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

Learn More →

Automated prostate cancer detection via comprehensive multi-parametric magnetic resonance imaging texture feature models

Automated prostate cancer detection via comprehensive multi-parametric magnetic resonance imaging... Background: Prostate cancer is the most common form of cancer and the second leading cause of cancer death in North America. Auto-detection of prostate cancer can play a major role in early detection of prostate cancer, which has a significant impact on patient survival rates. While multi-parametric magnetic resonance imaging (MP-MRI) has shown promise in diagnosis of prostate cancer, the existing auto-detection algorithms do not take advantage of abundance of data available in MP-MRI to improve detection accuracy. The goal of this research was to design a radiomics-based auto-detection method for prostate cancer via utilizing MP-MRI data. Methods: In this work, we present new MP-MRI texture feature models for radiomics-driven detection of prostate cancer. In addition to commonly used non-invasive imaging sequences in conventional MP-MRI, namely T2-weighted MRI (T2w) and diffusion-weighted imaging (DWI), our proposed MP-MRI texture feature models incorporate computed high-b DWI (CHB-DWI) and a new diffusion imaging modality called correlated diffusion imaging (CDI). Moreover, the proposed texture feature models incorporate features from individual b-value images. A comprehensive set of texture features was calculated for both the conventional MP-MRI and new MP-MRI texture feature models. We performed feature selection analysis for each individual modality and then combined best features from each modality to construct the optimized texture feature models. Results: The performance of the proposed MP-MRI texture feature models was evaluated via leave-one-patient-out cross-validation using a support vector machine (SVM) classifier trained on 40,975 cancerous and healthy tissue samples obtained from real clinical MP-MRI datasets. The proposed MP-MRI texture feature models outperformed the conventional model (i.e., T2w+DWI) with regard to cancer detection accuracy. Conclusions: Comprehensive texture feature models were developed for improved radiomics-driven detection of prostate cancer using MP-MRI. Using a comprehensive set of texture features and a feature selection method, optimal texture feature models were constructed that improved the prostate cancer auto-detection significantly compared to conventional MP-MRI texture feature models. Background cause of cancer death in men in the United States with Prostate cancer is the most common form of cancer diag- an estimated 29,480 deaths in 2014 [2]. Given that the nosed in North American men, with roughly 23,500 new median patient survival time for metastatic prostate can- cases in 2014 in Canada [1] and 233,000 new cases in 2014 cer ranges from 12.2 to 21.7 months [3], early diagnosis in the United States [2]. Furthermore, prostate cancer is of clinically significant prostate cancer would have signifi- the third leading cause of cancer death in Canadian men cant benefits to patient care. This is particularly true given with an estimated 4,000 deaths [1], and second leading that the five-year survival rate after diagnosis for patients with prostate cancer at the non-metastatic stage is 96 % in Canada [4]. *Correspondence: farzad.khalvati@sri.utoronto.ca Department of Medical Imaging, University of Toronto, Toronto, ON, Canada In the current clinical model, men with positive digital Physical Sciences, Sunnybrook Research Institute, Toronto, ON, Canada rectal exam (DRE) and elevated prostate-specific anti- Full list of author information is available at the end of the article gen (PSA) require multicore random biopsies for risk © 2015 Khalvati et al. Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 2 of 14 stratification. However, there is an ongoing controversy Radiologists’ interpretations of MP-MRI have shown to about the role of prostate PSA as a screening test in achieve good prostate cancer detection rates, reaching prostate cancer. Two recent major randomized clinical tri- accuracies of 80 % in the peripheral zone of the prostate als [5, 6] have demonstrated that PSA screening contains a gland [28]. Similarly, several algorithms have been pro- significant risk of overdiagnosis for prostate cancer where posed for auto-detection of prostate cancer using MP- it is estimated that 50 % of screened men are diagnosed MRI setting [13–15, 29–31]. These algorithms usually with prostate cancer. This leads to painful needle biopsies compute a set of low-level features from the MP-MRI data and subsequent potential overtreatment [5–8]. Moreover, to construct feature vectors. Next, a supervised classifier it has become increasingly clear that carrying out prostate is trained using the computed feature vectors from the biopsy procedures escalates hospital admission rates due training cases and their associated ‘ground-truth’ labels to infectious complications, regularly resulting in discom- (e.g., labeled healthy or cancerous). Finally, the trained fort and possible sexual dysfunction while with the chance classifier is used to classify new cases. The reported values of the needle missing cancerous tissue [9–11]. Neverthe- for accuracy of cancerous versus healthy tissue classifica- less, PSA testing has proven to reduce prostate cancer tion ranges from 64 % to 89 %, depending on the feature mortality by 20–30 % at long-term follow-ups [10]. There- sets and training/test data. fore, the PSA testing remains an important biomarker in Different texture features have been used in the litera- diagnosing prostate cancers that are clinically significant. ture for automatic detection and classification of prostate The remaining challenge is how to improve the prostate cancer. Most of the reviewed methods utilized texture cancer diagnosis to reduce the overdiagnosis of clinically features that are based on one or more of the follow- insignificant cancers. ing methods: First-Order statistical method, second-order Automatic detection of prostate cancer as part of a clin- statistical methods or Co-Occurrence Matrices [32, 33], ical decision support system can potentially help radiolo- steerable Gabor filter [34], Gradient based features (e.g., gists in interpreting images more accurately. Specifically, Kirsch [35]), fractal based features [36], run length matri- multi-parametric MR imaging (MP-MRI), which com- ces [37], and discrete cosine transform (DCT) [38]. Dif- bines two or more of T2-weighted MRI (T2w), diffusion- ferent classifiers are used for classification of pixels in weighted imaging (DWI), dynamic contrast enhanced prostate as cancerous and healthy among which support imaging (DCE), and spectroscopy has been investigated as vector machine (SVM), neural networks, naive Bayesian, a promising approach for prostate cancer diagnosis and and random forests are most frequently used. construction of detection algorithms [12–16]. By taking The usefulness of analysis of these texture features in advantage of the unique quantitative information pro- prostate cancer detection has been demonstrated in a vided by each individual imaging technique, MP-MRI can variety of applications. Madabhushi et al. [29] presented exploit the different characteristics of prostate tissue to the utility of combining multiple features in detecting improve differentiation between cancerous and surround- high likelihoods of prostatic adenocarcinoma from high- ing tissue. For example, cancerous tissue in the prostate resolution