017357. https://d oi.o rg/1 0.1 001/ & Bronstein, A. (2021). PILOT: Physics-Informed Learned jaman etwor kopen. 2020. 17357 Optimized Trajectories for Accelerated MRI. ArXiv190905773 Zbontar, J., Knoll, F., Sriram, A., Murrell, T., Huang, Z., Muckley, Physics. M. J., Defazio, A., Stern, R., Johnson, P., Bruno, M., Parente, Whitfield-Gabrieli, S., Ghosh, S. S., Nieto-Castanon, A., Saygin, Z., M., Geras, K. J., Katsnelson, J., Chandarana, H., Zhang, Z., Doehrmann, O., Chai, X. J., Reynolds, G. O., Hofmann, S. G., Drozdzal, M., Romero, A., Rabbat, M., Vincent, P., & Lui, Y. W. Pollack, M. H., & Gabrieli, J. D. E. (2016). Brain connectomics (2019). fastMRI: An Open Dataset and Benchmarks for Acceler- predict response to treatment in social anxiety disorder. Molecu- ated MRI. ArXiv181108839 Physics Statistics. lar Psychiatry, 21, 680–685. https:// doi. org/ 10. 1038/ mp. 2015. Zech, J. R., Badgeley, M. A., Liu, M., Costa, A. B., Titano, J. J., & 109 Oermann, E. K. (2018). Variable generalization performance of Whitfield-Gabr ieli, S., Wendelken, C., Nieto-Castañón, A., Bailey, S. a deep learning model to detect pneumonia in chest radiographs: K., Anteraper, S. A., Lee, Y. J., Chai, X.-Q., Hirshfeld-Becker, D. A cross-sectional study. PLoS Medicine, 15, e1002683. https:// R., Biederman, J., Cutting, L. E., & Bunge, S. A. (2020). Associa-doi. org/ 10. 1371/ journ al. pmed. 10026 83 tion of Intrinsic Brain Architecture With Changes in Attentional Zhang, X., Braun, U., Tost, H., & Bassett, D. S. (2020). Data-Driven and Mood Symptoms During Development. JAMA Psychiatry, 77, Approaches to Neuroimaging Analysis to Enhance Psychiatric 378–386. https:// doi. org/ 10. 1001/ jamap sychi atry. 2019. 4208 Diagnosis and Therapy. Biology Psychiatry Cognition Neurosci- Wijaya, S. K., Badri, C., Misbach, J., Soemardi, T. P., & Sutanno, V. ence Neuroimaging, 5, 780–790. https:// doi. org/ 10. 1016/j. bpsc. (2015). Electroencephalography (EEG) for detecting acute ischemic 2019. 12. 015. stroke, in: 2015 4th International Conference on Instrumentation, Zhang, Z., Romero, A., Muckley, M. J., Vincent, P., Yang, L., & Drozdzal, Communications, Information Technology, and Biomedical Engi- M. (2019) Reducing Uncertainty in Undersampled MRI Reconstruc- neering (ICICI-BME). Presented at the 2015 4th International tion with Active Acquisition. ArXiv190203051 Cs. Conference on Instrumentation, Communications, Information Zhou, H., Hu, R., Tang, O., Hu, C., Tang, L., Chang, K., Shen, Q., Technology, and Biomedical Engineering (ICICI-BME), pp. 42–48. Wu, J., Zou, B., Xiao, B., Boxerman, J., Chen, W., Huang, R. https:// doi. org/ 10. 1109/ ICICI- BME. 2015. 74013 12 Y., Yang, L., Bai, H. X., & Zhu, C. (2020). Automatic Machine Woon, W. L., Cichocki, A., Vialatte, F., & Musha, T. (2007). Tech- Learning to Differentiate Pediatric Posterior Fossa Tumors on niques for early detection of Alzheimer’s disease using spontane- Routine MR Imaging. American Journal of Neuroradiology, 41, ous EEG recordings. Physiological Measurement, 28, 335–347. 1279–1285. https:// doi. org/ 10. 3174/ ajnr. A6621 https:// doi. org/ 10. 1088/ 0967- 3334/ 28/4/ 001 Zhou, M., Scott, J., Chaudhury, B., Hall, L., Goldgof, D., Yeom, K. Xiao, T., Hua, W., Li, C., & Wang, S. (2019). Glioma Grading Pre- W., Iv, M., Ou, Y., Kalpathy-Cramer, J., Napel, S., Gillies, R., diction by Exploring Radiomics and Deep Learning Features, Gevaert, O., & Gatenby, R. (2018). Radiomics in Brain Tumor: in: Proceedings of the Third International Symposium on Image Image Assessment, Quantitative Feature Descriptors, and Computing and Digital Medicine, ISICDM 2019. Association Machine-Learning Approaches. American Journal of Neurora- for Computing Machinery, New York, NY, USA, pp. 208–213. diology, 39, 208–216. https:// doi. org/ 10. 3174/ ajnr. A5391 https:// doi. org/ 10. 1145/ 33648 36. 33648 77 Zwanenburg, A., Vallières, M., Abdalah, M. A., Aerts, H. J. W. L., Xu, J., Gong, E., Pauly, J., & Zaharchuk, G. (2017). 200x Low-dose Andrearczyk, V., Apte, A., Ashrafinia, S., Bakas, S., Beukinga, R. PET Reconstruction using Deep Learning. ArXiv171204119 Cs. J., Boellaard, R., Bogowicz, M., Boldrini, L., Buvat, I., Cook, G. Yang, G., Yu, S., Dong, H., Slabaugh, G., Dragotti, P. L., Ye, X., J. R., Davatzikos, C., Depeursinge, A., Desseroit, M.-C., Dinapoli, Liu, F., Arridge, S., Keegan, J., Guo, Y., & Firmin, D. (2018). N., Dinh, C. V., & Löck, S. (2020). The Image Biomarker Stand- DAGAN: Deep De-Aliasing Generative Adversarial Networks ardization Initiative: Standardized Quantitative Radiomics for High- for Fast Compressed Sensing MRI Reconstruction. IEEE Trans- Throughput Image-based Phenotyping. Radiology, 295, 328–338. actions on Medical Imaging, 37, 1310–1321. https:// doi. org/ 10. https:// doi. org/ 10. 1148/ radiol. 20201 91145 1109/ TMI. 2017. 27858 79 Ye, E., Sun, H., Leone, M. J., Paixao, L., Thomas, R. J., Lam, A. Publisher's Note Springer Nature remains neutral with regard to D., & Westover, M. B. (2020). Association of Sleep Electroen- jurisdictional claims in published maps and institutional affiliations. cephalography-Based Brain Age Index With Dementia. JAMA 1 3 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Neuroinformatics Springer Journals

How Machine Learning is Powering Neuroimaging to Improve Brain Health

Download PDF
22 pages

Loading next page...
 
/lp/springer-journals/how-machine-learning-is-powering-neuroimaging-to-improve-brain-health-7pWYBnTkBQ

References (243)

Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2022
ISSN
1539-2791
eISSN
1559-0089
DOI
10.1007/s12021-022-09572-9
Publisher site
See Article on Publisher Site

Abstract

This report presents an overview of how machine learning is rapidly advancing clinical translational imaging in ways that will aid in the early detection, prediction, and treatment of diseases that threaten brain health. Towards this goal, we aresharing the information presented at a symposium, “Neuroimaging Indicators of Brain Structure and Function - Closing the Gap Between Research and Clinical Application”, co-hosted by the McCance Center for Brain Health at Mass General Hospital and the MIT HST Neuroimaging Training Program on February 12, 2021. The symposium focused on the potential for machine learning approaches, applied to increasingly large-scale neuroimaging datasets, to transform healthcare delivery and change the trajec- tory of brain health by addressing brain care earlier in the lifespan. While not exhaustive, this overview uniquely addresses many of the technical challenges from image formation, to analysis and visualization, to synthesis and incorporation into the clinical workflow. Some of the ethical challenges inherent to this work are also explored, as are some of the regulatory requirements for implementation. We seek to educate, motivate, and inspire graduate students, postdoctoral fellows, and early career investigators to contribute to a future where neuroimaging meaningfully contributes to the maintenance of brain health. Keywords Machine learning · Deep learning · Clinical translational neuroimaging · Brain health · MRI · PET · EEG · Transcranial magnetic stimulation * Randy L. Gollub Computer Science and Artificial Intelligence Laboratory, rgollub@partners.org Massachusetts Institute of Technology, Cambridge, MA 02139, USA Department of Psychiatry and Martinos Center Institute of Systems Neuroscience, Medical Faculty, Heinrich for Biomedical Imaging, Department of Radiology, Heine University Düsseldorf, Düsseldorf, Germany Massachusetts General Hospital, Boston, MA 02114, USA Institute of Neuroscience and Medicine, Brain & Behaviour Harvard-MIT Health Sciences and Technology, (INM-7) Research Centre Jülich, Jülich, Germany Massachusetts Institute of Technology, Cambridge, MA 02139, USA Department of Radiology, Brigham and Women’s Hospital and Harvard Medical School, MA 02115 Boston, USA University of Massachusetts Boston, Boston, MA 02125, USA Department of Psychology, Northeastern University, Boston 02115, USA Department of Radiology, Boston Children’s Hospital, Harvard Medical School, Boston, MA 02115, USA Martinos, Radiology, MGH, MIT, HMS & EECS, Cambridge 02114, USA Department of Neurology and McCance Center for Brain Health / Harvard Medical School, Massachusetts General Department of Psychiatry, Brigham and Women’s Hospital Hospital, Boston 02114, USA and Harvard Medical School, Boston 02115, USA 6 15 Centre for Medical Image Computing, University College Center for Brain Circuit Therapeutics, Department London, London, UK of Neurology, Psychiatry, and Radiology, Brigham and Women’s Hospital and Harvard Medical School, Martinos Center for Biomedical Imaging, Department 02115 Boston, USA of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston 02114, USA Vol.:(0123456789) 1 3 944 Neuroinformatics (2022) 20:943–964 set of important questions that we believe researchers should Introduction bear in mind when working in this area. Machine learning models vary in the amount of domain Machine learning is contributing to rapid advances in clinical knowledge they incorporate and how they do so. Some mod- translational imaging to enable early detection, prediction, els explicitly enforce that their outputs are consistent with and treatment of diseases that threaten brain health. Brain the physics of an imaging or measurement process. Other diseases, including cerebrovascular disease, depression, methods act upon features explicitly chosen because they migraine headaches, and dementia, are leading causes of are known to be relevant for the task at hand. These feature- global disability (Vos et al., 2020). Continued progress in based methods are often applied to established clinical use neuroimaging and machine learning, and the collection cases to reduce time, manual labor, and/or person-to-person of increasingly large-scale data sets, promise to transform variation (Gajawelli et al., 2019). "End-to-end" approaches healthcare by providing non-invasive, reliable indicators of abstract out explicit feature definition to go from raw data brain health, resilience, and vulnerability long before clinical all the way to interpretable quantitative metrics, for exam- manifestations of disease. But many technical challenges ple, of brain health. All traditional step-by-step processes, remain. On February 12th 2021, the MGH McCance Center such as artifact removal, registration, conversion between for Brain Health at Mass General Hospital, together with the temporal and spectral domains, and feature extraction could Harvard-MIT Health Sciences and Technology Neuroim- be encompassed in a single machine learning pipeline. aging Training Program, co-hosted a virtual symposium, In general, models with more explicitly encoded domain “Neuroimaging Indicators of Brain Structure and Function— knowledge are less flexible in adapting to cases where the Closing the Gap Between Research and Clinical Application,” measurement process may be inaccurately characterized. to highlight some of these remaining challenges and machine That said, these methods are able to incorporate known learning approaches to overcome them. Recorded videos of the relationships, which can guide the learning process and symposium presentations and discussions are available link prevent nonsensical results. The methods covered in this to YouTube videos https:// www. youtu be. com/ playl ist? list= report represent examples from each of these categories as PL0A- NKHLV rNF82 vdjey yaBRo iXg77 lCeW. well as intermediate cases which explicitly incorporate some In this symposium report, we explore a spectrum of domain knowledge but also allow the model significant flex- machine learning applications in neuroimaging and use sym- ibility in producing unconstrained final outputs. posium presentations to illustrate key points. We cover both It is important to acknowledge that the field is dynamic recent advances and outstanding challenges, beginning with with each area undergoing rapid transitions. In some cases, image acquisition, and ending with computation of quantita- machine learning techniques are already deployed or in tive metrics and initial clinical utilization (Fig. 1): advanced stages of testing for deployment into existing clinical workflows. In other cases, efforts are focused on Section I describes how machine learning improves volu- early research, with the goal of discovering scientific insight metric image acquisition and reconstruction. or extracting meaningful features from the images that can Section II  describes machine learning approaches to be fed into machine learning methods to generate clinically image processing, focusing on image harmonization, and meaningful biomarkers (FDA-NIH Biomarker Working methods to detect deviations from healthy brain structure Group, 2016; Mateos-Pérez et al., 2018). In all presenta- and function. tions and throughout this report, our goal is to educate, Section III describes machine learning advances in inter- motivate and inspire graduate students, post-doctoral fel- pretation and analysis of non-volumetric EEG data. lows, and early career investigators to contribute to a future Section IV describes multiple approaches to assess brain where imaging meaningfully contributes to the maintenance health using deviations from healthy aging. of brain health. Section V describes how machine learning techniques can be applied to individual patient imaging, and other diagnostics, to personalize medical treatments that Section I: Machine Learning for Improved improve brain health. Volumetric Image Acquisition and Reconstruction Finally, Section VI explores the implications of deploy- ing neuroimaging indicators of brain health into a clinical Machine learning techniques can be used to improve one of workflow. In particular, we focus on regulatory approval the earliest steps in the neuroimaging pipeline even before pathways of machine learning algorithms and the ethical an image is viewed by a clinician or researcher: image considerations involved in collecting, algorithmically ana- formation. Typically, scanner-acquired measurements lyzing, and acting upon the derived information. We raise a 1 3 Neuroinformatics (2022) 20:943–964 945 Fig. 1 Schematic illustration of the spectrum of machine learn- summaries to metrics used to quantify brain health, such as the vol- ing applications in clinical translational neuroimaging. A typical ume of a structure or the estimated age of a subject. Finally, these volumetric neuroimaging workflow is shown for MRI. A patient quantitative metrics, once comprehensively validated, can be used is scanned, creating a signal (i.e. k-space data) which is converted to inform patient care through early detection of subtle abnormali- to an image via a reconstruction algorithm in preparation for clini- ties and to guide treatments such as targeted brain stimulation. These cal review by a radiologist. In some cases, the reconstructed image steps are not just useful at the individual patient level but can also undergoes further computational processing to produce higher-level drive population level analyses that can lead to insights regarding summaries, such as segmentations or registrations to an atlas. Option- healthy and disordered brain structure and function ally, in the future, further computational processing can convert these represent an encoding of the patient anatomy under the Once all scanner data is acquired, machine learning is physics