Access the full text.
Sign up today, get DeepDyve free for 14 days.
Functional magnetic resonance imaging (fMRI) has been shown successfully to assess and stratify patients with painful diabetic peripheral neuropathy (pDPN). This supports the idea of using neuroimaging as a mechanism-based technique to individualise therapy for patients with painful DPN. The aim of this study was to use deep learning to predict treatment response in patients with pDPN using resting state functional imaging (rs-fMRI). We divided 43 painful pDPN patients into responders and non-responders to lidocaine treatment (responders n = 29 and non-responders n = 14). We used rs-fMRI to extract functional connectivity features, using group independent component analysis (gICA), and performed automated treatment response deep learning classification with three-dimensional convolutional neural networks (3D-CNN). Using gICA we achieved an area under the receiver operating characteristic curve (AUC) of 96.60% and F1-Score of 95% in a ten-fold cross validation (CV) experiment using our described 3D-CNN algorithm. To our knowledge, this is the first study utilising deep learning methods to classify treatment response in pDPN. Keywords Functional magnetic resonance imaging · Resting state · Convolutional neural network · Painful diabetic peripheral neuropathy · Treatment response Introduction different applications is the identification of prominent delineated features on imaging datasets and there have been numerous stud- The application of machine learning (ML) methods to the analy- ies investigating the utility of CNN in this context. However, the ses of neuroimaging datasets has led to significant advances in classification of clinical phenotypes using raw medical images in the field. It has enabled a shift from population/cohort-based the absence of well-defined delineated features poses substantial analyses into more individualised biomarkers of disease or challenges for CNN and has not been fully investigated. functional brain states. From a clinical perspective, this has fun- In this study, we will examine a well-characterised cohort damental importance in diagnosis, prognosis and patient strati- of patients with chronic neuropathic pain secondary to dia- fication. One ML approach that is increasingly being used on betes. Neuropathy is one of the commonest complications of imaging datasets to detect lesions (Zhang et al., 2019b), automate diabetes causing disabling pain in the lower and upper limbs tissue segmentation (Liu et al., 2018) and classify different brain in up to a quarter of patients. This often results in consider- disorders (Farooq et al., 2017) is deep learning using convolu- able disability and suffering. Pharmacotherapy is the mainstay tional neural network (CNN). One thing in common across these of treatment to alleviate symptoms but it is often ineffective. Even with optimal pharmacotherapy, only a third of patients experience any meaningful pain relief (i.e. a 50% reduction in * Kevin Teh email@example.com pain intensity scores) (Finnerup et al., 2015; Katz et al., 2008). We recently demonstrated that combined ML and magnetic Academic Unit of Radiology, Department of Infection, resonance (MR) imaging could accurately predict patients into Immunity and Cardiovascular Disease, University sensory phenotypes (Teh et al., 2021) with an accuracy of 0.92. of Sheffield, Sheffield, UK 2 This has the potential to serve as a biomarker for use in patient Diabetes Research Department, Sheffield Teaching Hospitals stratification, leading to a more efficient treatment approach. NHS Foundation Trust, Sheffield, UK 3 We used a linear, support vector machine (SVM) algorithm Department of Oncology and Metabolism, University using features extracted from rs-fMRI and MRI volumetry. of Sheffield, Sheffield, UK Vol.:(0123456789) 1 3 Neuroinformatics Independent component analysis (ICA) is an effective Details of the clinical and neurophysiological assess- method to derive brain functional networks with both func- ment performed at Visit 1 are shown in Table 1. There were tional connectivity and spatial information (Nickerson et al., no statistically significant differences between responders 2017; Rajapakse et al., 2006). Unlike commonly used seed and non-responders in the assessments performed. Written based analysis approaches (seed to voxel and ROI to ROI), informed consent for the study was obtained before sub- ICA is a pure data-driven method, which generates highly jects participated in the study, which has prior approval by reproducible large-scale brain networks (Damoiseaux et al., the Sheffield Local Research Ethics Committee (Sheffield, 2006). A previous study using gICA