ex-vivo MRI (i.e., following radical prostate- gland may exhibit a moderate drop in signal in T2w [17] ctomy). In this method, the following feature sets were (which characterizes differences in transverse (spin-spin) extracted from 3D voxels to train an ensemble of classi- relaxation time of tissue), restricted diffusion in DWI [17] fiers: First- and second order statistical method, steerable (which characterizes diffusion of water in tissue), earlier Gabor filters, Gradient based features, and discrete cosine onset time, higher peak, and shorter peak time in DCE transform (DCT). The algorithm was applied to 5 MR [18] (which characterizes the concentration of an injected prostates and while specificity was high (e.g., 98 %), the gadolinium contrast agent over time as it passes into the sensitivity was reported to be low (36 % - 42 %). extracellular extravascular space of the tissue). Moreover, Duda et al. [39] defined multi-image texture analysis studies have demonstrated the ability of MP-MRI to direct (MITA) to characterize prostatic tissues in MR images. In biopsy with MRI/Ultrasound fusion techniques [19] and this method, multiple MRI acquisitions of corresponding to predict Gleason score [16] and tumour volume [20]. slices are performed to form a database of image n- Thepulse sequencethathas shownthe most promiseis tuples. This database includes T1-weighted (DCE), T2w, DWI in the peripheral zone and the combination of T2w and DWI of 19 patients. The process of validating any and DWI in the transition zone [21, 22]. The apparent MITA consisted of contouring Regions of Interest (ROI) diffusion coefficient (ADC) map in particular has shown by a clinician. Once all the ROIs were contoured, the cor- the most promise as a biomarker [16, 23–27]. Although responding slices from each sequence were combined to DCE is considered as part of MP-MRI, T2w+DWI is the form n-tuple images, from which texture features were most common MP-MRI because it has the most diagnos- extracted and concatenated in a feature vector used to tic value and does not require invasive contrast agent as train the classifiers. In addition to features used in [29], DCE does. fractal-based and run length features were also used. It Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 3 of 14 was shown that the MP-MRI performed better compared the other hand, improving the separability of cancerous to pairs of MRI modalities. Although accuracies of up to and healthy tissues in the images would have a signifi- 99 % was reported, the evaluation was only performed on cant impact on the performance of cancer auto-detection one slice (middle slice) for each modality. Moreover, the algorithms, potentially reducing the dependency on the ROIs used for classification and measuring the accuracy feature extraction methods. was considerably large (400 to 2,400 pixels). In this paper, we propose new MP-MRI texture fea- In another study, Litjens et al. [30] introduced the use of ture models that in addition to T2w and conventional a cascaded classifier in order to characterize benign con- DWI images, incorporate computed high-b diffusion- founders such as atrophy, inflammation, benign prostatic weighted imaging (CHB-DWI) [14] and the recently pro- hyperplasia (BPH), and prostatic intra-epithilial neoplasia posed correlated diffusion imaging (CDI) [42]. Compared (PIN) as the sources of challenge to diagnose malignant to DWI images, CHB-DWI and CDI have both shown prostate cancer. In this paper, the authors presented the initial promise to improve visual separability of cancer- biology behind the benign confounders and bridged it ous and healthy tissues in prostate, which can lead to with MRI sequences. The pathology annotations were improved performance of the proposed MP-MRI tex- propagated to MR images by registering the whole-mount ture feature models for detecting prostate cancer. One slides with MR images. Different features were extracted aspect of the proposed MP-MRI texture feature mod- from different MR images; second-order statistical and els is to use non-invasive modalities aiming for higher GaborfeaturesfromT2w,multi-scaleblobnessfilterfrom usability in the clinical practice. Hence, we did not use ADC images, and curve fitting and pharmacokinetic fea- DCE images in our models. Moreover, while most can- tures from DCE images [31]. The maximum relevance, cer detection algorithms use combined b-value images in minimum redundancy (mRMR) feature selection tech- the form of apparent diffusion coefficient (ADC) map, nique [40] was used to determine best features for sep- our proposed texture feature models utilize the individ- arating of cancer from three non-cancer classes (BPH, ual b-value images of DWI to extract additional sets of inflammation, and atrophy). A cascaded classifier was features leading to improved accuracies. For each modal- used to gradually determine whether the sample is can- ity, the best feature subsets are selected based on different cerous. MRI data of 31 patients with 44 corresponding performance evaluation criteria (sensitivity or specificity). histological H&E stained slides were used to evaluate the These best feature subsets are then combined to construct detection algorithm; a maximum accuracy of 76.4 % was the comprehensive feature set from which the final best achieved. feature subset is selected to be used by the classifier. To Tiwari et al. [13] proposed a method that combines the authors’ best knowledge, the proposed comprehensive structural and metabolic imaging data for separating texture feature models are the first that utilize all of the benign versus cancerous and high Gleason score ver- above-mentioned MP-MRI modalities and combine them sus low Gleason grade regions in MP-MRI that includes using best feature subsets to construct an optimal texture T2w and magnetic resonance spectroscopy (MRS). Sim- feature model. ilar set of features used in [29] was used with a random The proposed MP-MRI texture feature models are the forest classifier for detection where the evaluation was first attempt in designing comprehensive quantitative fea- performed on 29 patient studies; accuracy of 86 % was ture sequences or radiomics as a high dimensional mine- achieved. able feature space that can be used as both detection and Ozer et al. [15] extracted second-order statistical and prognostic tools for prostate cancer [43]. The proposed DCT features from 19 patients’ MP-MRI (T2w, DWI, and radiomics-driven models in this paper have been used DCE) and used two classifiers (SVM and Relevance Vector for prostate cancer detection and they can be augmented Machine (RVM) [41]) to autodetect prostate cancer. The for prognostic of prostate cancer as well. Studies on lung best achieved sensitivity and specificity were 78 % and and head-and-neck cancer patients have confirmed the 79 %, respectively. Glaister et al. [14] studied computed prognostic power of radiomics features when it comes high b-value DWI for localization of prostate cancer and to patient outcome prediction for personalized medicine it was found that using ultra-high b-value (≥ 2000s/mm ) [44, 45]. However, the prognostic capability of radiomics improves the separability of cancerous and healthy tissues features has not been fully investigated for prostate can- significantly. cer and this is a novel approach for identifying prostate The underlying challenge in all of these auto-detection tumours phenotypes. algorithms is whether there is enough separability In a previous work [46], the preliminary results for the between the cancerous and healthy tissues in a given proposed