governing the imaging system. For example, the useful in reconstructing the image of interest itself. Model- measurements acquired from an MRI scanner represent based optimization techniques for MRI, CT, and PET imaging the Fourier transform of the image of interest (Nishimura, have typically provided iteratively refined solutions to under- 2010), while the measurements acquired from PET and CT determined inverse problems (Griswold et al., 2002; Lustig scanners represent the Radon transform of the image of et al., 2008; Pruessmann et al., 1999). Recently, deep-learning interest (Ramm & Katsevich, 1996). Recovering the under- methods quickly estimate solutions to the inverse imaging lying image from the acquired scanner data requires solv- problems, including network architectures that explicitly employ ing an inverse problem. Due to time, patient comfort and the physics of the imaging system (Hammernik et al., 2018; safety considerations, or monetary constraints, often only Putzky et al., 2019; Schlemper et al., 2018). Efforts to collect a limited number of scanner measurements are acquired, large-scale public datasets of raw imaging data have accelerated making this inverse problem highly under-determined. advances in image reconstruction by enabling rapid model Further, the acquired signals may be corrupted by imper- prototyping and by simplifying and standardizing evaluation fections in the imaging process, such as patient motion or of varying approaches. For example, the FastMRI dataset system noise. Techniques from machine learning, includ- provides publicly available k-space data for reconstructing over ing (1) model-based optimization methods (Griswold six thousand human brain MRIs (Zbontar et al., 2019). et al., 2002; Lustig et al., 2008; Pruessmann et al., 1999), Beyond accelerated reconstruction from limited measurements, (2) data-driven learning methods (Quan et al., 2018; Yang several machine learning approaches have been proposed et  al., 2018), and (3) combinations of these two strate- for correcting artifacts arising during image acquisition and gies (Hammernik et al., 2018; Schlemper et al., 2018), are reconstruction. For example, both optimization and learning- promising approaches for mitigating these image forma- based approaches have been proposed for MRI denoising tion issues, enabling faster, higher-quality image creation (Anand & Sahambi, 2010; Manjon & Coupe, 2019) and motion for downstream analysis. correction (Chun et al., 2012; Haskell et al., 2018; Pipe, 1999). One such application of machine learning involves deciding At this symposium, Ms. Nalini Singh presented two neural which exact scanner measurements to acquire. Given time or network layer structures which can be used to build networks financial imaging budget constraints, machine learning can be which correct each of these artifacts while also being used for used to identify which subset of measurements will be the most accelerated reconstruction (Singh et al., 2020). Unlike many other informative for reconstructing the final image. For example, reconstruction methods, these layers incorporate convolutions on several approaches have been proposed for learning the optimal both the frequency space and image space features. By operating k-space acquisition pattern for a specified class of MRI scans, in both spaces, these layers both correct artifacts native to the often identifying different patterns for different anatomies frequency space and manipulate image space representations (Bahadir et al., 2020; Wang et al., 2021; Weiss et al., 2021). to form coherent image structures. Figure 2 shows a detailed For neuroimaging applications in particular, sampling trajec- diagram of the layers, and Fig. 3 shows example reconstructions tories could be optimized for specific structures of interest for demonstrating the positive impact of this method on the quality the clinical question being asked. More recent work aims to of the reconstructed image in representing the true brain anatomy. optimize the acquisition pattern with even greater specificity Several deep-learning based approaches have also been for each individual patient (Zhang et al., 2019). proposed for metal artifact reduction in CT imaging of other 1 3 946 Neuroinformatics (2022) 20:943–964 Fig. 2 Two joint layer architectures combining frequency and image resents frequency space features at the nth  layer, and v represents space representations, embedded within full network architectures image space features at the nth  layer. At each layer, Batch Normali- for MRI reconstruction. Red squares represent frequency space quan- zation (BN), a convolution, and an activation function are applied to tities, while blue squares represent image space quantities. u rep-both u and v , summarized by `F-Conv' or `I-Conv', respectively n n n anatomies (Gjesteby et al., 2017; Hu et al., 2019); these tech- While machine learning-based approaches promise to niques could be extended to improve brain CT imaging for improve the speed, value, and quality of brain image acquisition, patients with deep brain stimulation (DBS) devices in situ. In several challenges must be solved before they are incorporated PET imaging, deep-learning based approaches have demon- into standard clinical worko fl ws. For example, many current strated improved correction of attenuation effects both with reconstruction methods require large datasets of thousands of (Ladefoged et al., 2018; Liu et al., 2018a) and without (Dong high-quality acquired signals from a particular imaging proto- et al., 2020; Liu et al., 2018b) concordant anatomical imaging, col. This requirement makes it difficult to apply these methods or to enable low-dose PET (Xu et al., 2017). Each of these inno- to new imaging protocols for which large datasets have not yet vations makes critical contributions to improving the safety, been collected. New techniques are being developed to adapt quality and/or value of clinically meaningful information about these learning-based approaches to either require fewer train- brain health which can be gleaned from the imaging study. ing examples or to transfer the information from previously 1 3 Neuroinformatics (2022) 20:943–964 947 Fig. 3 Example reconstructions from 4 × undersampled data (row 1), by comparing the final two columns. The Interleaved and Alternating zoomed-in image patches (row 2), difference patches between recon- architectures produce two slightly different reconstructions, both of structions and ground truth images (row 3), and frequency space which better eliminate blurring and 'ringing' artifacts, where multiple reconstructions (row 4) are shown here to visually communicate the copies of the image appear stamped on top of each other impact of this reconstruction approach. It is most easily appreciated collected datasets of one protocol to a new protocol of inter- can be incorporated in each workflow step, but similar prin- est (Han et al., 2018). And, for any reconstruction method, ciples extend to other volumetric imaging techniques such uncertainty quantification techniques will be needed to high- as PET and CT. light regions of reconstructions with a high likelihood of error (Edupuganti et al., 2021). These uncertainty quantifications will Quality Assurance and Harmonization enable radiologists to understand when more detail is needed to identify a particular feature of interest (Edupuganti et al., 2021), Expert human labellers typically perform image quality possibly requiring re-imaging of the patient. assessment (QA), but this process is labor-intensive and can suffer from low inter-rater reliability. Carefully designed machine learning techniques promise to enable fast, easily Section II: Machine Learning Applications accessible, consistent QA. Previously proposed approaches for Volumetric Image Processing use carefully curated quality metrics as input features to various types of classifiers which label images as usable or Once brain imaging data is collected and reconstructed, unusable (Esteban et al., 2017; Küstner et al., 2018; Pizarro there are several steps in the image analysis pipeline where et al., 2016). Crowdsourcing approaches have also been used machine learning can improve the extraction of meaningful, to improve the accuracy of these automatic QA tools. The quantitative features related to brain health. In this section use of many non-expert, human raters as inputs to a convo- we explore some of these advances, giving an overview of lutional neural net improves the accuracy of classification each step and examples of how machine learning models are over a single site data set (Keshavan et al., 2019). Further, used. We focus on MRI to demonstrate how ML techniques a web-based API acting as a quality metric repository has 1 3 948 Neuroinformatics (2022) 20:943–964 Fig. 4 Atlas construction (concept in panel A) can enable quantifica- construction versus a prospectively gathered longitudinal cohort) tion of brain development across ages (panel B- schematically indi- and can detect abnormalities as outliers to normal (panel C). ADC- cating the benefits of using a clinical cohort of individuals for atlas Apparent Diffusion Coefficient increased the volume of quality metric labeled, multi-site images for a machine learning prediction task to be invariant data available to be used to develop new, more generalizable to the scanners on which they were acquired (Dinsdale et al., QA tools (Esteban et al., 2019). 2021). In addition to ensuring the quality of individual scans, batch effects affecting images acquired at different locations Quantification of Brain Health and Detection or times must be eliminated to perform large-scale, multi- of Abnormality site studies. Machine learning approaches provide flexible methods to detect and remove the relevant site-specific Machine learning approaches can also be used to charac- effects. One approach is to directly convert data acquired terize healthy brain characteristics and identify deviations in one setting to the data that would have been acquired from the norm. During the symposium, Dr. Yangming Ou in a different setting. During the symposium, Dr. Cetin- described how to construct ‘normal’ atlases using group- Karayumak presented such a retrospective harmonization wise unbiased image registration. Brain MRI atlases sum- technique which represents diffusion MRI (dMRI) data as a marize healthy brain anatomy and typical signal intensity combination of spherical harmonic basis functions. Rotation- profiles at the voxel-, regional-, fiber-, and whole-brain invariant features are derived at each voxel from the computed levels (Guimond et al., 2000) (Fig. 4A). Brain atlases con- basis function coefficients for each image, and a mapping structed from imaging data can be used in multiple ways is computed between the features of target and reference to quantify brain health. One example is the quantification scanners in order to harmonize them (Cetin Karayumak of normal childhood development (Ou et al., 2017; Sotardi et al., 2019). A different approach is to learn intermediate et al., 2021). A series of constructed atlases from cohorts representations invariant to the scanner on which any image of healthy subjects clustered by age can enable longitu- was acquired. These intermediate representations can then dinal quantification of brain development from data sets be used to reconstruct images without site-specific effects where every subject was scanned once (Fig.  4B). This is (Moyer et al., 2020). Alternatively, instead of removing or not only cost effective when constructed from clinically transforming site-specific effects at the image level, a third acquired brain scans, but also has the potential to incorpo- strategy is to encourage downstream features derived from the rate a more comprehensive range of healthy variation than 1 3 Neuroinformatics (2022) 20:943–964 949 data acquired in a single or set of pooled research studies. archives by using deep learning to transform lower- Another use of quantitative brain atlases is to detect subtle quality images into higher-quality ones, thus enabling abnormalities due to a wide range of disorders (Pinto et al., use of advanced image segmentation tools. In particular, 2018) (Fig. 4C). Atlas-quantified voxel-wise deviation val- several such tools are built for MP-RAGE scans, which ues can be used as features in classical machine classifiers are popular due to their SNR efficiency and contrast. (O’Muircheartaigh et al., 2020) or deep convolutional neural At this symposium, Dr. Juan Eugenio Iglesias presented networks (Baur et al., 2021) to further improve the accuracy an approach to synthesize isotropic 1  mm MP-RAGE and generality of atlas-based detection of deviations from volumes from low-resolution scans of arbitrary contrast, brain health. This strategy has been used for structural MRI enabling their segmentation and analysis with standard (Baur et al., 2021; O’Muircheartaigh et al., 2020) and diffu- neuroimaging tools (Iglesias et al., 2020). An example sion MRI (Pinto et al., 2018). is shown in Fig.  5, where a 5  mm axial FLAIR scan is transformed into a 1 mm isotropic MP-RAGE scan, Segmentation and subsequently segmented with FreeSurfer, which requires 1  mm isotropic T1 data—and thus could not Automatic segmentation of brain images enables quantita- have processed the FLAIR scan directly, due to MR tive estimation of the volumes of brain structures that can contrast mismatch and insufficient resolution. lead to other indicators of brain health. These quantitative estimates enable population studies as well as longitudinal Visualization analysis within individual subjects. Most previous work on brain segmentation has focused on MRI, which provides Visualization frameworks can foster deeper understanding and detailed images with an ever increasing range of specifi- facilitate interpretation of high-dimensional clinical imaging cally tuned contrasts for visualizing different details of brain data. Further, targeted visualizations allow developers to anatomy and function (Akkus et al., 2017). design and optimize computational algorithms. State-of-the-art Freesurfer is one example of a widely used package visualization tools deal with challenges such as large amounts of for brain MRI analysis and includes a machine learning data such as in diffusion and functional MRI, and the inevitable approach to segment many brain structures (Fischl et al., variation of file formats across different institutions. Web-based 2002) as part of a larger image analysis pipeline. This tech- tools such as Fiberweb (Ledoux et al., 2017) or XTK (Haehn nique involves finding the maximum a posteriori estimate et al., 2014) have contributed to brain imaging visualizations of a segmentation of an anatomical brain region (e.g. hip- and 3D rendering of connectivity in recent years. Many other pocampus), given the image to be segmented and a linear visualization tools not limited to DTI data emerged recently, transform mapping it to an expertly curated segmentation such as Neurolines (Al-Awami et al., 2014) to visualize 3D brain atlas. This technique only employs approximately one hun- tissue in 2D, and comparative visualizations for fMRI brain dred labeled scans for a specific atlas, but the entire segmen- images (Jönsson et al., 2019). tation procedure takes several hours. There has therefore At the symposium, Ms. Loraine Franke presented her been recent interest