analysis has shown that U.K.). resting state functional connectivity can successfully predict subsequent duloxetine treatment response of major depres- MRI Acquisition sive disorder (Fu et al., 2015). To capture these elusive fea- ture delineations, 3D group ICA features have been shown On the second visit prior to IV lidocaine treatment sub- to achieve superior classification accuracy (Qureshi et al., jects underwent MRI using a Phillips Achieva 3 Tesla 2019) when used in conjunction with CNN techniques. system (Phillips Medical Systems, Holland) with a Hence, the primary aim of this ML study is to train a 32-channel head coil. T1-weighted magnetisation pre- CNN-based classification model to differentiate respond - pared rapid acquisition gradient echo sequence was used ers and non-responders to neuropathic pain treatment using to acquire anatomical data with the following parameters: resting state (RS) maps as inputs to a 3D-CNN architecture. repetition time (TR) 7.2 ms, echo time (TE) 3.2 ms, f lip The secondary aim was to compare the classification per - angle 8°, and voxel size 0.9 mm3, yielding isotropic spa- formance of two different inputs (all group-level ICA com- tial resolution. Resting-state fMRI data was also acquired ponents vs statistically significant group-level components) while subjects fixated on a cross using a T2*-weighted for the 3D-CNN architecture. Table 1 Demographic, metabolic and neurophysiological assessments of study participants Method Responder Non Responder p Study Design and Participants n 29 14 Age (years) 57.48(8.82) 57.07(11.12) 0.896 Forty-three consecutive patients (responders n = 29 and non- Male sex (n, %) 19, 66 5,36 responders n = 14) with painful diabetic neuropathy were Type of diabetes (Type 9,20 3,11 recruited for this study. We used intravenous (IV) lidocaine 1/2) as a experimental model to examine treatment response in Duration of diabetes 20.41(14.7) 20.86(14.6) 0.926 this neuroimaging study. It is a recognised treatment option (years) in specialist centres for the management of intractable neu- Duration of pain (years) 10.37(7.51) 7.89(6.34) 0.292 ropathic pain i.e. when conventional treatment is ineffective Hba1c (mmol/mol) 70(18.42) 73(16.67) 0.714 (Kastrup et al., 1986, 1987; Viola et al., 2006) and included NTSS-6 score 16.01(3.29) 14.18(4.29) 0.176 in clinical guidelines (Tesfaye et al., 2011). It is also a good TCNS 21.19(4.49) 16(9.26) 0.07 experimental model for assessing neuroimaging markers of BMI 33.23(11.46) 30(5.6) 0.328 treatment response because 1) it has a central mechanisms of Becks 23.84(13.06) 11.71(14.15) 0.051 action (Abdi et al., 1998; Devor et al., 1992; Omana-Zapata DN4 7.86(1.49) 6.67(1.58) 0.055 et al., 1997); 2) a short half-life (40 min) – nullifying a direct Sural treatment effect on MRI measures; and 3) up to one half of Conduction velocity (m/s) 19.12(17.03) 14.98(18.73) 0.562 patients do not respond to treatment—enabling assessment of Amplitude (mAmp) 3.98(6.18) 1.9(5.7) 0.392 matched groups of responders and non-responders. Patients Common Peroneal Nerve completed a standardised questionnaire to determine treat- Conduction velocity (m/s) 26.35(17.97) 29.75(20.22) 0.646 ment response to IV lidocaine treatment. Responders were Amplitude (mAmp) 2.04(2.2) 1.39(2.29) 0.467 defined as patients who report at least a 30% reduction in Distal latency (ms) 3.73(2.83) 4.2(3.27) 0.693 pain intensity score (0 to 10 numeric rating scale, where Tibial Latency 3.7(2.94) 4.07(3.46) 0.829 0 = no pain and 10 = worse pain imaginable) post lidocaine Data presented as mean (standard deviation) unless otherwise stated. treatment. Study visits were conducted prior to IV lidocaine Groups were compared using Student’s t-tests treatment. Altogether, there were two study visits: Visit 1 was NTSS-6 Neuropathy Total Symptom Score-6 pain questionnaire, for detailed clinical and neurophysiological assessments and TCNS Toronto Clinical Neuropathy Score, DN4 Douleur Neuropathique Visit 2 for MR neuroimaging. en 4 Questions 1 3 Neuroinformatics pulse sequence, with TE = 35 ms; TR = 2600 ms, in- Deep Learning Features plane pixel dimensions = 1.8 mm × 1.8 mm, contiguous trans-axial slices thickness of 4 mm, a slice acquisition Multiple 3D spatial component maps were acquired by per- order = ascending (bottom-up), F OV = 128 × 128 mm, forming dual regression (Beckmann et al., 2009) on the pixel bandwidth of 2129, number of phase encoding steps group ICA. The 3D spatial maps of these selected func- of 94 and an echo train length of 53. tional networks were then inputted into our chosen 3D-CNN for classification of responders and non-responders to IV lidocaine treatment groups. This work examine differences