approach was reported. This paper is signifi- image. This means if the separability is poor, even sophis- cantly different than the initial work as follows. First, the ticated feature extraction algorithms may not have a sig- previous work only used T2w, ADC, CHB-DWI and CDI nificant effect on the accuracy of cancer detection. On whereas the current approach also utilizes four b-value Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 4 of 14 images. Second, in [46], only 19 features were used com- Diffusion-weighted imaging (DWI) pared to 96 features used in this work. As it will be seen in DWI is a promising imaging modality in which the sen- the “Results” section, using more data (more images using sitivity of tissue to Brownian motion of water molecules. b-value images and more features) makes the texture fea- The signal intensity is measured by applying pairs of ture model more accurate, in terms of the cancer detection opposing magnetic field gradient pulses, also known as accuracy. Third, in the previous work, feature selection lobe gradients [49]. The radio-frequency is excited by was not used while here we use feature selection for each applying a 180 degree pulse on the phase of all the spins. modality and also for combination of different modalities. The first gradient lobe, in turn, introduces a signal diphase Feature selection allows to build a more optimal texture in all the spins proportional to the gradient lobe area. feature model leading to more accurate results. Finally, The spins, then, evolve freely, divided into static spins only five patients datasets were used in the previous work and spins that move with respect to their relative posi- whereas 20 patients datasets have been used in this paper tion. The same intensity and polarity of the first gradient (6,535 cancerous and healthy tissue samples versus 40,975 lobe is used again for a second gradient lobe, where all the samples) allowing for a better validation of the proposed static spins align to the 90 degree pulse and the moving texture feature models. spins never recovering the phase. The moving spins create higher diphase among the spins, acquiring less signal than that of the static spins. The diffusion-weighted signal, S is Methods and materials formulated as: We propose MP-MRI texture feature models for prostate −bD cancer detection which take advantage of abundance of S = S e (1) data from different MR modalities to compute features used by the classifier. The goal is to combine features from where S is the signal intensity without the diffusion each imaging modality that best separates cancerous pix- weighting. The signal loss due to spins diphase, accord- els from healthy ones. In the following, we present the ing to Stejskal-Tanner sequence, can be controlled by b, imaging methods used in the proposed model, the feature which consists of amplitude and duration of the diffu- sets, and the proposed texture feature models. In addi- sion pulses, gradient intensity and the time between the tion, details about the image acquisition protocols and the two pulses as well as the gyromagnetic ratio, and D repre- performance measures are presented. sents the strength of the diffusion. The diffusion-weighted image (S) is usually generated with different b values which can be used to estimate apparent diffusion coef- Imaging methods ficient map (ADC) using the least-squares or maximum The main criteria for choosing imaging modalities used likelihood strategies [49]. The cancerous tissue in ADC is in the proposed texture feature models are twofold. First, usually represented by a darker intensity compared to the images that are part of well-known radiology reporting surrounding tissue. system. Second, they are acquired non-invasively, with no need for contrast agents, and can be collected in a single Computed high-b diffusion-weighted imaging (CHB-DWI) imaging session. Recently, a structured 5-scale reporting Previous research has shown that high b-value DWI system, PI-RADS, was proposed for consistent prostate 2 images (e.g., b-values greater than 1,000 s/mm ) allow MP-MRI reading [47] with subsequent studies confirm- for increased delineation between tumours and healthy ing its effectiveness with respect to biopsy results [48]. tissues [14, 50] which makes the prostate cancer detec- PI-RADS consisst of T2w, DWI (ADC) as well as DCE tion more robust. Nevertheless, due to hardware limita- images. Instead of using DCE which requires contrast tions, most MRI machines in practice do not produce agent, the proposed texture feature models use additional 2 DWI with b-values higher than 1,500 s/mm for prostate information available by DWI images which includes imaging. CHB-DWI is an alternative approach to obtain computed high b-value image, individual b-value images, high-b DWI in which a computational model is used to and correlated diffusion images. This subsection summa- reconstruct DWI at high b-values using low b-value DWI rizes the imaging methods used in the proposed MP-MRI acquisitions [14, 51]. For our experiments, we constructed feature models. 2 CHB-DWI with b-value at 2000s/mm using a Bayesian model with the same least squares estimation technique T2-weighted imaging (T2w) 2 used for ADC, extrapolating to the b-value of 2000s/mm . T2w is a MR imaging modality in which the sensitivity of tissue is characterized by measuring the relaxation time Correlated diffusion imaging (CDI) (spin-spin) of the applied magnetic field. The T2w image CDI [42] is a new diffusion magnetic resonance imaging of prostate usually shows a small reduction in signal in the modality, which takes advantage of the joint correlation in cancerous tissue [17]. signal attenuation across multiple gradient pulse strengths Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 5 of 14 Table 1 Summary of textural features used in the feature model and timings to not only reduce the dependency on the way diffusion gradient pulses are applied, but also improve Feature class Feature delineation between cancerous and healthy tissue. The First-order statistical features Mean, Standard deviation effectiveness of the delineation process depends on the Skewness models of the different types of tissue, since tumorous tis- Kurtosis sue has been empirically demonstrated to generate higher Energy, Contrast greyscale intensities at higher b-values. As such, in con- structing CDI, these properties are exploited where the Correlation, Variance utilized b-values are adjusted for a given application. The Inverse difference moment local correlation of signal attenuation across all b-values Sum average, Sum variance within a local sub-volume is calculated to better represent Sum entropy, Entropy the overall characterization of the water diffusion prop- Second-order statistical Difference variance erties of the tissue. The CDI signal is obtained via signal features (Haralick) Difference entropy mixing as follows [42]: Information measure of correlation n Homogeneity, Autocorrelation CDI(x) = ... S (x)...S (x)P(S (x), ... , S (x)| 0 n 0 n Dissimilarity, Cluster shade Cluster prominence V (x)) × dS (x)...dS (x) (2) 0 n Maximum probability Gabor filters 3 scales and 4 orientations where x denotes spatial location, b represents b values, S denotes the acquired signal, P denotes the conditional Kirsch filters 8 directions joint probability density function, and V (x) denotes the local subvolume around x. from different sets of images to construct radiomics fea- Feature extraction tures; a high-dimensional feature space that can be mined In order to separate the cancerous tissue from the healthy for