in neural network-based segmentation work on developing web-based interactive visualization methods, which provide segmentations on the order of sec- tools for diffusion tractography imaging data (Franke & onds (Akkus et al., 2017; Despotović et al., 2015) to address Haehn, 2020; Franke et al., 2020). Her open-source tool, the requirement to perform expert level segmentation on FiberStars (Franke et al., 2020) (Fig. 6) enables researchers large scale image data sets. to create low-dimensional cluster representations of high In these approaches, a convolutional neural network typi- dimensional data, select, visualize, and compare multiple cally directly predicts human-labeled segmentations from clusters across multiple patients, and visualize individual patches or volumes and requires many labeled images to patient fiber tracts. By using different projection techniques train. Furthermore, these approaches are extremely sensitive for multidimensional scaling such as t-SNE (van der Maaten to shifts in input image intensity. To apply these methods to & Hinton, 2008), PivotMDS (Brandes & Pich, 2007) and scans of a different contrast or resolution, additional labels others, the FiberStars tool lets the user interactively explore must be collected and used to retrain or fine-tune the net- high dimensional data. For example, FiberStars enables works. Thus, recent research has focused on unsupervised users to answer research questions with comparative ensem- deep learning approaches for training brain segmentation ble visualizations, especially for evaluating and testing networks, (Dalca et al., 2019) or on adapting trained net- hypotheses, or to analyze factors combined with pathologi- works to new imaging analysis task scenarios (Kamnitsas cal findings. FiberStars addresses a large class of complex et al., 2016). visualization challenges for multidimensional data or data It is now possible to aggregate larger cohorts of composed of collections of patients. useful brain image data from clinical and/or research 1 3 950 Neuroinformatics (2022) 20:943–964 Fig. 5 Left column: coronal plane of an MP-RAGE scan (top, slice produced by Dr. Iglesias’s tool. Right column: automated segmenta- thickness: 1  mm)) and corresponding coronal plane of an axial tion of the original and synthetic MP-RAGE volumes produced by FLAIR scan (bottom, slice thickness: 5 mm) from the ADNI dataset FreeSurfer (Fischl et al., 2002) (adni-info.org). Middle column: synthetic 1  mm MP-RAGE volume Some of the most advanced work focuses on the diag- Section III: Machine Learning Advances nostic needs for patients with epilepsy, an area for which in Interpretation and Analysis EEG is already in active clinical use. In the current standard of Non‑volumetric EEG Data of care, making diagnoses and therapeutic decisions relies on painstaking manual annotation of many hours of EEG Another important, and emerging, area where machine recording by highly-trained expert epileptologists (Si, 2020). learning approaches are enhancing the understanding of brain Machine learning methods enable automatic detection of health is in clinical applications of electroencephalography, markers of epilepsy in interictal (non-seizure) data using or EEG data. EEG is multidimensional time series data, specific spectral, morphological, or network-based features. where multiple electrodes are placed on the scalp resulting While some feature-based approaches attempt to replicate in simultaneous channels of data being collected at a high the eye of the expert using features like those epileptologists time resolution. EEG is currently in clinical use for multiple, observe; other end-to-end deep learning and neural network specific applications, such as for diagnosing and monitoring models attempt to glean undiscovered signatures of epilepsy sleep disorders, epilepsy, disorders of consciousness, stroke, from the raw data itself. One example of this is classifying real-time electroconvulsive therapy (ECT) patient monitoring, routine EEGs into normal vs. abnormal, where abnormal and anesthesia (Roy et al., 2019a, b). EEG has the unique is, by definition, heterogeneous and context-dependent (van advantages of being non-invasive, relatively inexpensive, and Leeuwen et al., 2019). Machine learning based clinical deci- more adaptable to naturalistic or ambulatory settings compared sion support for epileptologists for diagnosis and localiza- to other imaging modalities. In some cases, even a few EEG tion of epileptic foci are highly promising as they reveal electrodes in a specific location can yield enough information interrelationships between brain regions and activity that for inference, without the need for EEG across all of the cortex. are difficult to discern by eye. Therefore, machine learning approaches can not only greatly In contrast to epilepsy, where EEG is already being used streamline existing clinical applications of EEG, but they can clinically, machine learning approaches are expanding the also open the door to new applications such as earlier, less potential for EEG-based diagnostic biomarkers for other expensive, or more accessible diagnostics (Michel & Murray, diseases, such as Alzheimer’s Disease (Escudero et al., 2006; 2012; Miranda et al., 2019). 1 3 Neuroinformatics (2022) 20:943–964 951 Fig. 6 Split screen showing 3D representations of fiber tract anatomy ent patients is displayed with additional two-dimensional representa- given by fibers of dMRI scans across different subjects. The menu tion radial plots at the bottom of each patient’s panel showing scalar bar at the left facilitates toggling on and off visualization of differ - values associated with each of the fiber tracts. For each anatomical ent subjects (top left), cluster (middle left, showing the Callosum tract, the 2D radial plots show mean and standard deviations of the Forceps Major), and coloring of the 3D tract by a selected scalar different scalars on each axis. Demographic information about each value (bottom left, showing a measurement of fractional anisotropy patient is shown above the 3D visualization, for example, age, gen- (FA2) from the DTI scan). High values of fractional anisotropy are der, height and weight. Each patient is anonymized by a number seen colored in red while lower values are colored in blue. Inter-hemi- in the labels next to the anatomical fiber tract name in purple. Other sphere crossing of the third patient shows no red colors and therefore relevant measurements for analysis are mean fiber length, number of no high fractional anisotropy values. Tractography from five differ - fibers or fiber similarity Gallego-Jutglà et al., 2015; Jelles et al., 1999; Lehmann et al., National Sleep Research Resource (Sleep Data—National 2007; Tzimourta et al., 2021; Woon et al., 2007). However, Sleep Research Resource, 2021  https:// sleep dat a. or g/), these methods are further from clinical deployment than those the PhysioNet Computing in Cardiology Challenge 2018 for epilepsy, mostly in feature discovery stages. Analogous (Ghassemi et al., 2018), and the TUH Abnormal EEG corpus strides are being made to discover novel, cost-ee ff ctive, and (Alhussein et  al., 2019; Gemein et  al., 2020; Roy et  al., ambulatory EEG-based biomarkers for diagnosing stroke, 2019a, b; Temple University EEG Corpus Downloads, 2021). schizophrenia, and attention deficit hyperactivity disorder (Ahmadlou & Adeli, 2011; Hosseini et  al., 2020; Phang et al., 2020; Sastra Kusuina Wijaya et al., 2015). While these Section IV: Brain Health as Assessed directions have great potential for impact if successful, since by Deviations from Healthy Aging they are new clinical applications of EEG, their success depends on connecting sound and robust machine learning Another machine learning approach to characterize brain health algorithm design to underlying physiology, which can prove is to summarize an image or biosignal into a single metric that elusive. Interpretability will likely also come into play, since reflects brain health, such as brain age estimation (Fig.  7d) clinicians must be convinced of the specific clinical utility of (Cole et al., 2019). The difference between estimated brain EEG for each new application. age and actual chronologic age, known variously as predicted A recurring theme in the development of machine age difference (PAD), Brain Age Index (BAI) or ΔBrainAGE, learning methods that is the same for EEG, as it is for any has identified accelerated aging in individuals with cognitive other imaging modality, is the availability of large, labeled impairment (Liem et al., 2017; Poddar et al., 2019), traumatic datasets. Three such sources for large EEG datasets are the brain injuries (Cole et al., 2015), schizophrenia (Cole et al., 1 3 952 Neuroinformatics (2022) 20:943–964 Fig. 7 Machine learning (ML) can estimate a patient’s brain age and at various ages; (c) cross validation to quantify the accuracy of the quantify abnormal (accelerated or delayed) aging. (a) training sam- ML model; and (d) when applied to target patients, the ML model ples consisting of normal brain MRIs from a large set of individu- can quantify deviations from normal brain aging als; (b) ML algorithm that learns how a normal brain MRI appears 2018), Alzheimer's disease (Bashyam et al., 2020), and diabetes Another symposium speaker, Dr. Haoqi Sun, presented (Franke et al., 2013). Deviations from expected brain age have his work on a feature-based machine learning model that also been reported for more subtle changes due to social and takes advantage of the fact that brain activity as recorded environmental influences, including a protective decrease in by EEG during sleep naturally varies with age (Leone brain aging for long-term meditation practice (Luders et al., et al., 2021; Paixao et al., 2020; Sun et al., 2019; Ye et al., 2016), music-making (Rogenmoser et al., 2018), and a higher 2020). Features from both time and frequency domains level of education (Steffener et al., 2016), as well as accelerated of each sleep stage are used to compute an overall brain aging associated with smoking and alcohol consumption age. Figure 8 shows the scatter plot of chronological age (Guggenmos et al., 2017; Ning et al., 2020). vs. sleep EEG-predicted brain age, and eight example sleep At the symposium, Dr. Ou presented his recent work (He EEGs from across the lifespan with their chronological et al., 2020, 2021) on a novel, deep convolutional neural net- age and calculated brain age shown. Dr. Sun showed that work brain age prediction model that uses both morphologi- across two large sleep EEG datasets, people with significant cal and contrast-based changes in brain MRI data to estimate neurological or psychiatric disease show a mean excess brain brain age. This work was enabled by collating 11 different age (compared to chronological age) of 4 years compared data sets and carefully curating a very large, harmonized to healthy controls on a population level, while those with dataset that included enough healthy subjects of all ages hypertension or diabetes show a mean excess brain age of to train, test and validate the method (Fig. 7a). By explic- 3.5 years compared to healthy controls (Sun et al., 2019). itly splitting the T1-weighted brain MRI into morphometry Sun and colleagues have validated the association of (spatial information) and contrast (tissue based signal infor- significant differences between sleep EEG based age and mation) channels, his attention-driven multi-channel fusion chronological age in patients with dementia and MCI (Ye network (Fig. 7b) improved the accuracy of age estimation et al., 2020), people diagnosed with HIV under antiretroviral as compared to each channel alone, or naive fusion of two therapy (Leone et al., 2021), and all cause mortality (Paixao channels without their proposed attention mechanisms, when et al., 2020). applied to 16,705 normal brain MRIs acquired over the lifes- As with sleep EEG, features of brain activity under gen- pan (0–97 years of age) (He et al., 2021). The team cross eral anesthesia have also been demonstrated to change with validated their work against multiple published brain age age, allowing the EEG patterns measured during adminis- estimation algorithms and using multiple independent test tration of general anesthesia to be evaluated as a marker of data sets (Fig. 7c). A critical advantage of this end-to-end brain age (Akeju et al., 2015; Lee et al., 2017; Purdon et al., method is that it has the potential to differentiate between 2015). Similar ideas about indicators of brain health under abnormal aging associated with contrast change (e.g., general anesthesia are motivating the development of EEG lesions) and those associated with morphometric changes machine learning methods to monitor and assess disorders (e.g., atrophy). This is an important contribution toward of consciousness, since no other behavioral markers can be increasing the specificity of brain age estimator biomark - used (Engemann et al., 2018). ers, a major issue for this line of research (Kaufmann et al., A fundamental challenge in using brain age estimation as 2019). an index of brain health and/or meaningful clinical indicator 1 3 Neuroinformatics (2022) 20:943–964 953 Fig. 8 Illustration of sleep EEG-based brain age. (Left) The scatter stages) (top in each subplot), where the top, middle, and bottom rows plot of chronological age vs. brain age where the diagonal dashed are patients with young, middle, and old chronological age (CA, in red line indicates where chronological age equals brain age. The years) respectively; while the left, middle, and right columns are sub- mean absolute deviation (MAD) is 7.8  years and Pearson’s correla- jects with young, middle, and old brain age (BA, in years). Compari- tion R = 0.82. (Right) The confusion matrix of example EEG spectro- son within each row reveals different sleep EEG microstructures for grams (bottom in each subplot) and hypnogram (trajectory of sleep different brain ages while at similar chronological age is that the rate of age-related changes in brain structure status at individual level is an active area of research (Al and function (e.g. sleep) vary across the lifespan such that Zoubi et al., 2018; Cole & Franke, 2017; Mohajer et al., early and late life changes are more readily detected, but 2020; Varikuti et al., 2018). For example, in the case of sleep are very subtle between 30 and 60 years of age. For both EEG-based brain age, the density of sleep spindles (count/ MRI- and EEG-based brain age prediction, the sensitivity is hour) one of the features used in the model, appears to be a lowest during this part of the lifespan. Not surprisingly, one heritable trait based on the expression of CACNA1l, a gene promising application of MRI-based brain age prediction that is associated with both schizophrenia and sleep spindle is early detection of future neuropsychiatric disorders in formation (Merikanto et al., 2019). children and/or adolescents (Chung et al., 2018). The relative Despite these limitations, it is intriguing, and poten- stability of structural MRI measures bound the temporal tially clinically advantageous, that lifestyle choices such resolution of brain age estimates using that modality (Cole as exercise and sleep can modify these quantitative metrics & Franke, 2017; Karch et al., 2019), while the significantly of brain age in directions that reflect known associations higher night-to-night