Pre‑processing and Data Analysis in classification performance between an approach using all ICA spatial components (ASC) and another using a semi- Group ICA based analysis was performed using the CONN automatically selected meaningful independent component (version 18.b) (Whitfield-Gabrieli & Nieto-Castanon, 2012): time-series dataset (SSC). There were 8 spatial maps (SSC) functional connectivity toolbox software which was also identified when matched to a spatial template in CONN used to perform all pre-processing steps (using the default toolbox chosen with p < 0.1. These 8 chosen maps will be pre-processing pipeline), as well as subsequent statistical shown in more detail in the results section. We also com- analyses, on all subject scans. Using the CONN toolbox pre- pared the classification results of the 8 selected spatial processing pipeline, raw functional images were slice-time component maps with all 30 spatial component maps. Non corrected, realigned (motion corrected), unwarped, and co- post-processed resting state pre-processed image inputs registered to each subject's T1-weighted dataset in accord- were also used as a base comparison. ance with standard algorithms. Images were then normalised to Montreal Neurological Institute (MNI) coordinate space, spatially smoothed (5 mm full-width at half maximum), and Deep Learning and 3D‑CNN Framework resliced to yield 2 × 2 × 2 mm voxels and the resting state connectivity analysis was performed using CONN toolbox. We used a 3D-CNN based deep learning classification Further resting state analysis using gICA in CONN toolbox framework in this study based on a lightweight Voxnet was performed with variance normalisation pre-condition- (Maturana & Scherer, 2015) architecture. The framework ing, subject or condition concatenation of BOLD signal data was implemented on the TensorFlow library version 2.5 along temporal dimension, group level dimensionality reduc- with two bridged Nvidia Ge-Force GTX 2080Ti graphical tion, fast ICA for estimation of independent spatial compo- processing units (GPU). For the training model, we used the nents, and GICA1 back-projection for individual subject- Adam optimizer with a learning rate of 0.001 and setting level spatial map estimation (Beckmann et al., 2009). The epsilon at 0.1. Since the size of the dataset was relatively number of independent components to be estimated was set small for deep learning, to avoid model overfitting, we used to 30 and dimensionality reduction was set to 64. ten-fold cross-validation in this study to report the mean performance of the overall model. A modified version of the Voxnet classification frame- Proposed Framework work was used in this study. Specifically, we added batch normalization layers in the convolution layer. A dropout rate Figure 1 illustrates the procedures used for joint predic- of 0.3 was used in the fully connected layers with a batch tion of treatment response status in this study. Figure 1a size set at 4 and using 100 epochs. The parameters includ- shows the first step of the procedure in which the group- ing learning rate, epsilon value, dropout rate, batch size, and level 3D ICA spatial maps of all studied subjects were epoch size were optimised using the following ranges. For computed. Next, we employed an automatic clustering epsilon, we tuned it in the range of [0.1: 0.05: 1], for learn- tool (using CONN toolbox spatial match to template) to ing rate, we tuned it in the logarithmic range of [1, 0.1, 0.01, identify the useful group-level independent components 0.001, 0.0001, and 0.00001], for the dropout rate, we tuned (X ICs) for further analysis; we also compared all analysed it in the range [0.1: 0.05: 1], for the batch size, we optimized components (30 ICs). We then performed the two-stage it by the maximum available GPU memory, and the number dual regression (Fig. 1c) to extract: 1) subject-specific IC of epochs were tuned in the range of [10: 1: 200]. We used a time courses (stage 1) and; 2) subject-specific ICA spatial binary entropy loss function when training the deep learning maps (stage 2). Figure 1d presents the nested tenfold cross network, and an early stopping criterion was used to stop validation (CV) strategy used to evaluate the performance the network training with respect to the leave-one-out cross- of the proposed framework. The subject-specific ICA maps validation loss. The optimal parameter tuning was chosen to were applied to a 3D Convolutional Neural Network (3D- reflect the best performance metrics for the ASC and SSC CNN) for the classification task (Fig. 1e). 