different purposes such as detection or prognosis of one, a set of features is calculated on a given MR imag- cancer. Similar to conventional MP-MRI, the proposed ing modality (i.e., T2w, DWI, CHB-DWI, CDI, and indi- feature models include T2w and ADC modalities. They vidual b-value images). We incorporate four well-known also incorporate CHB-DWI, which has been shown to classes of texture features used in different studies to sep- increase separation between healthy and cancerous tis- arate cancerous and healthy tissues in prostate. These sue. As discussed in Section “Correlated diffusion imaging features include first- and second-order statistical features (CDI)”, as a new diffusion magnetic resonance modality, (Haralick [32, 33]), steerable Gabor filter features [34], and CDI has shown promise in separating healthy tissue from Kirsch filter features [35]. The first-order statistical fea- cancerous one. Although ADC incorporates all b-value tures include mean and standard deviation of grey-level images implicitly, individual b-value images may contain intensity, skewness, and kurtosis. Second-order statisti- information to help further distinguish healthy tissues cal features such as entropy and contrast are extracted from cancerous tissues. Therefore, we also incorporate from the gray-level co-occurrence matrix (GLCM) in four four b-value images into our proposed texture feature directions: 0 °, 45 °, 90 °, and 135 °. These texture features models. The following lists all the imaging modalities used include 18 features in each direction generating a total of by the proposed texture feature models for prostate cancer 72 features. Gabor features includes 12 features from three detection: scales and four orientations and Kirsch features include the maximum gradient in eight directions. As a result, I =T2w the proposed MP-MRI texture feature models consist of a I =ADC total of 96 features for each imaging modality: four from I = CHB-DWI: b-value at 2000s/mm first-order and 72 from second-order statistical features, I =CDI eight from Kirsch, and 12 from Gabor filters. Table 1 sum- I = b : b-value at 0s/mm 5 1 marizes all features used in the proposed texture feature I = b : b-value at 100s/mm 6 2 models. I = b : b-value at 400s/mm 7 3 I = b : b-value at 1000s/mm 8 4 Texture feature model Figure 1 shows the block diagram of the proposed texture For each modality, I , from the list above, the features feature models. The goal is to incorporate information described in Table 1 are calculated for a local window (e.g., Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 6 of 14 Fig. 1 Block diagram of the proposed texture feature models 3 × 3 pixels) sliding on the prostate gland. Each window on performance criteria used, the texture feature models is labeled either a tumour or non-tumour voxel. For each produce different results. imaging modality I , this gives a feature vector F . Once the best feature subsets for each imaging modality i i For each voxel in each image, the feature extraction was determined, the next step is to combine them to build function produces 96 features. A feature selection algo- different texture feature models (TFM) as follows: rithm determines a subset of features that contribute the TFM = T2w+ADC most to the separability of classes (e.g., cancerous vs. non- TFM = T2w+ADC+CHB-HBV cancerous tissues). This allows us to use the best features TFM = T2+CDI for each imaging modality when building the final texture TFM = T2w+ADC+CDI feature models. The feature selection algorithms usually TFM = T2w+ADC+HBV+CDI require the number of best features to be selected. For a TFM = T2w+ADC+HBV+CDI+b +b +b +b 6 1 2 3 4 given imaging modality I , to determine the optimal num- ber of features m , we perform an exhaustive search over The feature selection method is applied to each texture thefeature spacetoevaluatethe performanceofany num- feature model to build the final models. At this stage, the ber of features. This allows us to select m features as two performance criteria (sensitivity and specificity) are the feature vector F that produces the best results for a used to select the final best feature subsets for each tex- given imaging modality I . ture feature model. Algorithm 1 summarizes the texture To evaluate the performance of a given number of fea- feature model construction steps. tures, the accuracy or area under curve (AUC) for receiver operating characteristic (ROC) curve of the classification is usually used. Cancer cells in prostate usually constitute Algorithm 1 Texture Feature Model Construction a small fraction of the entire prostate gland (i.e., around 1: For each imaging modality I in training the set T = 1 %). This means that an accuracy of an algorithm may {I , I , ... , I }, apply feature extraction function: F = 1 2 n i be very high (e.g., 0.90) while it is unable to correctly Features(I ). locate the cancerous cells (i.e., low sensitivity). On the 2: For the feature set of each imaging modality in train- other hand, depending on the clinical procedures, differ- i ing set F , apply feature selection function: F = ent levels of sensitivity or specificity may be required. For F_Selection(F ) where m is thenumberofthe best i i example, for cancer screening programs, high sensitivity selected features for imaging modality I . (e.g., 0.90) is required where a moderate specificity (e.g., 3: Construct different combinations of selected features 0.60) is deemed to be adequate. On the other hand, for m 1 from different imaging modalities: F = F ∪ j i a procedure such as radical prostactomy, a high speci- m m 2 in F ... ∪ F where m = m + m + ... + m . 1 2 n i i 2 n ficity (e.g., 0.99) with moderate sensitivity (e.g., 0.60) is 4: Apply feature selection function to the constructed necessary to avoid unnecessary surgery. As a result, when m 0 m feature set F : F = F_Selection(F ) where m is j j j choosing the best feature subset, it is important to con- thenumberofthe finalbestselectedfeatures. sider different clinical scenarios by considering different 5: Apply classification to F . performance evaluation criteria for feature selection. To determine the best feature subsets, we examine two scenarios where in each scenario, it is assumed that either sensitivity or specificity has a higher priority in the per- For feature extraction function, we used the maximum formance evaluation of the proposed texture feature mod- relevance, minimum redundancy (mRMR) technique [40], els. As it will be seen in the results section, depending which is based on maximum relevance and minimum Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 7 of 14 redundancy of features. In this method, the feature subset ROI-based. In pixel-based approach [53], small neighbor- F is selected to satisfy the following criteria: hoods of pixels (e.g., 3 × 3) are considered to distinguish cancerous tissues from healthy ones. In other words, accu- racy determines the percentages of these neighborhood max D(F , c), D = MI(f ; c) (3) |F | m that were correctly labeled as cancerous or healthy. ROI- f ∈F based approach [29, 39, 54] is similar to pixel-based with the difference that it uses larger neighborhoods of pixels min R(F , c), R = MI(f ; f ) (4) i j m 2 |F | (e.g., 50 × 50) for calculating accuracy measures. f ,f ∈F i j In evaluating the performance of the proposed tex- where F is the best feature subset that we would like ture feature models in this paper, we use the pixel-based to find, c is the target class, f is a feature and MI is the approach so that the accuracy measurements are calcu- mutual information function. D and R