variability of sleep EEG-based brain with brain health. Studies have shown that actively exercis- age estimates is both a challenge to overcome if looking ing leads to an orchestra of changes in energy metabolism, for stability, but also a potential additional source of oxidative stress, inflammation, tissue repair, growth factor meaningful signal to exploit in future work (Arnal et al., response, and regulatory pathways in the brain (Contrepois 2020; Arnardottir et al., 2021; Hogan et al., 2021). et al., 2020). Sleep has a bidirectional relationship with the There are important caveats to the use of a brain age as a immune system (Irwin, 2019), therefore there is evidence marker of brain health since deviations from chronological for and reason to expect that exercise can improve sleep and age could be due to multiple factors. Brain age estimation thereby improve brain health, which will be reflected in nor - studies remain population-level statistical tests because malized sleep-based brain age biomarkers in people with current approaches lack the sensitivity to accurately assess evidence of accelerated aging. clinically meaningful deviation at the individual patient level. Because of these limitations, brain age is currently viewed as a screening tool where large deviations call for Section V: Application of Machine Learning further investigation. More work is required to improve Techniques for Diagnostics, Prognostication, the specificity and clinical utility of brain age estimation. and Personalization of Medical Treatments Combining EEG- and MRI-based brain-age estimation techniques with and without additional features (e.g. Imaging plays a key role in the clinical evaluation of genomic markers, demographics, socioeconomics status, and pathological changes that can be readily distinguished environmental factors) to more accurately predict disease from a healthy brain. Neuroradiologists routinely use 1 3 954 Neuroinformatics (2022) 20:943–964 neuroimaging modalities such as CT, MRI, and PET for informed by both data from a specific patient and aggregated both qualitative and quantitative assessment of diseases information from larger patient datasets (Calhoun et al., 2021; from infectious, autoimmune, oncological, degenerative, Vieira et al., 2017; Zhang et al., 2020). and vascular etiologies. However, despite standardization Data-driven precision therapeutics are already being efforts, manual assessment is subject to inter- and intra- translated to the clinic using transcranial magnetic stimula- rater variability (Filippi et  al., 1995; Provenzale & tion (TMS) and other targeted brain stimulation approaches. Mancini, 2012; Provenzale et al., 2009; van Horn et al., At the symposium, Dr. Shan Siddiqi presented on this work, 2021). As such, there is intense interest in automating highlighting that TMS targets for any given symptom may radiological assessment with machine learning. A popular be identified based on the location of brain lesions that cause approach is radiomics (Beig et al., 2020), which focuses the same symptom (Cash et al., 2020; Davey & Riehl, 2005). on the extraction of pertinent quantitative imaging features Complementing Dr. Siddiqi and his team’s research is a often followed by incorporation of these features into a large body of work focused on machine learning-based target predictive machine learning algorithm. These imaging optimization of field distributions for transcranial magnetic features are computational imaging descriptors reflecting and/or electric stimulation that factor in the biophysical measures such as size, shape, intensity distribution, and properties of biological tissues or feedback from real-time intensity heterogeneity (Zhou et al., 2018). Indeed, these fMRI. “The Automatic Neuroscientist” framework uses feature-based radiomic approaches have found success for real-time fMRI in combination with Bayesian optimization early detection (Sørensen et al., 2016), diagnosis (Kniep “to automatically design the optimal experiment to evoke a et al., 2019; Regenhardt et al., 2021; Tanioka et al., 2020; desired target brain state.” (Lorenz et al., 2016). Machine Zhou et al., 2020), prognostication (Macyszyn et al., 2016; learning techniques have been applied towards rt-fMRI neu- Stefano et al., 2020; Tang et al., 2020), treatment response rofeedback studies, where a neurofeedback signal can be prediction/assessment (Cai et al., 2020; Chang et al., 2016; derived using supervised learning methods such as linear Hofmeister et al., 2020), and non-invasive determination models and support vector machines (LaConte et al., 2007). of molecular markers (Beig et  al., 2018; Pan et  al., In addition, both data driven and hypothesis driven analy- 2019) for a wide variety of diseases. More recently, deep ses of functional connectivity data have been used to predict learning approaches (Chang et al., 2018a, b; Rauschecker clinical outcomes including treatment response in patients et al., 2020; Titano et al., 2018) have gained traction for (Whitfield-Gabrieli et al., 2016) as well as to predict pedi- similar tasks due to these approaches foregoing the need atric vulnerability to psychiatric disorders including psy- to pre-engineer imaging features. Some approaches have chosis (Collin et al., 2019, 2020), depression (Chai et al., even shown the utility of combining radiomics with deep 2015), anxiety, and ADHD (Collin et al., 2020; Cui et al., learning (Lao et al., 2017; Xiao et al., 2019). While there 2020). At the symposium, Dr. Susan Whitfield-Gabrieli is great promise for these automated approaches, they do presented on these approaches, sharing evidence that con- not come without pitfalls. Radiomics, in particular, has nectivity between the medial prefrontal cortex (MPFC) and been challenged by variability stemming from differences the dorsolateral prefrontal cortex (DLFPC) can be used as in image acquisition, pre-processing, segmentation, and a biomarker to predict attentional problems in a normative feature implementation (Hoebel et  al., 2021; Kalpathy- pediatric population as assessed four years later, where Cramer et  al., 2016; Schwier et  al., 2019). Approaches greater baseline MPFC-DLPFC connectivity predicted to rectify these sources of variability and harmonize worsening of attentional issues (Whitfield-Gabrieli et al., radiomic features have been an active area of study (Carré 2020) while decreased baseline subgenual anterior cingulate et al., 2020; Marcadent et al., 2020; Orlhac et al., 2018; (sgACC)—DLPFC connectivity predicted worsening of anx- Parmar et al., 2014; Zwanenburg et al., 2020). Similarly, iety/depression. As psychiatric neuroimaging research has deep learning approaches also suffer from a lack of evolved from the description of patient cohorts using simple generalizability across different image acquisition settings group comparisons towards a focus on individual differences and patient populations (AlBadawy et al., 2018; Chang and “predictive” analytics, preliminary studies suggest that et al., 2020; Zech et al., 2018). These challenges will need intra-individual fluctuations of brain activity provide better to be addressed before these automated approaches can be prediction of symptoms than group-based studies. Machine effectively utilized. learning integrated with experience-sampling can be used to Beyond its neuroradiologic applications to promote brain produce novel brain-based predictive models of state fluctua- health with early detection, diagnostics, or prognostication tions (e.g., fluctuations of mind wandering) which general- related to neurologic disease, machine learning also has izes to both healthy and clinical populations (Kucyi et al., applications towards brain health as it relates to precision 2021). Dr. Whitfield-Gabrieli also highlighted the use of medicine—i.e. the development of personalized interventional mindfulness based rt-fMRI neurofeedback as a non-invasive, therapies for a broader range of neuropsychiatric disorders personalized circuit therapeutic to reduce symptom severity 1 3 Neuroinformatics (2022) 20:943–964 955 in psychotic patients as well as for teens with major depres- Machine intelligence in medical imaging is one of the sive disorder and/or anxiety. (Bauer et al., 2020; Stoeckel most vibrant fields within the application of machine learn- et al., 2014) These pioneering studies provide strong motiva- ing in healthcare, and one of its biggest subfields is quanti- tion to pursue imaging based treatments. tative imaging (QI). QI refers to extraction of quantifiable Overall, machine learning based methods have potential features from medical images that serve as biomarkers for to augment diagnostic and treatment workflows. As with all specific physiological conditions, such as features relating clinical interventions, the overarching goal is to improve to aspects of brain health which have been discussed above. patient outcomes, either within a specific decision point or A premarket submission for a QI function requires a func- longitudinally. While promising, more rigorous prospective tion description including the level of automation (manual, and external validation studies in diverse clinical scenarios semi-automatic or fully automatic), a brief description of and populations are needed before these methods can be the training algorithm, quantitative performance specifica- deployed for widespread use. tions, and instructions used for semi-automatic labeling of the training set. The biggest part of the premarket submis- sion is the technical performance assessment which should Section VI: Additional Considerations include a definition of the QI function, its relationship to for Clinical Deployment the measurand, and the use conditions. For example, this could be a “brain age” assessment from MRI data applica- ble to images of a specific resolution collected on a specific Regulatory Framework MRI system. It should also specify the performance metrics and characterize the performance of the QI function under To deploy any of the advances highlighted in the symposium the predefined conditions. In the mentioned example, per - that use machine learning algorithms in clinical practice, formance metrics could include accuracy as measured in proposals must first clear the regulatory process as set by deviation between actual age and estimated age in a nor- the Center for Devices and Radiological Health within the mative cohort as well as bias or precision as measured in FDA that handles medical devices. Most machine learn- reproducibility or repeatability. A priori acceptance criteria ing methods in healthcare are categorized as software as a regarding these performance metrics should also be set along medical device (SaMD) which is a subcategory under soft- with restrictions and limitations on usage, and the results of ware related to medical devices under the medical device a study presented where the outcomes are compared to the umbrella. The pathway to market depends on the risk asso- predefined acceptance criteria. ciated with the software, which in turn depends primarily on 1) significance of information provided by SaMD to a healthcare decision and 2) state of healthcare situation or Ethical Considerations During Machine Learning condition; a more critical situation yields a higher risk rat- Model Development ing. Traditionally, SaMD algorithms need to be locked, i.e., give the same output for the same input, after they are sub- As the machine learning applications in this report mature in mitted to the FDA for premarket approval. This is impracti- their development, there are a number of vital ethical issues cal for machine learning software in situations where it is to be taken into consideration. While not the primary focus often desirable to continuously update the machine learning of the symposium, both the organizer, Dr. Randy Gollub models based on user data (e.g. to accommodate updates in and keynote speaker, Dr. Simon Eickhoff emphasized the scanner hardware and software). The FDA has proposed a importance of these aspects, pointing out a few examples of new regulatory framework based on a total product life cycle how, where, and why they are relevant. Some of these ethical approach, wherein the initial premarket submission outlines considerations have established guidelines or technical best the modifications that might take place in the future. The practices that need to be more widely used; others are ongoing manufacturer can then continuously update their machine discussions for which there is not yet a clear-cut solution (see learning models based on new user data without having to for example the Fair ML for Health Workshop that was held go through a new premarket submission provided that the during the NeurIPS 2019 Workshop (Fair ML for Health— update is within the SaMD Pre-Specifications and algo- Accepted Papers, 2021,  https:// www . f air m lf or h ealt h. com/ rithm change protocol (Digital Health Center of Excellence, accep ted- papers). It is crucial that scientists and researchers 2021, https:// www. fda. gov/ medic al- devic es/ digit al- health- participate actively in these discussions at each stage of center- excel lence/ sof tw ar e- medic al- device- samd). These development of these methods. We note that this section of guidelines are under active discussion, development, and our report is by no means comprehensive; for more in-depth refinement in collaboration with industry, academic, and discussions of these issues, see (Beauvais et al., 2021) and clinical leaders. (Chen et al., 2020). 