1 3 Neuroinformatics Fig. 1 Illustration of our proposed treatment response classification nents (8 and 30 IC components) were used to extract subject-specific framework using group ICA features extracted from rs-fMRI scans. time courses (stage 1) and subject-specific spatial ICA maps (stage Firstly, (A) shows the 30 group-level ICA networks from all studied 2) using dual regression method in (C). Step (D) presents the nested subjects. Next, we used CONN toolbox spatial to match template- tenfold cross validation strategy to train the classifier and regressors, clustering tool to identify the useful functional group-level ICA tune the hyperparameters and test their performances. Next the 4D maps. As a result, 8 IC selected spatial components (SSC) (p < 0.1) ICA spatial component maps (concatenated multiple 3D spatial com- were identified as statistically useful networks whereas 22 ICs were ponent maps) were fed into the 3D-CNN (E) for treatment response not used in the selected IC subset (B). Then these group-level compo- classification dataset. The chosen parameters were similarly used on the metrics we regard the AUC and F1 score to be the most pre-processed resting state image data (RSP). important metric in this work. Previous work has concluded that these measures fit best for imbalanced data problems Performance Measures (Jeni et al., 2013; Raeder et al., 2012). These, two perfor- mance measures are described further. Firstly, introducing One of the most crucial processes of defining a machine some acronyms TP, TN, FP and FN are the number of true learning model is model evaluation. For this work, we will positive (responders), true negative (non-responders), false compare standard metrics such as accuracy (ACC), precision positive and false negative samples, respectively. The met- (PR) and recall (RE) with average cross-validation AUC rics used are shown below: and F1-Score. We choose to include the AUC and F1-score TP + TN due to the slight imbalance in our dataset leaning towards ACC = (1) TP + TN + FP + FN treatment responders to lidocaine. Of all these performance 1 3 Neuroinformatics TP Also, when running our networks, we also calculated the PR = (2) TP + FP timing cost comparison. Using the deep learning configura- tion described above with our lightweight modified Voxnet CNN pipeline took 3.4 min to train 100 epoch per fold for TP RE = (3) the SSC, 5.1 min for ASC and 29.3 min for RSP. TP + FN where precision measures samples correctly classified as Network Component Analysis positive, and recall describes the proportion of all positive samples classified as positive. F1 score is interpreted as a Table 3 shows the ranking of each functional network as weighted average of the precision (PR) and recall (RE): the features of a deep learning framework. These compo- nents were chosen based on p < 0.1. The uncorrected p-value PR.RE F1 = 2 ⋅ (4) revealed the component’s significance. Using this signifi- PR + RE cance level we identified eight ICA spatial components AUC score (Area Under the receiver operating character- from a subset of 30. The 5 resting state functional networks istic Curve) characterises the area under the curve of sen- identified by the CONN network cortical ROIs (HCP-ICA) sitivities plotted against corresponding false positive rates through the ICA analysis were the cerebellar network (CEB), (FPR): default mode network (DMN), frontoparietal (FPN), sensori- motor network (SMN) and visual network (VN). They were FP FPR = (5) used to perform dual regression to generate subject-specific FP + TN time courses for connectivity analysis and spatial maps for classification. Results Discussion Classification Results To the best of our knowledge, this is the first study classify - ing treatment response using rs-fMRI and 3D-CNN deep Our results suggest that using ICA analysis specifically spa- learning architecture. The key findings of this work are 1) tial component maps can be used as a good discriminatory using ICA spatial component maps (ASC and SSC) per- predictor of lidocaine treatment response in painful DPN forms better than only using RSP as the input to our CNN patients. In our study, we compared using ASC, SSC and networks; 2) using all the group ICA spatial components RSP as the input to our CNN classification workflow. Com- (ASC) information performs better compared to the semi- paring the classification performances, the first thing we automatic selection of the highly relevant networks (SSC) noticed from Table 2 is the sub optimal results obtained from and 3) a lightweight 3D-CNN deep learning architecture for the pre-processed resting state data. We achieved a F1-Score classification uses imaging data more efficiently. With these of 69% and an AUC score of 44%. On the other hand, ASC approaches, we achieved better treatment response classifi- gave the best performing results achieving a F1-Score of cation results. 