are the relevance lated more precisely. As ground-truth, all MP-MR images and redundancy of features, respectively. Maximum rele- were reviewed and marked as healthy and cancerous vance guarantees that the selected features have the high- tissue by a radiologist with 18 and 13 years of experi- est shared information with the target class and minimum ence interpreting body and prostate MRI, respectively. In redundancy ensures that the redundant features are elim- addition, for cases with cancer, the MP-MRI images and inated. For the classifier, we used the SVM implemented expert annotations were compared to the corresponding in [52]. histopathology data, obtained through radical prostate- The proposed radiomics-driven cancer detection mod- ctomy with Gleason score seven and above, as ground- els combine a plethora of data from different imaging truth to confirm the accuracy of the MP-MRI markings. modalities of MP-MRI to construct comprehensive tex- ture feature models which can be used for both detection Results and prognosis purposes in prostate cancer. Figure 2 shows sensitivity and specificity for all 8 MP-MRI modalities using different number of best features (e.g., 10 Image data features to 96 features). For each modality, 40,975 sam- MRIdataof20patients(17 with cancer andthree ples (40,369 healthy and 606 cancerous samples confirmed without cancer) were acquired using a Philips Achieva by the radiologist) was used for the leave-one-patient-out 3.0T machine at Sunnybrook Health Sciences Centre, cross-validation. Toronto, Ontario, Canada. All data was obtained retro- Tables 3 and 4 show the quantitative results for dif- spectively under the local institutional research ethics ferent modalities and combinations of modalities. Using board (Research Ethics Board of Sunnybrook Health Sci- sensitivity as feature selection criteria (Table 3), the sen- ences Centre). For each patient, the following MP-MRI sitivity of the texture feature models reaches 0.86 using modalities were obtained (Table 2): T2w, DWI, and CDI. CDI alone. It is interesting to observe that CDI also out- The patients’ age ranged from 53 to 83. Table 2 sum- 1 performs the conventional MP-MRI (i.e., TFM )and marizes the information about the 20 patients’ datasets combination of conventional MP-MRI and CHB-DWI used in this research, which includes displayed field of (i.e., TFM ) (0.86 vs. 0.77 and 0.86 vs 0.69, respectively). view (DFOV), resolution, echo time (TE), and repeti- Although CDI alone gives the best results for sensitivity tion time (TR). Images were processed in the ProCanVAS (0.86), the full feature sets model (i.e., TFM ) produces the (Prostate Cancer Visual Analysis System) platform devel- best results when considering specificity, accuracy, and oped at Sunnybrook Research Institute, Toronto, ON, AUC as well (0.82, 0.82, and 0.86, respectively). Compar- Canada. Each modality (e.g., CDI) provided 40,975 sam- ing TFM to all other models in Table 3, at least 2 metrics ples used for the leave-one-patient-out cross-validation of out of 4 are significantly different than each of other mod- the algorithms. els. For example, comparing TFM to TFME ,the P values 6 5 for specificity and accuracy via Wilcoxon signed-rank test Evaluation metrics are 0.006 and 0.01, respectively. To evaluate the performance of cancer detection algo- Table 4 shows the performance results for using speci- rithms, two approaches may be used: pixel-based and ficity as performance evaluation criteria for feature selec- tion. It is observed that compared to the previous Table 2 Description of the prostate T2w, DWI, and CDI images approach (Table 3), the full feature sets model (TFM ) 2 3 Modality DFOV (cm)Resolution(mm ) TE (ms) TR (ms) improves the specificity by 0.06 (0.88). This was expected since the performance evaluation criteria used for fea- T2w 22 × 22 0.49 × 0.49 × 3 110 4,687 ture selection affects the final results. Thus, as discussed DWI 20 × 20 1.56 × 1.56 × 3 61 6,178 in Section “Texture feature model”, depending on the CDI 20 × 20 1.56 × 1.56 × 3 61 6,178 clinical scenario, one can choose different performance Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 8 of 14 Fig. 2 Performance results for different modalities (T2w, ADC, CHB-DWI, CDI, and 4 DWI images at different b values) across all features Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 9 of 14 Table 3 Evaluation results for prostate cancer detection: Feature selection based on Sensitivity (Results are shown with 95 % confidence interval) Imaging Number of Sensitivity Specificity Accuracy AUC modality features T2w 96 0.71 [0.54 0.89] 0.44 [0.39 0.49] 0.45 [0.40 0.50] 0.58 [0.48 0.68] CHB-DWI 90 0.73 [0.58 0.88] 0.78 [0.71 0.85] 0.77 [0.71 0.84] 0.79 [0.73 0.85] ADC 20 0.76 [0.64 0.88] 0.59 [0.51 0.67] 0.60 [0.52 0.67] 0.68 [0.63 0.74] CDI 96 0.86 [0.76 0.97] 0.80 [0.75 0.85] 0.79 [0.74 0.84] 0.85 [0.81 0.90] TFM = T2w+ADC 20 0.77 [0.64 0.91] 0.57 [0.49 0.65] 0.59 [0.51 0.66] 0.68 [0.62 0.74] TFM =T2w+ADC+CHB-DWI 208 0.69 [0.54 0.84] 0.79 [0.73 0.84] 0.78 [0.73 0.84] 0.78 [0.72 0.85] TFM =T2w+CDI 196 0.85 [0.75 0.96] 0.81 [0.76 0.86] 0.80 [0.76 0.85] 0.85 [0.81 0.90] TFM =T2w+ADC+CDI 216 0.86 [0.76 0.96] 0.81 [0.76 0.86] 0.80 [0.76 0.85] 0.85 [0.81 0.90] TFM =T2w+ADC 300 0.86 [0.75 0.96] 0.81 [0.77 0.86] 0.81 [0.77 0.85] 0.86 [0.83 0.90] +CHB-DWI+CDI TFM = T2w+ADC +CHB-DWI+CDI 416 0.86 [0.75 0.97] 0.82 [0.78 0.87] 0.82 [0.78 0.86] 0.86 [0.81 0.91] +b +b +b +b 1 2 3 4 evaluation criteria to better suit the clinical procedure sensitivity (0.90 vs. 0.87). Figure 4 shows the ROC curves requirements. Comparing TFM to all other models in for all six models as well as individual imaging modali- Table 4 (except for TFM ), at least 2 metrics out of four ties discussed in Section “Texture feature model”. It is seen are significantly different than each of other models. For that the combination of all imaging modalities, TFM , example, comparing TFM to TFM ,the P values for gives the best results in terms of AUC (0.90). This result 6 5 specificity and accuracy via Wilcoxon signed-rank test is significantly different with respect to any other imaging are 0.01. Comparing TFM to TFM , the two models are modality or texture feature model where P < 0.009. 6 3 significantly different with respect to AUC (P = 0.01). Table 5 shows the optimal results with the target of max- Tables 3 and 4 show the result when the goal was to imizing sensitivity, specificity, or AUC. As it can be seen, maximize sensitivity (Table 3) or specificity (Table 4). choosing a target yields the best result for the selected tar- Figure 3 shows the combinations of all eight imaging get. Setting AUC as the target maximizes the AUC (0.90) modalities (TFM )withbestfeature subsetsbased on sen- and at the same time generates more balanced results with sitivity and specificity with the objective of maximizing for respect to sensitivity and specificity (0.84 and 0.86). Using AUC. It can be seen that using specificity as performance sensitivity as the performance evaluation criteria maxi- evaluation criteria gives a higher best AUC compared to mizes the result for sensitivity (0.86). Using specificity as Table 4 Evaluation results for prostate cancer detection: Feature selection based on specificity (results are shown with 95 % confidence interval) Imaging Number of Sensitivity Specificity Accuracy AUC modality features T2w 10 0.66 [0.50 0.81] 0.47 [0.42 0.53] 0.48 [0.43 0.53] 0.57 [0.48 0.66] CHB-DWI 10 0.69 [0.52 0.86] 0.82 [0.75 0.88] 0.81 [0.75 0.87] 0.76 [0.68 0.84] ADC 96 0.73 [0.60 0.85] 0.62 [0.55 0.70] 0.63 [0.56 0.71] 0.70. [0.64 