1 3 956 Neuroinformatics (2022) 20:943–964 Data Sharing research use. With the increasing use of large datasets for multi- ple studies and across long periods of time, it is difficult to track Data sharing across institutions may eventually become nec- all of the downstream uses of a single person’s data. After data essary to create large enough datasets to train sophisticated collection, development and validation of new methods, these machine learning algorithms. To mitigate the risks of breach methods may eventually be commercialized. However, inherent of privacy, security, and confidentiality, robust de-identification to any trained machine learning model or algorithm is the data algorithms that retain all necessary imaging data elements are that was used for such training. The data is inextricably tied to essential, and all modalities of data must be scrutinized to ensure any intellectual property or commercial potential that results that there are not additional unintended sources of protected from the development process. Is it fair to allow patenting of information amongst them (e.g. private DICOM metadata tags). trained models or algorithms on data collected from people who Secure cloud servers and backup protocols, expert curation and did not consent to its possible use for commercial prot fi ? Should maintenance, and strict guidelines and training for researchers those people be included in any such profit? These questions on how to securely access, store, and dispose of data are all addi- must be answered as machine learning models become inte- tional tools to minimize the chances of confidentiality breaches grated into clinical pipelines. or loss of data. Federated learning methods which allow data to be stored only at the location where it was collected while allow- Bias in Datasets ing for multisite analysis are another means to support robust, yet protected, data sharing (K. Chang, Balachandar, et al., 2018; In machine learning, the phrase “garbage in, garbage out” Chang, Grinband, et al., 2018). For all these approaches, fre- reflects the fact that bias, noise, or flaws in the underlying quent communication between all institutions involved will data used to train a model will undoubtedly affect the qual- also ensure that everyone is kept apprised of possible issues in ity, accuracy, and validity of the results. Therefore, ensuring a timely manner and that any changes are implemented in an high quality data that is highly representative of the popula- organized and efficient manner. tions under study is paramount to the ethical and effective development of these methods. One key component of this Informed Consent for Expanded or Later Use of Data for assessments of brain health is ensuring adequate repre- sentation of traditionally underrepresented subpopulations It is becoming increasingly common for large datasets to be used in research, including underrepresented minorities, women, and reused in multiple studies and towards die ff rent machine low and middle income nationals, transgender and gender learning algorithms once they have been collected. This is non-conforming individuals, undocumented immigrants, mainly due to the cost in time, money, and resources to amass and pregnant women, especially from an intersectional lens. an entirely new dataset for each research question. Repurposing This is especially important because of the specific men- existing datasets across many studies is overall a very efficient tal and behavioral health issues which impact brain health and effective option; however, the wishes of those from whom in many of these subpopulations. It also includes consid- the data is collected must be respected. Most current informed erations in the study design itself to ensure these popula- consent paradigms are based on data being collected for a sin- tions are not excluded inadvertently by data acquisition gle study and therefore obtain informed consent from a patient methods, for example by failing to include more than the for that single study alone. However, this system needs to be traditional binary options when documenting gender. Even modified to reflect that it is likely a patient’s data could be used if the intentions of researchers are to include all popula- for many studies even decades into the future, most of which tions, other aspects of the study design can inadvertently cannot even be fathomed at the time of data collection. Patients be biased towards certain populations. For example, studies should, at minimum, have the ability to ‘opt out’ of having their that require mobile phone downloads of certain apps or track data used in future studies without their explicit consent. Many social media use exclude populations who do not have access current guidelines state that if the data is de-identified, it can to smartphones or social media. be shared and used for new studies without re-obtaining con- Even once a study is underway, oversight and periodic sent, often after obtaining a waiver of consent from the local assessments of study recruitment practices should be done Institutional Review Board. However, current trends warrant re- to check for inadvertent exclusion of certain populations. For examination of these guidelines with a goal of securing patient example, the inclusion and exclusion criteria of many stud- consent for wider, protected data use at the time of enrollment. ies, especially randomized controlled trials, are often writ- ten with purely scientific or clinical considerations in mind Intellectual Property and Commercialization relating to the treatment or diagnostic in question. However, they can result in a study population that is too restricted Clearer intellectual property guidelines are needed regarding and not reflective of the actual population of interest. Stud- models or algorithms developed from patient data collected for ies that involve multiple study visits at different times may 1 3 Neuroinformatics (2022) 20:943–964 957 discourage participation of those who do not have the access are non-trivial. How should such findings be handled? Do or flexibility to come to the research site several times. Any clinicians have an obligation to inform patients? What about researcher using existing datasets should hold to these same for measures such as one of the indices of ‘brain age’ for standards when checking the subject profile of the already which the implications are still under study? Clear clinical collected data and include this information, including limita- practice guidelines for the handling of sensitive informa- tions of the dataset, in any resulting publications. tion relating to brain health must be developed prior to the Technical approaches have been suggested as tools for deployment of any such algorithm. handling disproportionate representation of certain sub- While many of these ethical considerations are gray areas groups within large datasets. For example, some approaches for which we can only postulate guidelines and not clear force neural networks to learn intermediate representations answers, ongoing discussion will hopefully lead to new which cannot be used to predict a protected attribute of inter- best practices that adhere to the highest ethical standards est, e.g. gender or race (Dinsdale et al., 2021). on each of the issues discussed, to safeguard patients and their anatomic and physiological datasets. For researchers, Quantifying/understanding Uncertainty it is important to keep in mind that the misinterpretation or generalization of one poorly designed high-profile study or Whenever possible, researchers and scientists should attempt one breach in confidentiality can be enough for an entire to quantify the uncertainty of their model predictions using field to lose credibility. statistical tools such as the confidence interval. Before such algorithms are implemented, clinicians should receive training on how to interpret the results given the limita- Conclusion tions of any model, including uncertainty. There is much attention being focused on these issues, including annual The MGH McCance Center for Brain Health and Harvard- workshops that have been held since 2019 at the MICCAI MIT Health Sciences and Technology Neuroimaging Training meetings- “UNSURE Uncertainty for Safe Utilization of Program co-hosted virtual symposium, “Neuroimaging Machine Learning in Medical Imaging” with presentations Indicators of Brain Structure and Function—Closing the Gap and awards for work in areas such as risk management of Between Research and Clinical Application,” explored the machine learning systems in clinical pipelines, measurement recent explosion of machine learning approaches augmenting errors, methods for modeling noise in data, validation of the clinical and scientific neuroimaging pipeline. Researchers uncertainty estimates, calibration of uncertainty measures presented cutting-edge techniques for acquiring more informative and more (Greenspan et al., 2019). imaging data, more effectively analyzing this acquired data, and more precisely acting on the insights from this analysis Ethical Considerations at the Stage of Deploying to guide and individualize treatment decisions. The work a Machine Learning Model presented at this symposium highlighted several open research directions which must be explored in order to implement these Privacy and Confidentiality of Information techniques in practice. For example, progress in this field will require techniques for robust generalization of machine learning As machine learning tools for brain health emerge, an techniques to more realistic, heterogeneous datasets as well important question to answer is: who should have access as methods for identifying the uncertainty present in machine to the outputs of these methods? Should the patient have learning-based predictions and presenting this information to unfettered access to their own ‘brain health’ information? end users within a clinical workflow. Toward these ends, our Should all clinicians who might interact with that patient field will need to ensure the availability of sufficiently large, have access? If it will be integrated into the medical record, curated data sets; the ability to share valuable data sets thus how do we prevent it from affecting billing or insurance engaging a diverse, committed scientific community (Eickhoff practices? How might the information bias someone’s inter- et al., 2016); and responsible stewardship of brain imaging data action with a patient, especially without accurate reporting to ensure appropriate protections for individual privacy as well (with uncertainty)? These are all considerations that cannot as intellectual property and proper handling of bias in these data. be taken lightly and will have to be addressed to develop With a firm commitment in these directions, machine learning clinical best practices. promises to dramatically improve the early detection, prediction, and treatment of diseases that threaten brain health. Incidental Findings Note: Interested readers may view recorded videos of all symposium presentations and discussions at this As with genetic information and testing, the probabilities of link https://www .y outube. com/ pla ylis t?lis t=PL0A -NKHL V incidental findings in large datasets such as neuroimaging rNF82 vdjey yaBRo iXg77 lCeW. 1 3 958 Neuroinformatics (2022) 20:943–964 Acknowledgements We would like to thank Dr. Ana Namburete for acquisition and sleep staging. Sleep, 43, zsaa097. https:// doi. org/ 10. her helpful review and input on sections of this manuscript. We also 1093/ sleep/ zsaa0 97 would like to extend our appreciation to the Neuroimaging Training Arnardottir, E. S., Islind, A. S., & Óskarsdóttir, M. (2021). The Future Program students and faculty members who participated in the Sympo- of Sleep Measurements: A Review and Perspective. Sleep Medi- sium report preparation and/or presentations- Bruce Rosen, Jayashree cine Clinics, 16, 447–464. https:// doi. org/ 10. 1016/j. jsmc. 2021. Kalpathy-Cramer, John Gustaf Wilhelm Samuelsson, and Katharina 05. 004 Viktoria Hoebel. Bahadir, C. D., Wang, A. Q., Dalca, A. V., & Sabuncu, M. R. (2020). Deep-learning-based Optimization of the Under-sampling Pat- tern in MRI EEE TCP. Transactions Computational Imaging, Funding Open Access funding provided by the MIT Libraries. 6, 1139–1152. Bashyam, V. M., Erus, G., Doshi, J., Habes, M., Nasrallah, I.M., Truelove- Open Access This article is licensed under a Creative Commons Attri- Hill, M., Srinivasan, D., Mamourian, L., Pomponio, R., Fan, Y., bution 4.0 International License, which permits use, sharing, adapta- Launer, L. J., Masters, C.L., Maruff, P., Zhuo, C., Völzke, H., tion, distribution and reproduction in any medium or format, as long Johnson, S. C., Fripp, J., Koutsouleris, N., Satterthwaite, T. D., as you give appropriate credit to the original author(s) and the source, Davatzikos, C., on behalf of the ISTAGING Consortium, the P. A. provide a link to the Creative Commons licence, and indicate if changes disease C., ADNI, and CARDIA studies. (2020). MRI signatures of were made. The images or other third party material in this article are brain age and disease over the lifespan based on a deep brain network included in the article's Creative Commons licence, unless indicated and 14 468 individuals worldwide. Brain, 143, 2312–2324. https:// otherwise in a credit line to the material. If material is not included in doi. org/ 10. 1093/ brain/ awaa1 60. the article's Creative Commons licence and your intended use is not Bauer, C. C. C., Rozenkrantz, L., Caballero, C., Nieto-Castanon, A., permitted by statutory regulation or exceeds the permitted use, you will Scherer, E., West, M. R., Mrazek, M., Phillips, D. T., Gabrieli, need to obtain permission directly from the copyright holder. To view a J. D. E., & Whitfield-Gabrieli, S. (2020). Mindfulness training copy of this licence, visit http://cr eativ ecommons. or g/licen ses/ b y/4.0/ . preserves sustained attention and resting state anticorrelation between default-mode network and dorsolateral prefrontal cor- tex: A randomized controlled trial. Human Brain Mapping, 41, 5356–5369. https:// doi. org/ 10. 1002/ hbm. 25197 Baur, C., Wiestler, B., Muehlau, M., Zimmer, C., Navab, N., & References Albarqouni, S. (2021). Modeling Healthy Anatomy with Artificial Intelligence for Unsupervised Anomaly Detection in Ahmadlou, M., & Adeli, H. (2011). Functional community analysis of Brain MRI. Radiology Artificial Intelligence, 3,. https:// doi. org/ brain: A new approach for EEG-based investigation of the brain 10. 1148/ ryai. 20211 190169 pathology. Neuro Image, 58, 401–408. https://doi. or g/10. 1016/j. Beauvais, M. J. S., Knoppers, B. M., & Illes, J. (2021). A marathon, not neuro image. 2011. 04. 070. a sprint – neuroimaging. Open Science and Ethics Neuroimage, Akeju, O., Pavone, K. J., Thum, J. A., Firth, P. G., Westover, M. B., 236. https:// doi. org/ 10. 1016/j. neuro image. 2021. 118041 Puglia, M., Shank, E. S., Brown, E. N., & Purdon, P. L. (2015). Beig, N., Bera, K., & Tiwari, P. (2020). Introduction to radiomics and Age-dependency of sevoflurane-induced electroencephalogram radiogenomics in neuro-oncology: implications and challenges. dynamics in children. British Journal of Anaesthesia, 115, i66– Neuro-Oncology Advance 2, iv3–iv14. https:// doi. org/ 10. 1093/ i76. https:// doi. org/ 10. 1093/ bja/ aev114 noajnl/ vdaa1 48 Akkus, Z., Galimzianova, A., Hoogi, A., Rubin, D. L., & Erickson, Beig, N., Patel, J., Prasanna, P., Hill, V., Gupta, A., Correa, R., Bera, K., B. J. (2017). Deep Learning for Brain MRI Segmentation: State Singh, S., Partovi, S., Varadan, V., Ahluwalia, M., Madabhushi, A., of the Art and Future Directions. Journal of Digital Imaging, & Tiwari, P. (2018). Radiogenomic analysis of hypoxia pathway is 30, 449–459. https:// doi. org/ 10. 1007/ s10278- 017- 9983-4 predictive of overall survival in Glioblastoma. Science and Reports, Al Zoubi, O., Ki Wong, C., Kuplicki, R. T., Yeh, H., Mayeli, A., Refai, 8, 7. https:// doi. org/ 10. 1038/ s41598- 017- 18310-0 H., et al. (2018). Predicting Age From Brain EEG Signals—A Brandes, U., & Pich, C. (2007). Eigensolver Methods for Progressive Machine Learning Approach. Frontiers Aging Neuroscience, 10, Multidimensional Scaling of Large Data, in: Kaufmann, M., 184. https:// doi. org/ 10. 3389/ fnagi. 2018. 00184. Wagner, D. (Eds.), Graph Drawing, Lecture Notes in Computer Al-Awami, A. K., Beyer, J., Strobelt, H., Kasthuri, N., Lichtman, J. Science. Springer, Berlin, Heidelberg, pp. 42–53. https://doi. or g/ W., Pfister, H., & Hadwiger, M. (2014). NeuroLines: A Subway 10. 1007/ 978-3- 540- 70904-6_6 Map Metaphor for Visualizing Nanoscale Neuronal Connectivity. Cai, J., Zheng, J., Shen, J., Yuan, Z., Xie, M., Gao, M., Tan, H., Liang, IEEE Transactions on Visualization and Computer Graphics, 20, Z., Rong, X., Li, Y., Li, H., Jiang, J., Zhao, H., Argyriou, A. 2369–2378. https:// doi. org/ 10. 1109/ TVCG. 2014. 23463 12 A., Chua, M. L. K., & Tang, Y. (2020). A Radiomics Model for AlBadawy, E. A., Saha, A., & Mazurowski, M. A. (2018). Deep learning Predicting the Response to Bevacizumab in Brain Necrosis after for segmentation of brain tumors: Impact of cross-institutional train- Radiotherapy. Clinical Cancer Research, 26, 5438–5447. https:// ing and testing. Medical Physics, 45, 1150–1158. https:// doi. org/ 10. doi. org/ 10. 