95% and a mean balanced AUC of 96.60% in a ten-fold CV experiment using the described 3D-CNN algorithm. As compared to SSC it also achieved good performance with Table 3 SSC networks selected based on comparison of network F1-Score of 90% and AUC of 85%. The rest of the results activities between responders to lidocaine and non-responders at for the three groups are shown in Table 2. p < 0.1 Component Name Component p-value Pearson(r) Number Table 2 Described performance metrics comparing ASC and SSC Cerebellar ICA1 0.014 0.373 results Default Mode ICA10 0.003 0.446 ASC SSC RSP FrontoParietal ICA5 0.058 0.292 Accuracy 0.9307 0.8529 0.6743 Sensorimotor ICA9 0.099 0.254 Precision 0.9237 0.8460 0.6744 Sensorimotor ICA26 0.058 0.291 Recall 0.9787 0.9623 0.7136 Sensorimotor ICA29 0.026 0.339 F1-Score 0.9504 0.9004 0.6934 Visual ICA6 0.009 0.394 AUC 0.9661 0.8499 0.4375 Visual ICA11 0.001 0.579 1 3 Neuroinformatics Our analyses show a clear disparity between RSP and In our literature search, we did not find any treatment both the ICA analysed spatial component maps (ASC and response classification of pDPN using machine learning. SSC) giving an enhanced performance score. When com- We, however, found two closest studies focusing on pain paring ASC and SSC there is an improvement for the ASC and treatment response classification using ML. The first score with a 5% improvement in the F1-Score and a 11.6% study used a SVM classifier to differentiate lower back pain improvement for the AUC. The extra 22 spatial components and healthy volunteers (Lamichhane et al., 2021). They in the ASC which are noisier (not satisfying the p < 0.1 achieved an average classification accuracy of 74.51% and condition of the 8 networks shown in Fig. 2) is shown to an AUC = 0.787. The second study used deep learning (DL) have contributed to the performance metric boost of the neural networks in classifying chronic pain patients and ASC method. Using deep learning we were able to auto- pain-free controls (Santana et al., 2019). Their best model matically discover the intricate structure differences of the using the Ann4brain architecture and MSDL parcellation responders and non responders imaging features especially had a balanced accuracy of 86.8% and AUC of 0.918. Due true when larger imaging information was used during the to the dier ff ences in patient cohorts, we cannot directly com - training step as shown by (LeCun et al., 2015). Hence from pare these results with ours. However, with our high testing our results we surmise that when a dataset is closely matched accuracy of 93% and AUC of 97%, we believe we can trans- (in our case painful diabetic neuropathy patients) using more late our painful DPN CNN model to other disease cohorts. ICA components provides better performance. In future we For example, with other neuropathy patients (i.e. idiopathic will explore SSC classification performance more using dif- and chemotherapy painful neuropathy). ferent p value network selection (i.e. p < 0.2, p < 0.3 etc.) to Another important factor in our study is that we set the further our assertion. number of ICA components to 30, as is conventionally used We did not explicitly compare our Voxnet based light- in a recent study (Qureshi et al., 2019). Recently, it has been weight model with other popular deeper learning networks shown that the detected disease-related differences in func- such as Resnet or VGG-16 (Simonyan & Zisserman, 2014). tional connectivity may alter as a function of ICA model We calculated the trainable parameters of VGG-16 to be order, specifically how many ICA components to use (Abou around 2.5 billion compared to around 1.4 million using Elseoud et al., 2011). Their multi-level ICA exploration of Voxnet with identical 4D ICA spatial component input data- unmedicated seasonal affective disorders functional con- set. Hence due to GPU memory limitations, we did not com- nectivity enables optimisation of sensitivity to brain disor- pare our proposed CNN model with these deeper networks. ders and found 45 RS networks as the optimum component When the computation time was compared between SSC number to use. To further explore the effect of chosen RS and ASC there was an extra 1.7 min computation time per ICA components, a recent study compared 2, 10 and 45 ICA fold when using the ASC as the input data. We consider this chosen components. They reached an accuracy of 95% for a good use of extra computational time given the boost of the 2 components, 93% for the 10 components and 92% for performance when using the ASC input dataset. the 45 components. In our study, the performance decreases Fig. 2 Spatial to