0.76] CDI 10 0.82 [0.69 0.94] 0.85 [0.80 0.89] 0.84 [0.80 0.88] 0.84 [0.78 0.89] TFM = T2w+ADC 110 0.72 [0.59 0.86] 0.63 [0.55 0.70] 0.64 [0.56 0.71] 0.69 [0.63 0.75] TFM =T2w+ADC 40 0.66 [0.50 0.82] 0.77 [0.71 0.83] 0.77 [0.71 0.82] 0.73 [0.65 0.81] +CHB-DWI 20 0.78 [0.65 0.91] 0.86 [0.82 0.90] 0.86 [0.82 0.89] 0.84 [0.78 0.90] TFM =T2w+CDI 40 0.77 [0.63 0.90] 0.86 [0.82 0.90] 0.85 [0.81 0.89] 0.84 [0.79 0.89] TFM =T2w+ADC+CDI TFM =T2w+ADC 50 0.78 [0.64 0.91] 0.86 [0.82 0.90] 0.85 [0.82 0.89] 0.84 [0.78 0.90] +CHB-DWI+CDI TFM =T2w+ADC +CHB-DWI+CDI 130 0.80 [0.69 0.91] 0.88 [0.85 0.92] 0.88 [0.84 0.91] 0.88 [0.83 0.93] +b +b +b +b 1 2 3 4 Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 10 of 14 the performance evaluation criteria maximizes the result for specificity (0.88) and AUC (0.90), depending on the selected target. Figure 5 shows an example for all four modalities which include T2w, ADC, CHB-DWI, and CDI. As it can be seen, CDI (Fig. 5d) is the only modality that clearly shows a bright nodule where a tumour is located (confirmed by histopathology data - Fig. 6). Discussion Automated prostate cancer detection has been investi- gated by different research groups in the field. The under- lying building blocks of these algorithms consist of feature extraction and classification applied to local windows of pixels in the images. Most algorithms in the literature usually incorporate few imaging sequences into their pro- posed feature models. For example, the work presented in Fig. 4 ROC for different texture feature models [39] used three sequences (T1w, T2w, and DCE) to build the texture feature model. In contrast, in this paper, we have taken one step further by incorporating information from all available MR imaging data which includes T2w, One important aspect in the clinical workflow for ADC, and different b-value images of DWI (i.e., b-values prostate cancer detection is the targeted clinical proce- at 0, 100, 400, and 1000s/mm ). Moreover, we incorpo- dure. For example, cancer screening programs impose rated computed high-b DWI (CHB-DWI) [14] as well as different performance requirements compared to proce- correlated diffusion imaging (CDI) [42] into our model. dures such as radical prostatectomy. We designed the Adding these two extra imaging modalities enriched our proposed texture feature model accounting for such texture feature models in terms of the diversity of data requirements where the performance of the model can where 6 different models were developed and evalu- be optimized for sensitivity, specificity, or the area under ated (Sections “Texture feature model” and “Results”). the ROC curve. For example, to use the proposed tex- As a quantitative radiomics approach for prostate can- ture feature model for cancer screening, sensitivity can be cer detection, we used a comprehensive texture feature used as the performance evaluation criteria to steer the model which incorporated eight different imaging modal- feature selection process which would lead to best result ities where each modality contributed with its best feature for sensitivity (0.86) with reasonable results for specificity subset to the ultimate texture feature model in which all (0.82). For cases where higher specificity is required, one modalities were combined. can use specificity as the performance evaluation criteria to optimize the results for specificity (0.88) with accept- able sensitivity (0.80). Our experiments showed that using specificity as the performance evaluation criteria can also maximize the results for AUC (0.90) which leads to a bal- anced results for sensitivity and specificity; 0.84 and 0.86, respectively. The fact that the proposed model is flexible in terms of optimizing the results for the procedure it is used for makes it more practical. This is another novel aspect of the proposed model in this paper which to the authors’ best knowledge has not been fully explored in the literature. The limitations of our research include a relatively lim- ited number of datasets (20 patients) and targeting only Gleason score of seven and above. Evaluating the pro- posed model using a larger dataset and considering lower Gleason scores (e.g., six) will add more confidence to the reliability of the model which will be done as future Fig. 3 AUC based on using sensitivity and specificity as performance work. Other limitation is that the proposed model was not evaluation criteria assessed by clinicians to investigate whether it improves Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 11 of 14 Table 5 Evaluation results for prostate cancer detection: Feature selection based on Sensitivity and Specificity (Results are shown with 95 % confidence interval) Target Performance evaluation criteria Sensitivity Specificity AUC Sensitivity Sensitivity 0.86 [0.75 0.97] 0.82 [0.78 0.87] 0.86 [0.81 0.91] Specificity Specificity 0.80 [0.69 0.91] 0.88 [0.85 0.92] 0.88 [0.83 0.93] AUC Specificity 0.84 [0.76 0.91] 0.86 [0.82 0.91] 0.90 [0.88 0.93] the clinical readings by radiologists. Similar to the work Section “Results”, boosted the results significantly. We reported in [55], clinical assessment of the proposed auto- have developed an enhanced version of CDI, called dual- detection model will be performed to evaluate its effect stage correlated diffusion imaging (D-CDI) which has on the clinicians’ performance. Finally, given the fact that shown promise in enhancing separability of cancerous and cancerous pixels are a small fraction of the entire prostate healthy tissue in prostate MRI compared to CDI [56]. gland, it is possible that the reported specificity results are As future work, we will incorporate D-CDI to the pro- an overestimation. A larger and more diversified dataset posed texture feature model to investigate the effect on will help to investigate this more thoroughly. performance. We will also investigate developing a hybrid Our proposed texture feature model incorporated CDI morphological-textural feature model for prostate cancer as one of the imaging modalities which as shown in where in addition to texture analysis, the morphological Fig. 5 a T2w does not clearly show a tumour although there is mild signal alteration in the left peripheral zone (arrow). b ADC does not clearly show a tumour (arrow). c CHB-DWI of 2000 s/mm shows no tumour (arrow). d CDI clearly shows a bright nodule (arrow) corresponding to tumour Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 12 of 14 Fig. 6 Corresponding axial hematoxylin and eosin stained tissue showing a Gleason 7 (4+3) tumor circled in red corresponding to the lesion identified best on the CDI images in Fig. 5-d characteristics (e.g., shape) of candidate regions are taken optimal texture feature models. A SVM classifier was into account to detect cancer. A preliminary work on trained via leave-one-patient-out setting to classify the morphological feature model has been presented in [57] new cases. The proposed MP-MRI texture feature models upon which we will extend and build the hybrid model. showed promise in accurate detection of prostate cancer. The b-value images of DWI are usually distorted due to Endnote patient movement during the image acquisition which The conventional MP-MRI refers to the combination may reduce cancer separability. We have presented pre- of T2w and DWI, which is represented as ADC. Thus, liminary results for co-registering the b-value images to throughout the paper, T2w+DWI and T2w+ADC are compensate for patient movement [58]. As future work, used interchangeably. we will incorporate this co-registration algorithm into our proposed texture feature