1158/ 1078- 0432. CCR- 20- 1264 1002/ mp. 12752 Calhoun, V. D., Pearlson, G. D., & Sui, J. (2021). Data-driven Alhussein, M., Muhammad, G., & Hossain, M. S. (2019). EEG Pathol- approaches to neuroimaging biomarkers for neurological and ogy Detection Based on Deep Learning. IEEE Access, 7, 27781– psychiatric disorders: Emerging approaches and examples. Cur- 27788. https:// doi. org/ 10. 1109/ ACCESS. 2019. 29016 72 rent Opinion in Neurology, 34, 469–479. https://doi. or g/10. 1097/ Anand, C. S., & Sahambi, J. S. (2010). Wavelet domain non-linear WCO. 00000 00000 000967 filtering for MRI denoising. Magnetic Resonance Imaging, 28, Carré, A., Klausner, G., Edjlali, M., Lerousseau, M., Briend-Diop, J., 842–861. https:// doi. org/ 10. 1016/j. mri. 2010. 03. 013 Sun, R., Ammari, S., Reuzé, S., Alvarez Andres, E., Estienne, Arnal, P. J., Thorey, V., Debellemaniere, E., Ballard, M. E., Bou Hernandez, T., Niyoteka, S., Battistella, E., Vakalopoulou, M., Dhermain, F., A., Guillot, A., Jourde, H., Harris, M., Guillard, M., Van Beers, Paragios, N., Deutsch, E., Oppenheim, C., Pallud, J., & Robert, C. P., Chennaoui, M., & Sauvet, F. (2020). The Dreem Headband (2020). Standardization of brain MR images across machines and compared to polysomnography for electroencephalographic signal 1 3 Neuroinformatics (2022) 20:943–964 959 protocols: Bridging the gap for MRI-based radiomics. Science and Cole, J. H., Ritchie, S. J., Bastin, M. E., Valdés Hernández, M. C., Reports, 10, 12340. https:// doi. org/ 10. 1038/ s41598- 020- 69298-z Muñoz Maniega, S., Royle, N., Corley, J., Pattie, A., Harris, S. Cash, R. F. H., Weigand, A., Zalesky, A., Siddiqi, S. H., Downar, J., Fitzger- E., Zhang, Q., Wray, N. R., Redmond, P., Marioni, R. E., Starr, ald, P .B., & Fox, M D. (2020). Using Brain Imaging to Improve J. M., Cox, S. R., Wardlaw, J. M., Sharp, D. J., & Deary, I. J. Spatial Targeting of Transcranial Magnetic Stimulation for Depres- (2018). Brain age predicts mortality. Molecular Psychiatry, 23, sion Biological Psychiatry S0006322320316681. https:// doi. org/ 10. 1385–1392. https:// doi. org/ 10. 1038/ mp. 2017. 62 1016/j. biops ych. 2020. 05. 033 Collin, G., Nieto-Castanon, A., Shenton, M. E., Pasternak, O., Kelly, S., Cetin Karayumak, S., Bouix, S., Ning, L., James, A., Crow, T., Shenton, Keshavan, M. S., Seidman, L. J., McCarley, R. W., Niznikiewicz, M., et al. (2019). Retrospective harmonization of multi-site diffu- M. A., Li, H., Zhang, T., Tang, Y., Stone, W. S., Wang, J., & sion MRI data acquired with different acquisition parameters. Neuro Whitfield-Gabrieli, S. (2019). Brain functional connectivity data Image, 184, 180–200. https://d oi.o rg/1 0.1 016/j.n euroi mage.2 018.0 8. enhance prediction of clinical outcome in youth at risk for psy- 073. chosis. NeuroImage Clin., 26, 102108. https:// doi. org/ 10. 1016/j. Chai, X. J., Hirshfeld-Becker, D., Biederman, J., Uchida, M., Doehrmann, nicl. 2019. 102108 O., Leonard, J. A., et al. (2015). Functional and structural brain cor- Collin, G., Seidman, L. J., Keshavan, M. S., Stone, W. S., Qi, Z., Zhang, T., relates of risk for major depression in children with familial depres- Tang, Y., Li, H., Anteraper, S. A., Niznikiewicz, M. A., McCarley, sion. Neuro Image Clinical, 8, 398–407. https:// doi. org/ 10. 1016/j. R. W., Shenton, M. E., Wang, J., & Whitfield-Gabrieli, S. (2020). nicl. 2015. 05. 004. Functional connectome organization predicts conversion to psychosis Chang, K., Balachandar, N., Lam, C., Yi, D., Brown, J., Beers, A., in clinical high-risk youth from the SHARP program. Molecular Psy- Rosen, B., Rubin, D. L., & Kalpathy-Cramer, J. (2018). Dis- chiatry, 25, 2431–2440. https://doi. or g/10. 1038/ s41380- 018- 0288-x tributed deep learning networks among institutions for medical Contrepois, K., Wu, S., Moneghetti, K. J., Hornburg, D., Ahadi, S., Tsai, M.-S., imaging. Journal of the American Medical Informatics Associa- Metwally, A. A., Wei, E., Lee-McMullen, B., Quijada, J. V., Chen, S., tion, 25, 945–954. https:// doi. org/ 10. 1093/ jamia/ ocy017 Christle, J. W., Ellenberger, M., Balliu, B., Taylor, S., Durrant, M. G., Chang, K., Beers, A. L., Brink, L., Patel, J. B., Singh, P., Arun, N. T., Hoe- Knowles, D. A., Choudhry, H., Ashland, M., & Snyder, M. P. (2020). bel, K. V., Gaw, N., Shah, M., Pisano, E. D., Tilkin, M., Coombs, Molecular Choreography of Acute Exercise. Cell, 181, 1112-1130.e16. L. P., Dreyer, K. J., Allen, B., Agarwal, S., & Kalpathy-Cramer, J. https:// doi. org/ 10. 1016/j. cell. 2020. 04. 043 (2020). Multi-Institutional Assessment and Crowdsourcing Evalua- Cui, H., Giuliano, A. J., Zhang, T., Xu, L., Wei, Y., Tang, Y., Qian, tion of Deep Learning for Automated Classification of Breast Den- Z., Stone, L. M., Li, H., Whitfield-Gabrieli, S., Niznikiewicz, sity. Journal of the American College of Radiology, 17, 1653–1662. M., Keshavan, M. S., Shenton, M. E., Wang, J., & Stone, W. https:// doi. org/ 10. 1016/j. jacr. 2020. 05. 015 S. (2020). Cognitive dysfunction in a psychotropic medication- Chang, K., Zhang, B., Guo, X., Zong, M., Rahman, R., Sanchez, D., naïve, clinical high-risk sample from the ShangHai-At-Risk-for- et al. (2016). Multimodal imaging patterns predict survival in Psychosis (SHARP) study: Associations with clinical outcomes. recurrent glioblastoma patients treated with bevacizumab. Neuro- Schizophr. Res. Biomarkers in the Attenuated Psychosis Syn- Oncology, 18, 1680–1687. https:// doi. or g/ 10. 1093/ neuonc/ drome, 226, 138–146. https:// doi. org/ 10. 1016/j. schres. 2020. 06. now086. 018 Chang, P., Grinband, J., Weinberg, B. D., Bardis, M., Khy, M., Cadena, Dalca, A. V., Yu, E., Golland, P., Fischl, B., Sabuncu, M. R., & Iglesias, G., Su, M.-Y., Cha, S., Filippi, C. G., Bota, D., Baldi, P., Poisson, J. E. (2019). Unsupervised Deep Learning for Bayesian Brain L. M., Jain, R., & Chow, D. (2018). Deep-Learning Convolu- MRI Segmentation. ArXiv190411319v2. tional Neural Networks Accurately Classify Genetic Mutations in Davey, K., & Riehl, M. (2005). Designing transcranial magnetic stimu- Gliomas. American Journal of Neuroradiology, 39, 1201–1207. lation systems. IEEE Transactions on Magnetics, 41, 1142–1148. https:// doi. org/ 10. 3174/ ajnr. A5667https:// doi. org/ 10. 1109/ TMAG. 2004. 843326 Chen, I.Y., Pierson, E., Rose, S., Joshi, S., Ferryman, K., & Ghassemi, M. Despotović, I., Goossens, B., & Philips, W. (2015). MRI segmenta- (2020). Ethical Machine Learning in Health Care. ArXiv200910576 Cs. tion of the human brain: Challenges, methods, and applications. Chun, S. Y., Reese, T. G., Ouyang, J., Guerin, B., Catana, C., Zhu, X., Computational and Mathematical Methods in Medicine, 2015, et al. (2012). MRI-based nonrigid motion correction in simul- 450341. https:// doi. org/ 10. 1155/ 2015/ 450341 taneous PET/MRI. Journal Nuclear Medicine, 53, 1284–1291. Digital Health Center of Excellence Software as a Medical Device https:// doi. org/ 10. 2967/ jnumed. 111. 092353. (SaMD). FDA. (2021). https:// www. fda. gov/ medic al- devic es/ Chung, Y., Addington, J., Bearden, C. E., Cadenhead, K., Cornblatt, digit al- health- center- excel lence/ softw are- medic al- device- samd B., Mathalon, D. H., McGlashan, T., Perkins, D., Seidman, L.J., (Accessed 15 June 2021). Tsuang, M., Walker, E., Woods, S.W., McEwen, S., van Erp, T. Dinsdale, N. K., Jenkinson, M., & Namburete, A. I. L. (2021). Deep G. M., & Cannon, T. D. (2018). North American Prodrome Lon- learning-based unlearning of dataset bias for MRI harmonisation gitudinal Study (NAPLS) Consortium and the Pediatric Imaging, and confound removal. Neuro Image, 228,117689. https:// doi. Neurocognition, and Genetics (PING) Study Consortium. Use of org/ 10. 1016/j. neuro image. 2020. 117689 Machine Learning to Determine Deviance in Neuroanatomical Dong, X., Lei, Y., Wang, T., Higgins, K., Liu, T., Curran, W. J., Mao, H., Maturity Associated With Future Psychosis in Youths at Clini- Nye, J. A., & Yang, X. (2020). Deep learning-based attenuation cally High Risk. JAMA Psychiatry, 75, 960–968. https://doi. or g/ correction in the absence of structural information for whole-body 10. 1001/ jamap sychi atry. 2018. 1543. positron emission tomography imaging. Physics in Medicine & Cole, J. H., & Franke, K. (2017). Predicting Age Using Neuroimaging: Biology, 65, 055011. https:// doi. org/ 10. 1088/ 1361- 6560/ ab652c Innovative Brain Ageing Biomarkers. Trends in Neurosciences, Edupuganti, V., Mardani, M., Vasanawala, S., & Pauly, J. (2021). Uncer- 40, 681–690. https:// doi. org/ 10. 1016/j. tins. 2017. 10. 001 tainty Quantification in Deep MRI Reconstruction. IEEE Transac- Cole, J. H., Leech, R., & Sharp, D. J. (2015). Prediction of brain age tions on Medical Imaging, 40, 239–250. https:// doi. org/ 10. 1109/ suggests accelerated atrophy after traumatic brain injury. Annals TMI. 2020. 30250 65 of Neurology, 77, 571–581. https:// doi. org/ 10. 1002/ ana. 24367 Eickhoff, S., Nichols, T. E., Van Horn, J. D., & Turner, J. A. (2016). Cole, J. H., Marioni, R. E., Harris, S. E., & Deary, I. J. (2019). Brain Sharing the wealth: Neuroimaging data repositories. Neuro age and other bodily “ages”: Implications for neuropsychiatry. Image, 124, 1065–1068. https:// doi. org/ 10. 1016/j. neuro image. 2015. 10. 079. Molecular Psychiatry, 24, 266–281. https:// doi. or g/ 10. 1038/ s41380- 018- 0098-1 1 3 960 Neuroinformatics (2022) 20:943–964 Engemann, D. A., Raimondo, F., King, J. -R., Rohaut, B., Louppe, image-domain metal artifact reduction, in: Developments in G., Faugeras, F., et al. (2018). Robust EEG-based cross-site and X-Ray Tomography XI. Presented at the Developments in X-Ray cross-protocol classification of states of consciousness. Brain, Tomography XI, International Society for Optics and Photonics, 141, 3179–3192. https:// doi. org/ 10. 1093/ brain/ awy251. 103910W. https:// doi. org/ 10. 1117/ 12. 22744 27 Escudero, J., Abásolo, D., Hornero, R., Espino, P., & López, M. (2006). Greenspan, H., Tanno, R., Erdt, M., Arbel, T., Baumgartner, C., Dalca, Analysis of electroencephalograms in Alzheimer’s disease patients A., Sudre, C. H., Wells, W. M., Drechsler, K., & Linguraru, M. with multiscale entropy. Physiological Measurement, 27, 1091– G. (2019). Uncertainty for Safe Utilization of Machine Learn- 1106. https:// doi. org/ 10. 1088/ 0967- 3334/ 27/ 11/ 004 ing in Medical Imaging and Clinical Image-Based Procedures: Esteban, O., Birman, D., Schaer, M., Koyejo, O. O., Poldrack, R. A., First International Workshop, UNSURE 2019, and 8th Interna- & Gorgolewski, K. J. (2017). MRIQC: Advancing the automatic tional Workshop, CLIP 2019, Held in Conjunction with MICCAI prediction of image quality in MRI from unseen sites. PloS One, 2019, Shenzhen, China, October 17, 2019, Proceedings Springer 12,. Nature. Esteban, O., Blair, R. W., Nielson, D. M., Varada, J. C., Marrett, Griswold, M. A., Jakob, P. M., Heidemann, R. M., Nittka, M., Jellus, S., Thomas, A. G., et al. (2019). Crowdsourced MRI quality V., Wang, J., Kiefer, B., & Haase, A. (2002). Generalized auto- metrics and expert quality annotations for training of humans calibrating partially parallel acquisitions (GRAPPA). Magnetic and machines. Science Data, 6, 1–7. https:// doi. org/ 10. 1038/ Resonance in Medicine, 47, 1202–1210. https://doi. or g/10. 1002/ s41597- 019- 0035-4mrm. 10171 Fair ML for Health - Accepted Papers. (2021). https://www .f airmlf orhealt h. Guggenmos, M., Schmack, K., Sekutowicz, M., Garbusow, M., com/ accep ted- papers (Accessed 28 July 2021). Sebold, M., Sommer, C., et al. (2017). Quantitative neurobio- FDA-NIH Biomarker Working Group. (2016). BEST (Biomarkers, logical evidence for accelerated brain aging in alcohol depend- EndpointS, and other Tools) Resource. Food and Drug Admin- ence. Translational Psychiatry, 7, 1–7. https:// doi. org/ 10. 1038/ istration (US), Silver Spring (MD).s41398- 017- 0037-y. Filippi, M., Horsfield, M.A., Bressi, S., Martinelli, V., Baratti, C., Reganati, Guimond, A., Meunier, J., & Thirion, J. -P. (2000). Average Brain Models: P., Campi, A., Miller, D.H., & Comi, G. (1995). Intra- and inter- A Convergence Study. Computer Vision and Image Understanding, observer agreement of brain MRI lesion volume measurements in 77, 192–210. https:// doi. org/ 10. 1006/ cviu. 1999. 0815. multiple sclerosis. A comparison of techniques. Brain Journal Neu- Haehn, D., Rannou, D., Ahtam, B., Grant, P., & Pienaar, R. (2014). rology, 118(Pt 6), 1593–1600. https:// doi. org/ 10. 1093/ brain/ 118.6. Neuroimaging in the Browser using the X Toolkit Front. Neuro- 1593 informatics 8. https://doi. or g/10. 3389/ conf. fninf. 2014. 08. 00101 Fischl, B., Salat, D. H., Busa, E., Albert, M., Dieterich, M., Haselgrove, Hammernik, K., Klatzer, T., Kobler, E., Recht, M. P., Sodickson, D. C., van der Kouwe, A., Killiany, R., Kennedy, D., Klaveness, K., Pock, T., & Knoll, F. (2018). Learning a variational network S., Montillo, A., Makris, N., Rosen, B., & Dale, A. M. (2002). for reconstruction of accelerated MRI data. Magnetic Resonance Whole brain segmentation: Automated labeling of neuroanatomi- in Medicine, 79, 3055–3071. https://doi. or g/10. 1002/ mr m.26977 cal structures in the human brain. Neuron, 33, 341–355. https:// Han, Y. S., Yoo, J., Ye, J. C. (2018). Deep Learning with Domain doi. org/ 10. 1016/ s0896- 6273(02) 00569-x Adaptation for Accelerated Projection-Reconstruction MR. Franke, K., Gaser, C., Manor, B., & Novak, V. (2013). Advanced ArXiv170301135 Cs. http:// arxiv. org/ abs/ 1703. 01135 BrainAGE in older adults with type 2 diabetes mellitus. Fron- Haskell, M. W., Cauley, S. F., & Wald, L. L. (2018). Targeted Motion tiers Aging Neuroscience 5 https:// doi. org/ 10. 3389/ fnagi. 2013. Estimation and Reduction (TAMER): Data Consistency Based 00090 Motion Mitigation for MRI using a Reduced Model Joint Optimi- Franke, L., & Haehn, D. (2020). Modern Scientific Visualizations on the zation. IEEE Transactions on Medical Imaging, 37, 1253–1265. Web. Informatics, 7, 37. https://doi. or g/10. 3390/ inf ormatic s7040 037 https:// doi. org/ 10. 1109/ TMI. 2018. 27914 82 Franke, L., Weidele, D. K. I., Zhang, F., Cetin-Karayumak, S., Pieper, He, S., Gollub, R. L., Murphy, S. N., Perez, J. D., Prabhu, S., Pienaar, S., O’Donnell, L. J., Rathi, Y., & Haehn, D. (2020). FiberStars: R., et al. (2020). Brain Age Estimation Using LSTM on Children’s Visual Comparison of Diffusion Tractography Data between Brain MRI. Proceeding IEEE International Symposium Biomedical Multiple Subjects. ArXiv200508090 Cs. Imaging, 2020, 420–423. https:// doi. org/ 10. 1109/ isbi4 5749. 2020. Gajawelli, N., Tsao, S., Kromnick, M., Nelson, M., & Leporé, N. 90983 56. (2019). Image Postprocessing Adoption Trends in Clinical Medi- He, S., Pereira, D., David Perez, J., Gollub, R. L., Murphy, S. N., Prabhu, S., cal Imaging. Journal of the American College of Radiology, 16, Pienaar, R., Robertson, R. L., Ellen, Grant, P., & Ou Y. (2021). Multi- 945–951. https:// doi. org/ 10. 1016/j. jacr. 2019. 01. 005 channel Attention-Fusion Neural Network for Brain Age Estimation: Gallego-Jutglà, E., Solé-Casals, J., Vialatte, F. -B., Elgendi, M., Accuracy, Generality, and Interpretation with 16,705 Healthy MRIs Cichocki, A., & Dauwels, J. (2015). A hybrid feature selection across. Lifespan Medical Image Analysis 102091 https:// doi. org/ 10. approach for the early diagnosis of Alzheimer’s disease. Journal 1016/j. media. 2021. 102091 of Neural Engineering, 12,016018. https://doi. or g/10. 