match template (CONN toolbox gICA) to select the best SSC based on p < 0.1 1 3 Neuroinformatics with smaller component number, which is the opposite of In summary, in the last few years CNNs have been used the described study. We will investigate a wider number of very successfully with rs-fMRI to classify different disease set ICA component analysis to further affirm our findings. phenotypes (Qureshi et al., 2019). We have demonstrated the Our study also used gICA, a data driven method, to inves- effectiveness of using a 3D lightweight CNN method, which tigate the association between brain networks. Here, we not only saves on computational time but also gives high decomposed the signal from whole brain voxels to spatially prediction scores. We believe that the presented results show and temporally independent components. Future work will that rs-fMRI has great application potential as a biomarker in involve using different indices of functional brain alterations, neuropathic pain (Cauda et al., 2009, 2010; Sağ et al., 2019; including amplitude of low-frequency fluctuations (ALFF), Zhang et al., 2019a). Finally, we have successfully shown that regional homogeneity (ReHo), and regional functional cor- we are able to differentiate responders and non-responders to relation strength (RFCS) which has been successfully dem- neuropathic pain treatment using 3D gICA resting state (RS) onstrated in a CNN migraine study (Yang et al., 2018). In maps as inputs to a 3D-CNN architecture. our gICA analysis method, we only compared the analysis of SSC and ASC. Here, we concatenated the 3D spatial compo- nents together (8 and 30), which consists of different resting Information Sharing Statement state networks (RSN). However, in future to fully understand and interpret our results further we will also explore rest- The analysed RS data that supports the findings of this study ing state networks individually between our two groups. Of can be downloaded at https://doi. or g/10. 5281/ zenodo. 63255 87 . particular interest to our pDPN dataset are the default mode, Pre-processed or raw resting state data associated with this work sensorimotor and fronto parietal networks. This will enable can be provided upon request. us to identify and rank individual networks based on the Acknowledgements The authors thank the radiographers at the Univer- binary classification performance as shown in a recent work sity of Sheffield Academic Department of Magnetic Resonance Imag- on Autism spectrum disorder (Yang et al., 2021). To further ing for their hard work, skills, and contributions. The authors also this study we will analyse the resting state data with higher thank the study volunteers who spent considerable time participating regional sensitivity using a region of interest (ROI) to ROI in this study. analysis approach. The outcome of the ROI-ROI analysis Author Contributions K.T ran all the MR resting state and machine will also be expanded further with the use of linear classifi- learning analysis and is the guarantor of this work and, as such, had ers for example linear SVM instead of black box models like full access to all the data in the study and takes responsibility for the deep learning for better interpretability as shown recently integrity of the data and the accuracy of the data analysis and wrote the (Teh et al., 2021). manuscript. P.A and S.T. reviewed and revised the manuscript. D.S. is the principal investigator of the used dataset, recruited participants, This study has some limitations. Our dataset of 43 sam- undertook clinical and neurophysiological assessments, researched and ples is relatively small. However, in our previous study, we analyzed clinical and MR data. have shown with similar sample sizes using resting state ROI to ROI image analysis to have enough sample size power to Funding This study was supported by the European Foundation for the clearly differentiate between lidocaine treatment responses Study of Diabetes (Microvascular Complications Project Grant), the National Institute of Health Research Efficacy and Mechanism Evalu- (Wilkinson et al., 2020). Another limitation is the lack of ation Programme (NIHR 129921) and University of Sheffield, Health a separate testing dataset. To address this, the results pre- Education England, Knowledge Exchange fund. sented are the ten-fold cross-validated performance metrics of our CNN method. In future work, we intend to apply Declarations our algorithm to other treatment response datasets to exter- nally validate our algorithm. This will include out of sam- Conflict of Interest No potential conflicts of interest relevant to this ple testing with within disease datasets (painful DPN) and article were reported. also