models to investigate the effect Competing interests on the accuracy of cancer detection. Finally, normalized The authors have declared that no competing interests exist. entropy has been shown to be a strong predictor of patient survival rate for lung and renal cell cancers in CT images Authors’ contributions Conceived and designed the methodology: FKh, AW, MAH. Performed the [59]. The future work will also involve investigating the experiments: FKh, AW. Analyzed the data: FKh, AW, MAH. Wrote the paper: efficacy of the proposed texture feature model in this FKh, AW, MAH. All authors read and approved the final manuscript. paper for normalized MP-MRI entropy characterization Acknowledgements of prostate cancer. This research has been supported by the Ontario Institute of Cancer Research (OICR), Canada Research Chairs programs, Natural Sciences and Engineering Conclusion Research Council of Canada (NSERC), and the Ministry of Research and Innovation of Ontario. The authors would like to thank Prof. Rajiv Chopra, In this paper, we introduced new multi-parametric MRI Department of Radiology, UT Southwestern Medical Center, for providing texture feature models for prostate cancer detection. Our histology images. new MP-MRI texture feature models add two new imag- Author details ing modalities, computed high-b DWI and correlated Department of Medical Imaging, University of Toronto, Toronto, ON, Canada. diffusion imaging, to the most commonly used MP-MRI, Physical Sciences, Sunnybrook Research Institute, Toronto, ON, Canada. T2w+ADC. As a quantitative radiomics approach for Department of Systems Design Engineering, University of Waterloo, Waterloo, ON, Canada. automatic detection of prostate cancer, a comprehensive set of texture features were calculated for the conventional Received: 2 December 2014 Accepted: 9 July 2015 MP-MRI and new MP-MRI texture feature models. A fea- ture selection method was used to select the optimal fea- References tures for each modality. Two different performance eval- 1. Canadian Cancer Society. Canadian Cancer Statistics. 2014. http://www. uation criteria were used for feature selection to reflect cancer.ca/~/media/cancer.ca/CW/cancer%20information/cancer different clinical workflows in prostate cancer diagnosis. %20101/Canadian%20cancer%20statistics/Canadian-Cancer-Statistics- 2014--EN.pdf. Thebestfeature subsetswerethencombinedtoconstruct Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 13 of 14 2. American Cancer Society. Cancer Facts and Figures. 2014. http://www. 22. Oto A, Kayhan A, Jiang Y, Tretiakova M, Yang C, Antic T, et al. Prostate cancer.org/research/cancerfactsstatistics/cancerfactsfigures2014/. cancer: differentiation of central gland cancer from benign prostatic 3. Jemal A, Bray F, Center MM, Ferlay J, Ward E, Forman D. Global cancer hyperplasia by using diffusion-weighted and dynamic contrast-enhanced statistics. CA: A Cancer J Clin. 2011;61(2):69–90. MR imaging. Radiology. 2010;257:715–23. 4. Canadian Cancer Society. Canadian Cancer Statistics. 2011. http://www. 23. Shigemura K, Yamanaka N, Yamashita M. Can diffusion-weighted cancer.ca/~/media/cancer.ca/CW/cancer%20information/cancer magnetic resonance imaging predict a high Gleason score of prostate %20101/Canadian%20cancer%20statistics/Canadian-Cancer-Statistics- cancer? Korean J Urol. 2013;54:234–38. 2011-EN.pdf. 24. Caivano R, Rabasco P, Lotumolo A, Cirillo P, D’Antuono F, Zandolino A, 5. Andriole GL, Crawford D, Grubb RL, Buys SS, Chia D, Church TR, et al. et al. Comparison between Gleason score and apparent diffusion PLCO Project Team. Mortality results from a randomized prostate-cancer coefficient obtained from diffusion-weighted imaging of prostate cancer screening trial. N Engl J Med. 2009;360:1310–9. patients. Cancer Invest. 2013;31:625–9. 6. Schroder FH, Hugosson J, Roobol MJ, Tammela TL, Ciatto S, Nelen V, 25. Bae H, Yoshida S, Matsuoka Y, Nakajima H, Ito E, Tanaka H, et al. Apparent et al. ERSPC Investigators. Screening and prostate-cancer mortality in a diffusion coefficient value as a biomarker reflecting morphological and randomized European study. N Engl J Med. 2009;360:1320–8. biological features of prostate cancer. Int Urol Nephrol. 2014;46(3):555–61. 7. Greene KL, Albertsen PC, Babaian RJ, Albertsen PC, Carter HB, Gann PH, 26. Nagarajan R, Margolis D, Raman S, Sheng K, King C, Reiter R, et al. et al. Prostate Specific Antigen Best Practice Statement 2009 Update. J Correlation of Gleason scores with diffusion-weighted imaging findings Urol. 2009;182(5):2232–41. of prostate cancer. Adv Urol. 2012;2012:374805. 8. Vickers AJ, Sjoberg DD, Ulmert D, Vertosick E, Roobol MJ, Thompson I, 27. Bittencourt LK, Barentsz JO, de Miranda LC, Gasparetto EL. Prostate MRI: et al. Empirical estimates of prostate cancer overdiagnosis by age and diffusion-weighted imaging at 1.5T correlates better with prostatectomy prostate-specific antigen. BMC Medicine. 2014;12:26. Gleason Grades than TRUS-guided biopsies in peripheral zone tumours. 9. Loeb S, Vellekoop A, Ahmed HU, Catto J, Emberton M, Nam R, et al. Eur Radiol. 2012;22:468–75. Systematic review of complications of prostate biopsy. Eur Urol. 28. Langer DL, van der Kwast TH, Evans AJ, Sun L, Yaffe MJ, Trachtenberg J, 2013;64(6):876–92. et al. Intermixed normal tissue within prostate cancer: effect on MR 10. Schroder FH, Hugosson J, Roobol MJ, Tammela TLJ, Ciatto S, Nelen V, imaging measurements of apparent diffusion coefficient and T2–sparse et al. ERSPC Investigators. “Prostate-cancer mortality at 11 years of versus dense cancers. Radiology. 2008;249:900–8. follow-up. N Engl J Med. 2012;366(11):981–90. 29. Madabhushi A, Feldman MD, Metaxas DN, Tomaszeweski J, Chute D. 11. Nam RK, Saskin R, Lee Y, Liu Y, Law C, Klotz LH, et al. Increasing hospital Automated detection of prostatic adenocarcinoma from high-resolution admission rates for urological complications after transrectal ultrasound ex vivo MRI. IEEE Trans Med Img. 2005;24(12):1611–25. guided prostate biopsy. Division of Urology, Sunnybrook Research 30. Litjensa GJS, Elliottb R, Shihc N, Feldmanc M, Barentsza JO, Institute. 2013. doi: 10.1016/j.juro.2009.11.043. Hulsbergen - van de Kaaa CA, et al. Distinguishing prostate cancer from 12. Arumainayagam N, Ahmed HU, Moore CM, Freeman A, Allen C, benign confounders via a cascaded classifier on multi-parametric MRI. Sohaib SA, et al. Multiparametric MR Imaging for Detection of Clinically Medical Imaging. 2014. doi:10.1117/12.2043751. Significant Prostate Cancer: A Validation Cohort Study with Transperineal 31. Vos PC, Barentsz JO, Karssemeijer N, Huisman HJ. Automatic Template Prostate Mapping as the Reference Standard. Radiology. computer-aided detection of prostate cancer based on multiparametric 2013;268(3):761–769. magnetic resonance image analysis. Phys Med Biol. 2012;57:1527–42. 13. Tiwari P, Kurhanewicz J, Madabhushi A. Multi-kernel graph embedding 32. Haralick RM, Shanmugan K, Dinstein I. Textural features for image for detection, gleason grading of prostate cancer via MRI-mrs. Med Image classification. IEEE Trans Syst Man Cybern. 1973;SMC-3:610–21. Anal. 2013;17:219–35. 33. Soh L, Tsatsoulis C. Texture Analysis of SAR Sea Ice Imagery Using Gray 14. Glaister J, Cameron A, Wong A, Haider MA. Quantitative investigative Level Co-Occurrence Matrices. IEEE Trans Geosci Remote Sensing. analysis of tumour separability in the prostate gland using ultra-high 1999;37(2):. doi:10.1109/36.752194. b-value computed diffusion imaging. In: EMBC’2012; 2012. p. 420–423. 34. Gabor D. Theory of communication. J Inst Electr Eng. 1946;93:429–41. doi:10.1109/EMBC.2012.6345957. 