1088/ 1741- Hoebel, K. V., Patel, J. B., Beers, A. L., Chang, K., Singh, P., Brown, J. M., 2560/ 12/1/ 016018 et al. (2021). Radiomics Repeatability Pitfalls in a Scan-Rescan MRI Gemein, L. A. W., Schirrmeister, R. T., Chrabąszcz, P., Wilson, D., Study of Glioblastoma. Radiology Artificial Intelligence, 3,. https:// Boedecker, J., Schulze-Bonhage, A., Hutter, F., & Ball, T. doi. org/ 10. 1148/ ryai. 20201 90199 (2020). Machine-learning-based diagnostics of EEG pathology. Hofmeister, J., Bernava, G., Rosi, A., Vargas, M. I., Carrera, E., Montet, NeuroImage, 220, 117021. https://doi. or g/10. 1016/j. neur oimag e. X., Buergermeister, S., Poletti, P. -A., Platon, A., Lovblad, K -O., 2020. 117021 & Machi, P. (2020). Clot-Based Radiomics Predict a Mechanical Ghassemi, M. M., Moody, B. E., Lehman, L. -W. H., Song, C., Li, Q., Thrombectomy Strategy for Successful Recanalization in Acute Sun, H., Mark, R. G., Westover, M. B., & Clifford, G. D. (2018). Ischemic Stroke. Stroke, 51, 2488–2494. https:// doi. org/ 10. 1161/ You Snooze, You Win: the PhysioNet/Computing in Cardiology STROK EAHA. 120. 030334. Challenge 2018, in: 2018 Computing in Cardiology Conference Hogan, J., Sun, H., Paixao, L., Westmeijer, M., Sikka, P., Jin, J., Tesh, R., (CinC). Presented at the 2018 Computing in Cardiology Confer- Cardoso, M., Cash, S. S., Akeju, O., Thomas, R., & Westover, M. B. ence (CinC), pp. 1–4. https:// doi. org/ 10. 22489/ CinC. 2018. 049 (2021). Night-to-night variability of sleep electroencephalography- based brain age measurements. Clinical Neurophysiology, 132, 1–12. Gjesteby, L., Yang, Q., Xi, Y., Shan, H., Claus, B., Jin, Y., Man, https:// doi. org/ 10. 1016/j. clinph. 2020. 09. 029 B. D., & Wang, G. (2017). Deep learning methods for CT 1 3 Neuroinformatics (2022) 20:943–964 961 Hosseini, M.-P., Hemingway, C., Madamba, J., McKee, A., Ploof, N., Küstner, T., Gatidis, S., Liebgott, A., Schwartz, M., Mauch, L., Martirosian, Schuman, J., & Voss, E. (2020). Review of Machine Learning Algo- P., Schmidt, H., Schwenzer, N. F., Nikolaou, K., Bamberg, F., Yang, rithms for Brain Stroke Diagnosis and Prognosis by EEG Analysis. B., & Schick, F. (2018). A machine-learning framework for auto- ArXiv200808118 Cs Eess. http:// arxiv. org/ abs/ 2008. 08118 matic reference-free quality assessment in MRI. Magnetic Resonance Hu, Z., Jiang, C., Sun, F., Zhang, Q., Ge, Y., Yang, Y., Liu, X., Zheng, Imaging, 53, 134–147. https:// doi. org/ 10. 1016/j. mri. 2018. 07. 003 H., & Liang, D. (2019). Artifact correction in low-dose dental CT LaConte, S. M., Peltier, S. J., & Hu, X. P. (2007). Real-time fMRI imaging using Wasserstein generative adversarial networks. Medi- using brain-state classification. Human Brain Mapping, 28, cal Physics, 46, 1686–1696. https:// doi. org/ 10. 1002/ mp. 13415 1033–1044. https:// doi. org/ 10. 1002/ hbm. 20326 Iglesias, J. E., Billot, B., Balbastre, Y., Tabari, A., Conklin, J., Alexander, Ladefoged, C. N., Marner, L., Hindsholm, A., Law, I., Højgaard, L., & D. C., Golland, P., Edlow, B. L., & Fischl, B. (2020). Joint super- Andersen, F. L. (2018). Deep Learning Based Attenuation Cor- resolution and synthesis of 1 mm isotropic MP-RAGE volumes from rection of PET/MRI in Pediatric Brain Tumor Patients: Evalua- clinical MRI exams with scans of different orientation, resolution and tion in a Clinical Setting. Frontiers in Neuroscience, 12, 1005. contrast. ArXiv201213340 Cs Eess.https:// doi. org/ 10. 3389/ fnins. 2018. 01005 Irwin, M. R. (2019). Sleep and inflammation: Partners in sickness and Lao, J., Chen, Y., Li, Z.-C., Li, Q., Zhang, J., Liu, J., & Zhai, G. (2017). in health. Nature Reviews Immunology, 19, 702–715. https://d oi. A Deep Learning-Based Radiomics Model for Prediction of org/ 10. 1038/ s41577- 019- 0190-z Survival in Glioblastoma Multiforme. Science and Reports, 7, Jelles, B., van Birgelen, J. H., Slaets, J. P. J., Hekster, R. E. M., Jonkman, 10353. https:// doi. org/ 10. 1038/ s41598- 017- 10649-8 E. J., & Stam, C. J. (1999). Decrease of non-linear structure in the Ledoux, L -P., Morency, F. C., Cousineau, M., Houde, J-C., Whittingstall, EEG of Alzheimer patients compared to healthy controls. Clinical K., & Descoteaux, M. (2017). Fiberweb. Diffusion Visualization and Neurophysiology, 110, 1159–1167. https:// doi. org/ 10. 1016/ S1388- Processing in the Browser Frontiers Neuroinformatics, 11. https:// 2457(99) 00013-9doi. org/ 10. 3389/ fninf. 2017. 00054 Jönsson, D., Bergström, A., Forsell, C., Simon, R., Engström, M., Lee, J. M., Akeju, O., Terzakis, K., Pavone, K. J., Deng, H., Houle, Ynnerman, A., & Hotz, I. (2019). A Visual Environment for T. T., Firth, P. G., Shank, E. S., Brown, E. N., & Purdon, P. Hypothesis Formation and Reasoning in Studies with fMRI L. (2017). A Prospective Study of Age-dependent Changes in and Multivariate Clinical Data. The Eurographics Association. Propofol-induced Electroencephalogram Oscillations in Chil- https:// doi. org/ 10. 2312/ vcbm. 20191 232 dren. Anesthesiology, 127, 293–306. https:// doi. org/ 10. 1097/ Kalpathy-Cramer, J., Mamomov, A., Zhao, B., Lu, L., Cherezov, D., ALN. 00000 00000 001717 Napel, S., Echegaray, S., Rubin, D., McNitt-Gray, M., Lo, P., Lehmann, C., Koenig, T., Jelic, V., Prichep, L., John, R. E., Wahlund, L. Sieren, J.C., Uthoff, J., Dilger, S.K.N., Driscoll, B., Yeung, -O., et al. (2007). Application and comparison of classification algo- I., Hadjiiski, L., Cha, K., Balagurunathan, Y., Gillies, R., & rithms for recognition of Alzheimer’s disease in electrical brain activ- Goldgof, D. O (2016). Radiomics of Lung Nodules: A Multi- ity (EEG). Journal of Neuroscience Methods, 161, 342–350. https:// Institutional Study of Robustness and Agreement of Quantita-doi. org/ 10. 1016/j. jneum eth. 2006. 10. 023. tive Imaging Features. Tomogrography Ann Arbor Michigan, Leone, M. J., Sun, H., Boutros, C. L., Liu, L., Ye, E., Sullivan, L., 2(430–437). https:// doi. org/ 10. 18383/j. tom. 2016. 00235 Thomas, R. J., Robbins, G. K., Mukerji, S. S., & Westover, Kamnitsas, K., Baumgartner, C., Ledig, C., Newcombe, V. F. J., Simpson, M. B. (2021). HIV Increases Sleep-based Brain Age Despite J. P., Kane, A. D., Menon, D. K., Nori, A., Criminisi, A., Rueckert, Antiretroviral Therapy. Sleep zsab058. https://d oi.o rg/1 0.1 093/ D., & Glocker, B. (2016). Unsupervised domain adaptation in brain sleep/ zsab0 58 lesion segmentation with adversarial networks. ArXiv161208894 Cs. Liem, F., Varoquaux, G., Kynast, J., Beyer, F., Kharabian Masouleh, http:// arxiv. org/ abs/ 1612. 08894 S., Huntenburg, J. M., Lampe, L., Rahim, M., Abraham, A., Karch, J. D., Filevich, E., Wenger, E., Lisofsky, N., Becker, M., Butler, Craddock, R. C., Riedel-Heller, S., Luck, T., Loeffler, M., O., Mårtensson, J., Lindenberger, U., Brandmaier, A. M., & Kühn, Schroeter, M. L., Witte, A. V., Villringer, A., & Margulies, D. S. (2019). Identifying predictors of within-person variance in MRI- S. (2017). Predicting brain-age from multimodal imaging data based brain volume estimates. NeuroImage, 200, 575–589. https:// captures cognitive impairment. NeuroImage, 148, 179–188. doi. org/ 10. 1016/j. neuro image. 2019. 05. 030https:// doi. org/ 10. 1016/j. neuro image. 2016. 11. 005 Kaufmann, T., van der Meer, D., Doan, N. T., Schwarz, E., Lund, M. J., Liu, F., Jang, H., Kijowski, R., Bradshaw, T., & McMillan, A. B. Agartz, I., Alnæs, D., Barch, D. M., Baur-Streubel, R., Bertolino, (2018a). Deep Learning MR Imaging-based Attenuation A., Bettella, F., Beyer, M. K., Bøen, E., Borgwardt, S., Brandt, C. Correction for PET/MR Imaging. Radiology, 286, 676–684. L., Buitelaar, J., Celius, E. G., Cervenka, S., Conzelmann, A., & https:// doi. org/ 10. 1148/ radiol. 20171 70700 Westlye, L. T. (2019). Common brain disorders are associated with Liu, F., Jang, H., Kijowski, R., Zhao, G., Bradshaw, T., & McMillan, heritable patterns of apparent aging of the brain. Nature Neurosci- A. B. (2018b). A deep learning approach for 18F-FDG PET ence, 22, 1617–1623. https:// doi. org/ 10. 1038/ s41593- 019- 0471-7 attenuation correction. EJNMMI Physics, 5, 24. https://d oi.o rg/ Keshavan, A., Yeatman, J. D., & Rokem, A. (2019). Combining Citizen 10. 1186/ s40658- 018- 0225-8. Science and Deep Learning to Amplify Expertise in Neuroimag- Lorenz, R., Monti, R. P., Violante, I. R., Anagnostopoulos, C., Faisal, ing. Frontiers Neuroinformatics, 13, 29. https://doi. or g/10. 3389/ A. A., Montana, G., & Leech, R. (2016). The Automatic Neu- fninf. 2019. 00029. roscientist: A framework for optimizing experimental design Kniep, H. C., Madesta, F., Schneider, T., Hanning, U., Schönfeld, M. with closed-loop real-time fMRI. NeuroImage, 129, 320–334. H., Schön, G., Fiehler, J., Gauer, T., Werner, R., & Gellissen, S. https:// doi. org/ 10. 1016/j. neuro image. 2016. 01. 032 (2019). Radiomics of Brain MRI: Utility in Prediction of Meta- Luders, E., Cherbuin, N., & Gaser, C. (2016). Estimating brain age static Tumor Type. Radiology, 290, 479–487. https://doi. or g/10. using high-resolution pattern recognition: Younger brains in 1148/ radiol. 20181 80946 long-term meditation practitioners. NeuroImage, 134, 508– Kucyi, A., Esterman, M., Capella, J., Green, A., Uchida, M., Bieder- 513. https:// doi. org/ 10. 1016/j. neuro image. 2016. 04. 007 man, J., Gabrieli, J. D. E., Valera, E. M., & Whitfield-Gabrieli, Lustig, M., Donoho, D. L., Santos, J. M., & Pauly, J. M. (2008). S. (2021). Prediction of stimulus-independent and task-unrelated Compressed Sensing MRI. IEEE Signal Processing Magazine, thought from functional brain networks. Nature Communications, 25, 72–82. https:// doi. org/ 10. 1109/ MSP. 2007. 914728 12, 1793. https:// doi. org/ 10. 1038/ s41467- 021- 22027-0 1 3 962 Neuroinformatics (2022) 20:943–964 Macyszyn, L., Akbari, H., Pisapia, J. M., Da, X., Attiah, M., Pigrish, V., of Aging, 88, 150–155. https://doi. or g/10. 1016/j. neur obiola ging. et al. (2016). Imaging patterns predict patient survival and molecular 2019. 12. 015 subtype in glioblastoma via machine learning techniques. Neuro- Pan, C.-C., Liu, J., Tang, J., Chen, X., Chen, F., Wu, Y.-L., et al. (2019). Oncology, 18, 417–425. https:// doi. org/ 10. 1093/ neuonc/ nov127. A machine learning-based prediction model of H3K27M muta- Manjon, J. V., & Coupe, P. (2019). MRI denoising using Deep Learn- tions in brainstem gliomas using conventional MRI and clinical ing and Non-local averaging. ArXiv191104798 Math. features. Radiotherapy Oncology, 130, 172–179. https://doi. or g/ Marcadent, S., Hofmeister, J., Preti, M. G., Martin, S. P., Van De 10. 1016/j. radonc. 2018. 07. 011. Ville, D., & Montet, X. (2020). Generative Adversarial Net- Parmar, C., Rios Velazquez, E., Leijenaar, R., Jermoumi, M., Carvalho, works Improve the Reproducibility and Discriminative Power S., Mak, R. H., Mitra, S., Shankar, B. U., Kikinis, R., Haibe-Kains, of Radiomic Features. Radiology Artificial Intelligence, 2, B., Lambin, P., & Aerts, H. J. W. L. (2014). Robust Radiomics e190035. https:// doi. org/ 10. 1148/ ryai. 20201 90035 feature quantification using semiautomatic volumetric segmenta- Mateos-Pérez, J. M., Dadar, M., Lacalle-Aurioles, M., Iturria- tion. PloS One, 9, e102107. https:// doi. org/ 10. 1371/ journ al. pone. Medina, Y., Zeighami, Y., & Evans, A. C. (2018). Structural 01021 07 neuroimaging as clinical predictor: A review of machine learn- Phang, C. -R., Noman, F., Hussain, H., Ting, C. -M., & Ombao, H. (2020). ing applications. NeuroImage Clinical, 20, 506–522. https:// A Multi-Domain Connectome Convolutional Neural Network for doi. org/ 10. 1016/j. nicl. 2018. 08. 019 Identifying Schizophrenia From EEG Connectivity Patterns. IEEE Merikanto, I., Utge, S., Lahti, J., Kuula, L., Makkonen, T., Lahti‐ Journal of Biomedical and Health Informatics, 24, 1333–1343. Pulkkinen, M., & Pesonen, A. K. (2019). Genetic risk fac-https:// doi. org/ 10. 1109/ JBHI. 2019. 29412 22. tors for schizophrenia associate with sleep spindle activity in Pinto, A. L. R., Ou, Y., Sahin, M., & Grant, P. E. (2018). Quantitative healthy adolescents. Journal of Sleep Research, 28 https://doi. Apparent Diffusion Coefficient Mapping May Predict Seizure org/ 10. 1111/ jsr. 12762 Onset in Children With Sturge-Weber Syndrome. Pediatric Michel, C. M., & Murray, M. M. (2012). Towards the utilization of EEG Neurology, 84, 32–38. https:// doi. org/ 10. 1016/j. pedia trneu rol. as a brain imaging tool. NeuroImage, 61, 371–385. https:// doi. org/ 2018. 04. 004 10. 1016/j. neuro image. 2011. 12. 039 Pipe, J. G. (1999). Motion correction with PROPELLER MRI: Applica- Miranda, P., D Cox, C., Alexander, M., Danev, S., & R. T., Lakey, tion to head motion and free-breathing cardiac imaging. Magnetic J., (2019). Overview of current diagnostic, prognostic, and Resonance in Medicine, 42, 963–969. https://doi. or g/10. 1002/ (sici) therapeutic use of EEG and EEG-based markers of cognition, 1522- 2594(199911) 42:5% 3c963:: aid- mrm17% 3e3.0. co;2-l mental, and brain health. Integrative Molecular. Medicine, 6. Pizarro, R. A., Cheng, X., Barnett, A., Lemaitre, H., Verchinski, B. https:// doi. org/ 10. 15761/ IMM. 10003 78 A., Goldman, A. L., Xiao, E., Luo, Q., Berman, K. F., Callicott, Mohajer, B., Abbasi, N., Mohammadi, E., Khazaie, H., Osorio, R. S., J. H., Weinberger, D. R., & Mattay, V. S. (2016). Automated Rosenzweig, I., Eickhoff, C. R., Zarei, M., Tahmasian, M., & Quality Assessment of Structural Magnetic Resonance Brain Eickhoff, S. B. (2020). Gray matter volume and estimated brain Images Based on a Supervised Machine Learning Algorithm. age gap are not linked with sleep-disordered breathing. Human Front. Neuroinformatics, 10, 52. https:// doi. org/ 10. 3389/ fninf. Brain Mapping, 41, 3034–3044. https:// doi. org/ 10. 1002/ hbm. 2016. 00052 24995 Poddar, J., Pradhan, M., Ganguly, G., & Chakrabarti, S. (2019). Bio- Moyer, D., Ver Steeg, G., Tax, C. M. W., & Thompson, P. M. (2020). chemical deficits and cognitive decline in brain aging: Interven- Scanner invariant representations for diffusion MRI harmoniza- tion by dietary supplements. Journal of Chemical Neuroanatomy, tion. Magnetic Resonance in Medicine, 84, 2174–2189. https:// 95, 70–80. https:// doi. org/ 10. 1016/j. jchem neu. 2018. 04. 002 doi. org/ 10. 1002/ mrm. 28243 Provenzale, J. M., Ison, C., & Delong, D. (2009). Bidimensional meas- Ning, K., Zhao, L., Matloff, W., Sun, F., & Toga, A. W. (2020). Asso- urements in brain tumors: Assessment of interobserver variabil- ciation of relative brain age with tobacco smoking, alcohol con- ity. AJR. American Journal of Roentgenology, 193, W515-522. sumption, and genetic variants. Science and Reports, 10, 10. https:// doi. org/ 10. 2214/ AJR. 09. 2615 https:// doi. org/ 10. 1038/ s41598- 019- 56089-4 Provenzale, J. M., & Mancini, M. C. (2012). Assessment of intra- Nishimura, D. G. (2010). Principles of magnetic resonance imaging. observer variability in measurement of high-grade brain tumors. Self-Published. Journal of Neuro-Oncology, 108, 477–483. https:// doi. org/ 10. O’Muircheartaigh, J., Robinson, E. C., Pietsch, M., Wolfers, T., Aljabar, 1007/ s11060- 012- 0843-2 P., Grande, L. C., Teixeira, R. P. A. G., Bozek, J., Schuh, A., Pruessmann, K. P., Weiger, M., Scheidegger, M. B., & Boesiger, P. Makropoulos, A., Batalle, D., Hutter, J., Vecchiato, K., Steinweg, J. (1999). SENSE: Sensitivity encoding for fast MRI. Magnetic K., Fitzgibbon, S., Hughes, E., Price, A. N., Marquand, A., Reuckert, Resonance in Medicine, 42, 952–962. D., & Edwards, A. D. (2020). Modelling brain development to detect Purdon, P. L., Pavone, K. J., Akeju, O., Smith, A. C., Sampson, A. L., white matter injury in term and preterm born neonates. Brain, 143, Lee, J., Zhou, D. W., Solt, K., & Brown, E. N. (2015). The Ageing 467–479. https:// doi. org/ 10. 