between disease datasets (idiopathic and chemotherapy Open Access This article is licensed under a Creative Commons Attri- induced neuropathy). In addition, although we have shown bution 4.0 International License, which permits use, sharing, adapta- that a 3D-CNN model trained on rs-fMRI datasets accurately tion, distribution and reproduction in any medium or format, as long classifies patients according to their treatment response (i.e. as you give appropriate credit to the original author(s) and the source, based on analgesic response), we are unable to determine provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are or explore the impact of neuropathy or diabetes itself on included in the article's Creative Commons licence, unless indicated cortical changes described. To address this, future studies otherwise in a credit line to the material. If material is not included in should consider including appropriate control subjects (e.g. the article's Creative Commons licence and your intended use is not participants with diabetes and no neuropathy or subjects permitted by statutory regulation or exceeds the permitted use, you will with neuropathic pain from different aetiologies). 1 3 Neuroinformatics need to obtain permission directly from the copyright holder. To view a Liu, F., Zhou, Z., Jang, H., Samsonov, A., Zhao, G., & Kijowski, R. copy of this licence, visit http://cr eativ ecommons. or g/licen ses/ b y/4.0/ . (2018). Deep convolutional neural network and 3D deformable approach for tissue segmentation in musculoskeletal magnetic reso- nance imaging. Magnetic Resonance in Medicine, 79(4), 2379–2391. Maturana, D., & Scherer, S. (2015). Voxnet: A 3d convolutional neural References network for real-time object recognition. 2015 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS). Nickerson, L. D., Smith, S. M., Öngür, D., & Beckmann, C. F. (2017). Abdi, S., Lee, D. H., & Chung, J. M. (1998). The anti-allodynic effects Using dual regression to investigate network shape and amplitude of amitriptyline, gabapentin, and lidocaine in a rat model of neu- in functional connectivity analyses. Frontiers in Neuroscience, ropathic pain. Anesthesia & Analgesia, 87(6), 1360–1366. 11, 115. Abou Elseoud, A., Littow, H., Remes, J., Starck, T., Nikkinen, J., Nissilä, Omana-Zapata, I., Khabbaz, M. A., Hunter, J. C., & Bley, K. R. (1997). J., Tervonen, O., Timonen, M., & Kiviniemi, V. J. (2011). Group- QX-314 inhibits ectopic nerve activity associated with neuro- ICA model order highlights patterns of functional brain connectivity. pathic pain. Brain Research, 771(2), 228–237. Frontiers in Systems Neuroscience, 5, 37. Qureshi, M. N. I., Oh, J., & Lee, B. (2019). 3D-CNN based discrimi- Beckmann, C. F., Mackay, C. E., Filippini, N., & Smith, S. M. (2009). nation of schizophrenia using resting-state fMRI. Artificial Intel- Group comparison of resting-state FMRI data using multi-subject ligence in Medicine, 98, 10–17. ICA and dual regression. NeuroImage, 47(Suppl 1), S148. Raeder, T., Forman, G., & Chawla, N. V. (2012). Learning from imbal- Cauda, F., D’Agata, F., Sacco, K., Duca, S., Cocito, D., Paolasso, I., anced data: Evaluation matters. In Data mining: Foundations and Isoardo, G., & Geminiani, G. (2010). Altered resting state atten- intelligent paradigms (pp. 315–331). Springer. tional networks in diabetic neuropathic pain. Journal of Neurol- Rajapakse, J. C., Tan, C. L., Zheng, X., Mukhopadhyay, S., & Yang, K. ogy, Neurosurgery & Psychiatry, 81(7), 806–811. (2006). Exploratory analysis of brain connectivity with ICA. IEEE Cauda, F., Sacco, K., D’Agata, F., Duca, S., Cocito, D., Geminiani, Engineering in Medicine and Biology Magazine, 25(2), 102–111. G., Migliorati, F., & Isoardo, G. (2009). Low-frequency BOLD Sağ, A. T., Has, A. C., Öztekin, N., Temuçin, Ç. M., & Oğuz, K. K. fluctuations demonstrate altered thalamocortical connectivity in (2019). Tracking pain in resting state networks in patients with diabetic neuropathic pain. BMC Neuroscience, 10(1), 1–14. hereditary and diabetic neuropathy. Archives of Neuropsychiatry, Damoiseaux, J. S., Rombouts, S., Barkhof, F., Scheltens, P., Stam, C. J., 56(2), 92. Smith, S. M., & Beckmann, C. F. (2006). Consistent resting-state Santana, A. N., Cifre, I., De Santana, C. N., & Montoya, P. (2019). networks across healthy subjects. Proceedings of the National Using deep learning and resting-state fMRI to classify chronic Academy of Sciences, 103(37), 13848–13853. pain conditions. Frontiers in Neuroscience, 13, 1313. Devor, M., Wall, P. D., & Catalan, N. (1992). Systemic lidocaine Simonyan, K., & Zisserman, A. (2014). Very deep convolutional net- silences ectopic neuroma and DRG discharge without blocking works for large-scale image recognition. arXiv preprint arXiv: nerve conduction. Pain, 48(2), 261–268. 1409. 1556 Farooq, A., Anwar, S., Awais, M., & Rehman, S. (2017). A deep CNN Teh, K., Wilkinson, I. D., Heiberg-Gibbons, F., Awadh, M., Kelsall, based multi-class classification of Alzheimer's disease using MRI. A., Pallai, S., Sloan, G., Tesfaye, S., & Selvarajah, D. (2021). 2017 IEEE International Conference on Imaging systems and Somatosensory network functional connectivity differentiates techniques (IST). clinical pain phenotypes in diabetic neuropathy. Diabetologia, Finnerup, N. B., Attal, N., Haroutounian, S., McNicol, E., Baron, R., 64(6), 1412–1421. Dworkin, R. H., Gilron, I., Haanpää, M., Hansson, P., & Jensen, Tesfaye, S., Vileikyte, L., Rayman, G., Sindrup, S. H., Perkins, B., T. S. (2015). Pharmacotherapy for neuropathic pain in adults: Baconja, M., Vinik, A., Boulton, A., & Toronto Expert Panel on A systematic review and meta-analysis. The Lancet Neurology, Diabetic Neuropathy. (2011). Painful diabetic peripheral neuropa- 14(2), 162–173. thy: Consensus recommendations on diagnosis, assessment and Fu, C. H., Costafreda, S. G., Sankar, A., Adams, T. M., Rasenick, M. management. Diabetes/metabolism Research and Reviews, 27(7), M., Liu, P., Donati, R., Maglanoc, L. A., Horton, P., & Marangell, 629–638. L. B. (2015). Multimodal functional and structural neuroimaging Viola, V., Newnham, H. H., & Simpson, R. W. (2006). Treatment of investigation of major depressive disorder following treatment intractable painful diabetic neuropathy with intravenous ligno- with duloxetine. BMC Psychiatry, 15(1), 1–11. caine. Journal of Diabetes and Its Complications, 20(1), 34–39. Jeni, L. A., Cohn, J. F., & De La Torre, F. (2013). Facing imbalanced Whitfield-Gabrieli, S., & Nieto-Castanon, A. (2012). Conn: a func- data - recommendations for the use of performance metrics. 2013 tional connectivity toolbox for correlated and anticorrelated brain Humaine Association Conference on Affective Computing and networks. Brain Connectivity, 2(3), 125–141. Intelligent Interaction. Wilkinson, I. D., Teh, K., Heiberg-Gibbons, F., Awadh, M., Kelsall, Kastrup, J., Angelo, H., Petersen, P., Dejgård, A., & Hilsted, J. (1986). A., Shillo, P., Sloan, G., Tesfaye, S., & Selvarajah, D. (2020). Treatment of chronic painful diabetic neuropathy with intravenous Determinants of treatment response in painful diabetic peripheral lidocaine infusion. British Medical Journal (Clinical Research neuropathy: A combined deep sensory phenotyping and multi- Ed.), 292(6514), 173. modal brain MRI study. Diabetes, 69(8), 1804–1814. Kastrup, J., Petersen, P., Dejgård, A., Angelo, H. R., & Hilsted, J. Yang, H., Zhang, J., Liu, Q., & Wang, Y. (2018). Multimodal MRI- (1987). Intravenous lidocaine infusion—a new treatment of based classification of migraine: Using deep learning convolu- chronic painful diabetic neuropathy? Pain, 28(1), 69–75. tional neural network. Biomedical Engineering Online, 17(1), Katz, J., Finnerup, N. B., & Dworkin, R. H. (2008). Clinical trial out- 1–14. come in neuropathic pain: Relationship to study characteristics. Yang, M., Cao, M., Chen, Y., Chen, Y., Fan, G., Li, C., Wang, J., & Neurology, 70(4), 263–272. Liu, T. (2021). Large-scale brain functional network integration Lamichhane, B., Jayasekera, D., Jakes, R., Glasser, M. F., Zhang, J., for discrimination of autism using a 3-D deep learning model. Yang, C., Grimes, D., Frank, T. L., Ray, W. Z., & Leuthardt, E. Frontiers in Human Neuroscience, 15, 277. C. (2021). Multi-modal biomarkers of low back pain: A machine Zhang, Q., Zhang, P., Yan, R., Xu, X., Mao, C., Liu, X., Li, F., Ma, J., learning approach. NeuroImage: Clinical, 29, 102530. Ye, L., & Yao, Z. (2019a). A single-blinded trial using resting- LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, state functional magnetic resonance imaging of brain activity in 521(7553), 436–444. 1 3 Neuroinformatics patients with type 2 diabetes and painful neuropathy. Diabetes pathologically proven datasets. Computerized Medical Imaging Therapy, 10(1), 135–147. and Graphics, 77, 101645. Zhang, S., Han, F., Liang, Z., Tan, J., Cao, W., Gao, Y., Pomeroy, M., Ng, K., & Hou, W. (2019b). An investigation of CNN mod- Publisher's Note Springer Nature remains neutral with regard to els for differentiating malignant from benign lesions using small jurisdictional claims in published maps and institutional affiliations. 1 3
Neuroinformatics – Springer Journals
Published: Aug 26, 2022
Keywords: Functional magnetic resonance imaging; Resting state; Convolutional neural network; Painful diabetic peripheral neuropathy; Treatment response
Access the full text.
Sign up today, get DeepDyve free for 14 days.