1093/ brain/ awz412 Brain: Age-dependent changes in the electroencephalogram dur- Orlhac, F., Boughdad, S., Philippe, C., Stalla-Bourdillon, H., Nioche, ing propofol and sevoflurane general anaesthesia. British Journal C., Champion, L., Soussan, M., Frouin, F., Frouin, V., & Buvat, I. of Anaesthesia, 115, i46–i57. https:// doi. org/ 10. 1093/ bja/ aev213 (2018). A Postreconstruction Harmonization Method for Multicenter Putzky, P., Karkalousos, D., Teuwen, J., Miriakov, N., Bakker, B., Radiomic Studies in PET. Journal of Nuclear Medicine, 59, 1321– Caan, M., & Welling, M. (2019). i-RIM applied to the fastMRI 1328. https:// doi. org/ 10. 2967/ jnumed. 117. 199935 challenge. ArXiv191008952. Ou, Y., Zöllei, L., Retzepi, K., Castro, V., Bates, S. V., Pieper, S., Quan, T. M., Nguyen-Duc, T., & Jeong, W.-K. (2018). Compressed Andriole, K. P., Murphy, S. N., Gollub, R. L., & Grant, P. E. Sensing MRI Reconstruction Using a Generative Adversarial Net- (2017). Using clinically acquired MRI to construct age-specific work With a Cyclic Loss. IEEE Transactions on Medical Imaging, ADC atlases: Quantifying spatiotemporal ADC changes from 37, 1488–1497. https:// doi. org/ 10. 1109/ TMI. 2018. 28201 20 birth to 6-year old. Human Brain Mapping, 38, 3052–3068. Ramm, A. G., & Katsevich, A. I. (1996). The Radon Transform and https:// doi. org/ 10. 1002/ hbm. 23573 Local Tomography. CRC Press. Paixao, L., Sikka, P., Sun, H., Jain, A., Hogan, J., Thomas, R., & Rauschecker, A. M., Rudie, J. D., Xie, L., Wang, J., Duong, M. T., Botzolakis, E. J., Kovalovich, A. M., Egan, J., Cook, T. C., Bryan, Westover, M. B. (2020). Excess brain age in the sleep electro- R. N., Nasrallah, I. M., Mohan, S., & Gee, J. C. (2020). Artificial encephalogram predicts reduced life expectancy. Neurobiology 1 3 Neuroinformatics (2022) 20:943–964 963 Intelligence System Approaching Neuroradiologist-level Differen- Sun, H., Paixao, L., Oliva, J. T., Goparaju, B., Carvalho, D. Z., van tial Diagnosis Accuracy at Brain MRI. Radiology, 295, 626–637. Leeuwen, K. G., Akeju, O., Thomas, R. J., Cash, S. S., Bianchi, https:// doi. org/ 10. 1148/ radiol. 20201 90283 M. T., & Westover, M. B. (2019). Brain age from the electro- Regenhardt, R. W., Bretzner, M., Zanon Zotin, M. C., Bonkhoff, A. K., encephalogram of sleep. Neurobiology of Aging, 74, 112–120. Etherton, M. R., Hong, S., Das, A. S., Alotaibi, N. M., Vranic, J. https:// doi. org/ 10. 1016/j. neuro biola ging. 2018. 10. 016 E., Dmytriw, A. A., Stapleton, C. J., Patel, A. B., Kuchcinski, G., Tang, T., Jiao, Y., Cui, Y., Zhao, D., Zhang, Y., Wang, Z., Meng, Rost, N. S., & Leslie-Mazwi, T. M. (2021). Radiomic signature of X., Yin, X.-D., Yang, Y.-J., Teng, G., & Ju, S. (2020). Penum- DWI-FLAIR mismatch in large vessel occlusion stroke. Journal of bra-based radiomics signature as prognostic biomarkers for Neuroimaging. https:// doi. org/ 10. 1111/ jon. 12928 thrombolysis of acute ischemic stroke patients: A multicenter Rogenmoser, L., Kernbach, J., Schlaug, G., & Gaser, C. (2018). Keep- cohort study. Journal of Neurology, 267, 1454–1463. https:// ing brains young with making music. Brain Structure & Func-doi. org/ 10. 1007/ s00415- 020- 09713-7 tion, 223, 297–305. https:// doi. org/ 10. 1007/ s00429- 017- 1491-2 Tanioka, S., Ishida, F., Yamamoto, A., Shimizu, S., Sakaida, H., Toyoda, Roy, S., Kiral-Kornek, I., & Harrer, S. (2019a). ChronoNet: A Deep M., et al. (2020). Machine Learning Classification of Cerebral Recurrent Neural Network for Abnormal EEG Identification, in: Aneurysm Rupture Status with Morphologic Variables and Hemo- Riaño, D., Wilk, S., ten Teije, A. (Eds.), Artificial Intelligence in dynamic Parameters. Radiology Artificial Intelligence, 2, e190077. Medicine, Lecture Notes in Computer Science. Springer Inter-https:// doi. org/ 10. 1148/ ryai. 20191 90077 national Publishing, Cham, pp. 47–56. https:// doi. org/ 10. 1007/ Temple University EEG Corpus (2021). https://www .isip. picon epr ess.com/ 978-3- 030- 21642-9_8proje cts/ tuh_ eeg/ html/ downl oads. shtml (Accessed 29 Apr 2021). Roy, Y., Banville, H., Albuquerque, I., Gramfort, A., Falk, T. H., & Titano, J. J., Badgeley, M., Schefflein, J., Pain, M., Su, A., Cai, M., Swin- Faubert, J. (2019b). Deep learning-based electroencephalog- burne, N., Zech, J., Kim, J., Bederson, J., Mocco, J., Drayer, B., raphy analysis: A systematic review. Journal of Neural Engi- Lehar, J., Cho, S., Costa, A., & Oermann, E. K. (2018). Automated neering, 16, 051001. https://doi. or g/10. 1088/ 1741- 2552/ ab260c deep-neural-network surveillance of cranial images for acute neu- Schlemper, J., Caballero, J., Hajnal, J.V., Price, A., & Rueckert, D. rologic events. Nature Medicine, 24, 1337–1341. https:// doi. org/ 10. (2018). A Deep Cascade of Convolutional Neural Networks 1038/ s41591- 018- 0147-y for Dynamic MR Image Reconstruction. IEEE Translational Tzimourta, K. D., Christou, V., Tzallas, A. T., Giannakeas, N., Astrakas, Medicine Imaging 491–503. L. G., Angelidis, P., & Tsipouras, M. G. (2021). Machine Learn- Schwier, M., van Griethuysen, J., Vangel, M. G., Pieper, S., Peled, S., ing Algorithms and Statistical Approaches for Alzheimer’s Disease Tempany, C., Aerts, H. J. W. L., Kikinis, R., Fennessy, F. M., & Analysis Based on Resting-State EEG Recordings. A Systematic Fedorov, A. (2019). Repeatability of Multiparametric Prostate Review International Journal of Neural Systems, 2130002. https:// MRI Radiomics Features. Science and Reports, 9, 9441. https:// doi. org/ 10. 1142/ S0129 06572 13000 23 doi. org/ 10. 1038/ s41598- 019- 45766-z van Horn, N., Kniep, H., Broocks, G., Meyer, L., Flottmann, F., Bechstein, Si, Y. (2020). Machine learning applications for electroencephalograph M., Götz, J., Thomalla, G., Bendszus, M., Bonekamp, S., Pfaff, J. signals in epilepsy: A quick review. Acta Epileptologica, 2, 5. A. R., Dellani, P. R., Fiehler, J., & Hanning, U. (2021). ASPECTS https:// doi. org/ 10. 1186/ s42494- 020- 00014-0. Interobserver Agreement of 100 Investigators from the TEN- Singh, N. M., Iglesias, J. E., Adalsteinsson, E., Dalca, A. V., & Golland, SION. Study Clinical Neuroradiology. https:// doi. org/ 10. 1007/ P. (2020). Joint Frequency and Image Space Learning for Fourier s00062- 020- 00988-x Imaging. ArXiv200701441 Cs Eess. van der Maaten, L., & Hinton, G. (2008). Visualizing Data using t-SNE. Sleep Data - National Sleep Research Resource – NSRR (2021). https:// Journal of Machine Learning Research, 9, 2579–2605. sleep data. org/ (Accessed Apr 29 2021). van Leeuwen, K. G., Sun, H., Tabaeizadeh, M., Struck, A. F., van Putten, Sørensen, L., Igel, C., Liv Hansen, N., Osler, M., Lauritzen, M., M. J. A. M., & Westover, M. B. (2019). Detecting abnormal electro- Rostrup, E., Nielsen, M., Alzheimer’s Disease Neuroimaging encephalograms using deep convolutional networks. Clinical Neuro- Initiative and the Australian Imaging Biomarkers and Lifestyle physiology, 130, 77–84. https:// doi. org/ 10. 1016/j. clinph. 2018. 10. 012 Flagship Study of Ageing. (2016). Early detection of Alzhei- Varikuti, D. P., Genon, S., Sotiras, A., Schwender, H., Hoffstaedter, F., Patil, mer’s disease using MRI hippocampal texture. Human Brain K. R., Jockwitz, C., Caspers, S., Moebus, S., Amunts, K., Davatzikos, Mapping, 37, 1148–1161. https:// doi. org/ 10. 1002/ hbm. 23091 C., & Eickhoff, S. B. (2018). Evaluation of non-negative matrix fac- Sotardi, S., Gollub, R. L., Bates, S. V., Weiss, R., Murphy, S. N., Grant, torization of grey matter in age prediction. NeuroImage, 173, 394– P. E., & Ou, Y. (2021). Voxelwise and Regional Brain Apparent 410. https:// doi. org/ 10. 1016/j. neuro image. 2018. 03. 007 Diffusion Coefficient Changes on MRI from Birth to 6 Years of Vieira, S., Pinaya, W. H. L., & Mechelli, A. (2017). Using deep learning Age. Radiology, 298, 415–424. https:// doi. org/ 10. 1148/ r adiol. to investigate the neuroimaging correlates of psychiatric and neu- 20202 02279 rological disorders: Methods and applications. Neuroscience and Stefano, A., Comelli, A., Bravatà, V., Barone, S., Daskalovski, I., Savoca, Biobehavioral Reviews, 74, 58–75. https:// doi. org/ 10. 1016/j. neubi G., Sabini, M. G., Ippolito, M., & Russo, G. (2020). A preliminary orev. 2017. 01. 002 PET radiomics study of brain metastases using a fully automatic seg- Vos, T., Lim, S. S., Abbafati, C., Abbas, K. M., Abbasi, M., Abbasi- mentation method. BMC Bioinformatics, 21, 325. https://d oi.o rg/1 0. fard, M., Abbasi-Kangevari, M., Abbastabar, H., Abd-Allah, F., 1186/ s12859- 020- 03647-7 Abdelalim, A., Abdollahi, M., Abdollahpour, I., Abolhassani, H., Steffener, J., Habeck, C., O’Shea, D., Razlighi, Q., Bherer, L., & Stern, Y. Aboyans, V., Abrams, E. M., Abreu, L. G., Abrigo, M. R. M., (2016). Differences between chronological and brain age are related Abu-Raddad, L. J., Abushouk, A. I., & Murray, C. J. L. (2020). to education and self-reported physical activity. Neurobiology of Global burden of 369 diseases and injuries in 204 countries and Aging, 40, 138–144. https://doi. or g/10. 1016/j. neur obiola ging. 2016. territories, 1990–2019: A systematic analysis for the Global Bur- 01. 014 den of Disease Study 2019. The Lancet, 396, 1204–1222. https:// Stoeckel, L. E., Garrison, K. A., Ghosh, S. S., Wighton, P., Hanlon, doi. org/ 10. 1016/ S0140- 6736(20) 30925-9 C. A., Gilman, J. M., et al. (2014). Optimizing real time fMRI Wang, G., Luo, T., Nielsen, J.-F., Noll, D. C., & Fessler, J. A. (2021). neurofeedback for therapeutic discovery and development. Neu- B-spline Parameterized Joint Optimization of Reconstruction roImage Clinical, 5, 245–255. https:// doi. or g/ 10. 1016/j. nicl. and K-space Trajectories (BJORK) for Accelerated 2D MRI. 2014. 07. 002. ArXiv210111369. 1 3 964 Neuroinformatics (2022) 20:943–964 Weiss, T., Senouf, O., Vedula, S., Michailovich, O., Zibulevsky, M., Network Open, 3, e2017357–e2017357. https://d oi.o rg/1 0.1 001/ & Bronstein, A. (2021). PILOT: Physics-Informed Learned jaman etwor kopen. 2020. 17357 Optimized Trajectories for Accelerated MRI. ArXiv190905773 Zbontar, J., Knoll, F., Sriram, A., Murrell, T., Huang, Z., Muckley, Physics. M. J., Defazio, A., Stern, R., Johnson, P., Bruno, M., Parente, Whitfield-Gabrieli, S., Ghosh, S. S., Nieto-Castanon, A., Saygin, Z., M., Geras, K. J., Katsnelson, J., Chandarana, H., Zhang, Z., Doehrmann, O., Chai, X. J., Reynolds, G. O., Hofmann, S. G., Drozdzal, M., Romero, A., Rabbat, M., Vincent, P., & Lui, Y. W. Pollack, M. H., & Gabrieli, J. D. E. (2016). Brain connectomics (2019). fastMRI: An Open Dataset and Benchmarks for Acceler- predict response to treatment in social anxiety disorder. Molecu- ated MRI. ArXiv181108839 Physics Statistics. lar Psychiatry, 21, 680–685. https:// doi. org/ 10. 1038/ mp. 2015. Zech, J. R., Badgeley, M. A., Liu, M., Costa, A. B., Titano, J. J., & 109 Oermann, E. K. (2018). Variable generalization performance of Whitfield-Gabr ieli, S., Wendelken, C., Nieto-Castañón, A., Bailey, S. a deep learning model to detect pneumonia in chest radiographs: K., Anteraper, S. A., Lee, Y. J., Chai, X.-Q., Hirshfeld-Becker, D. A cross-sectional study. PLoS Medicine, 15, e1002683. https:// R., Biederman, J., Cutting, L. E., & Bunge, S. A. (2020). Associa-doi. org/ 10. 1371/ journ al. pmed. 10026 83 tion of Intrinsic Brain Architecture With Changes in Attentional Zhang, X., Braun, U., Tost, H., & Bassett, D. S. (2020). Data-Driven and Mood Symptoms During Development. JAMA Psychiatry, 77, Approaches to Neuroimaging Analysis to Enhance Psychiatric 378–386. https:// doi. org/ 10. 1001/ jamap sychi atry. 2019. 4208 Diagnosis and Therapy. Biology Psychiatry Cognition Neurosci- Wijaya, S. K., Badri, C., Misbach, J., Soemardi, T. P., & Sutanno, V. ence Neuroimaging, 5, 780–790. https:// doi. org/ 10. 1016/j. bpsc. (2015). Electroencephalography (EEG) for detecting acute ischemic 2019. 12. 015. stroke, in: 2015 4th International Conference on Instrumentation, Zhang, Z., Romero, A., Muckley, M. J., Vincent, P., Yang, L., & Drozdzal, Communications, Information Technology, and Biomedical Engi- M. (2019) Reducing Uncertainty in Undersampled MRI Reconstruc- neering (ICICI-BME). Presented at the 2015 4th International tion with Active Acquisition. ArXiv190203051 Cs. Conference on Instrumentation, Communications, Information Zhou, H., Hu, R., Tang, O., Hu, C., Tang, L., Chang, K., Shen, Q., Technology, and Biomedical Engineering (ICICI-BME), pp. 42–48. Wu, J., Zou, B., Xiao, B., Boxerman, J., Chen, W., Huang, R. https:// doi. org/ 10. 1109/ ICICI- BME. 2015. 74013 12 Y., Yang, L., Bai, H. X., & Zhu, C. (2020). Automatic Machine Woon, W. L., Cichocki, A., Vialatte, F., & Musha, T. (2007). Tech- Learning to Differentiate Pediatric Posterior Fossa Tumors on niques for early detection of Alzheimer’s disease using spontane- Routine MR Imaging. American Journal of Neuroradiology, 41, ous EEG recordings. Physiological Measurement, 28, 335–347. 1279–1285. https:// doi. org/ 10. 3174/ ajnr. A6621 https:// doi. org/ 10. 1088/ 0967- 3334/ 28/4/ 001 Zhou, M., Scott, J., Chaudhury, B., Hall, L., Goldgof, D., Yeom, K. Xiao, T., Hua, W., Li, C., & Wang, S. (2019). Glioma Grading Pre- W., Iv, M., Ou, Y., Kalpathy-Cramer, J., Napel, S., Gillies, R., diction by Exploring Radiomics and Deep Learning Features, Gevaert, O., & Gatenby, R. (2018). Radiomics in Brain Tumor: in: Proceedings of the Third International Symposium on Image Image Assessment, Quantitative Feature Descriptors, and Computing and Digital Medicine, ISICDM 2019. Association Machine-Learning Approaches. American Journal of Neurora- for Computing Machinery, New York, NY, USA, pp. 208–213. diology, 39, 208–216. https:// doi. org/ 10. 3174/ ajnr. A5391 https:// doi. org/ 10. 1145/ 33648 36. 33648 77 Zwanenburg, A., Vallières, M., Abdalah, M. A., Aerts, H. J. W. L., Xu, J., Gong, E., Pauly, J., & Zaharchuk, G. (2017). 200x Low-dose Andrearczyk, V., Apte, A., Ashrafinia, S., Bakas, S., Beukinga, R. PET Reconstruction using Deep Learning. ArXiv171204119 Cs. J., Boellaard, R., Bogowicz, M., Boldrini, L., Buvat, I., Cook, G. Yang, G., Yu, S., Dong, H., Slabaugh, G., Dragotti, P. L., Ye, X., J. R., Davatzikos, C., Depeursinge, A., Desseroit, M.-C., Dinapoli, Liu, F., Arridge, S., Keegan, J., Guo, Y., & Firmin, D. (2018). N., Dinh, C. V., & Löck, S. (2020). The Image Biomarker Stand- DAGAN: Deep De-Aliasing Generative Adversarial Networks ardization Initiative: Standardized Quantitative Radiomics for High- for Fast Compressed Sensing MRI Reconstruction. IEEE Trans- Throughput Image-based Phenotyping. Radiology, 295, 328–338. actions on Medical Imaging, 37, 1310–1321. https:// doi. org/ 10. https:// doi. org/ 10. 1148/ radiol. 20201 91145 1109/ TMI. 2017. 27858 79 Ye, E., Sun, H., Leone, M. J., Paixao, L., Thomas, R. J., Lam, A. Publisher's Note Springer Nature remains neutral with regard to D., & Westover, M. B. (2020). Association of Sleep Electroen- jurisdictional claims in published maps and institutional affiliations. cephalography-Based Brain Age Index With Dementia. JAMA 1 3

Journal

Neuroinformatics – Springer Journals

Published: Oct 1, 2022

Keywords: Machine learning; Deep learning; Clinical translational neuroimaging; Brain health; MRI; PET; EEG; Transcranial magnetic stimulation

There are no references for this article.