35. Kirsch RA. Computer determination of the constituent structure of 15. Ozer S, Haider MA, Langer DL, van der Kwast TH, Evans AJ, Wernick MN, biological images. Comput Biomed Res. 1971;4:315–28. et al. Prostate cancer localization with multispectral MRI based on 36. Chen C, Daponte JS, Fox MD. Fractal feature analysis and classification in relevance vector machines. In: 2009 IEEE International Symposium on medical imaging. IEEE Trans Med Imag. 1989;8(2):133–42. Biomedical Imaging From Nano to Macro; 2009. p. 73–76. 37. Galloway MM. Texture analysis using gray level run lengths. Comp Graph doi:10.1109/ISBI.2009.5192986. Im Proc. 1975;4(2):172–9. 16. Turkbey B, Mani H, Shah V, Rastinehad AR, Bernardo M, Pohida T, et al. 38. Khayam SA. The Discrete, Cosine Transform (DCT): Theory and Multiparametric 3T prostate magnetic resonance imaging to detect Application: Department of Electrical & Computer Engineering, Michigan cancer: histopathological correlation using prostatectomy specimens State University; 2003. processed in customized magnetic resonance imaging based molds. J 39. Duda D, Kretowski M, Mathieu R, Crevoisier RD, Bezy-Wendling J. Urol. 2011;186:1818–24. Multi-Image Texture Analysis in Classification of Prostatic Tissues from 17. Haider MA, van der Kwast TH, Tanguay J, Evans AJ, Hashmi A, MRI. Preliminary Results. Information Technologies in Biomedicine, 139 Lockwood G, et al. Combined T2-weighted and diffusion-weighted MRI Volume 3, Advances in Intelligent Systems and Computing 283, 2014. for localization of prostate cancer. AJR Am J Roentgenol. 2007;189(2): doi:10.1007/978-3-319-06593-9_13. 323–8. 40. Peng H, Long F, Ding C. Feature selection based on mutual information 18. Langer DL, van der Kwast TH, Evans AJ, Plotkin A, Trachtenberg J, criteria of maxdependency, max-relevance, and min-redundancy. IEEE Wilson BC, et al. Prostate Tissue Composition and MR Measurements: Trans Pattern Anal Mach Intell. 2005;27:1226–38. Investigating the Relationships between ADC, T2, K-trans, v-e, and 41. Tipping ME. Sparse Bayesian learning and the relevance vector machine. Corresponding Histologic Features. Radiology. 2010;255(2):485–94. J Mach Learn Res. 2001;1:211–244. 19. Vourganti S, Rastinehad A, Yerram NK, Nix J, Volkin D, Hoang A, et al. 42. Wong A, Glaister J, Cameron A, Haider M. Correlated Diffusion Imaging. Multiparametric magnetic resonance imaging and ultrasound fusion BMC Medical Imaging. 201313–26. doi:10.1186/1471-2342-13-26. biopsy detect prostate cancer in patients with prior negative transrectal 43. Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RG, ultrasound biopsies. J Urol. 2012;188:2152–7. Granton P, et al. Radiomics: extracting more information from medical 20. Turkbey B, Mani H, Aras O, Rastinehad AR, Shah V, Bernardo M, et al. images using advanced feature analysis. Eur J Cancer. 2012;48(4):441–6. Correlation of magnetic resonance imaging tumor volume with doi:10.1016/j.ejca.2011.11.036. histopathology. J Urol. 2012;188: 44. Aerts HJ, Velazquez ER, Leijenaar RT, Parmar C, Grossmann P, 1157–63. Carvalho S, et al. Decoding tumour phenotype by noninvasive imaging 21. Tan CH, Wei W, Johnson V, Kundra V. Diffusion-weighted MRI in the using a quantitative radiomics approach. Nat Commun. 2014;5:4006. detection of prostate cancer: meta-analysis. AJR Am J Roentgenol. 45. Gevaert O, Xu J, Hoang CD, Leung AN, Xu Y, Quon A, et al. Non-small 2012;199:822–829. cell lung cancer: identifying prognostic imaging biomarkers by leveraging Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 14 of 14 public gene expression microarray data–methods and preliminary results. Radiology. 2012;264(2):387–96. doi:10.1148/radiol.12111607, 2012. 46. Khalvati F, Modhafar A, Cameron A, Wong A, Haider MA. Multi- Parametric Diffusion Magnetic Resonance Imaging Texture Feature Model for Prostate Cancer Analysis. International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), Computational Diffusion MRI Workshop. Boston, Mass., USA; 2014. p. 79–88. 47. Barentsz JO, Richenberg J, Clements R, Choyke P, Verma S, Villeirs G, et al. ESUR prostate MR guidelines 2012. Eur Radiol. 2012;22(4):746–57 22:746–757, 2012. 48. Portalez D, Mozer P, Cornud F, Renard-Penna R, Misrai V, Thoulouzan M, et al. “Validation of the European Society of Urogenital Radiology scoring system for prostate cancer diagnosis on multiparametric magnetic resonance imaging in a cohort of repeat biopsy patients. Eur Urol. 2012;62(6):986–96. doi:10.1016/j.eururo.2012.06.044. Epub. 49. Koh DM, Padhani AR. Diffusion-weighted MRI: a new functional clinical technique for tumour imaging. Br J Radiol. 2006;79:633–5. 50. Rosenkrantz AB, Taneja SS. Radiologist, be aware: ten pitfalls that confound the interpretation of multiparametric prostate MRI. AJR Am J Roentgenol. 2014;202:109–20. 51. Shafiee MJ, Haider SA, Wong A, Lui D, Cameron A, Modhafar A, et al. Apparent Ultra-High b-value Diffusion-Weighted Image Reconstruction via Hidden Conditional Random Fields. IEEE Trans Med Imaging. 2015;34(5):1111–24. 52. Chang C, Lin C. LIBSVM: A library for support vector machines. ACM Trans Intell Syst Technol. 2011;2(3):27:1–27:27. 53. Ozer S, Langer DL, Liu X, Haider MA. Supervised and unsupervised methods for prostate cancer segmentation with multispectral MRI. Med Phys. 2010;37(3):1873–83. 54. Rezaeilouyeh H, Mahoor MH, Zhang J, La Rosa FG, Chang S, Werahera PN. Diagnosis of Prostatic Carcinoma on Multiparametric Magnetic Resonance Imaging Using Shearlet Transform. In: 36 th Intenational Conference of the IEEE Engineering In Medicine and Biology Society (EMBC); 2014. p. 6442–6445. doi:10.1109/EMBC.2014.6945103. 55. Hambrock T, Vos PC, Hulsbergen-van de Kaa CA, Barentsz JO, Huisman HJ. Prostate Cancer: Computer-aided Diagnosis with Multiparametric 3-T MR Imaging–Effect on Observer Performance. Radiology. 2013;266:521–30. 56. Wong A, Khalvati F, Haider MA. Dual-Stage Correlated, Diffusion Imaging. International Symposium on Biomedical Imaging. Brooklyn, NY, USA; 2015, pp. 75–78. 57. Cameron A, Modhafar A, Khalvati F, Lui D, Shafiee MJ, Wong A, et al. Multi-Parametric MRI Prostate Cancer Analysis via a Hybrid Morphological-Textural Model. In: International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). Chicago, Ill., USA; 2014. p. 3357–3360. 58. Cho DS, Khalvati F, Wong A, Haider MA. Prostate DWI Co-Registration via Maximization of Hybrid Statistical Likelihood and Cross-Correlation for Improved ADC and Computed Ultra-High b-value DWI Calculation. In: Proceedings of International Society for Magnetic Resonance in Medicine (ISMRM). Toronto, ON, Canada; 2015. p. 3748. 59. Khalvati F, Wong A, Bjarnason GA, Haider MA. Semi-Automatic Normalized Entropy Characterization of Metastatic Renal Cell Cancer via Spatio-Textural Tumour Classification. In: International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). Chicago, Ill., USA; 2014. Submit your next manuscript to BioMed Central and take full advantage of: • Convenient online submission • Thorough peer review • No space constraints or color figure charges • Immediate publication on acceptance • Inclusion in PubMed, CAS, Scopus and Google Scholar • Research which is freely available for redistribution Submit your manuscript at www.biomedcentral.com/submit http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png BMC Medical Imaging Springer Journals

Automated prostate cancer detection via comprehensive multi-parametric magnetic resonance imaging texture feature models

Loading next page...
 
/lp/springer-journals/automated-prostate-cancer-detection-via-comprehensive-multi-parametric-g3MXAIJ6vW

References (67)

Publisher
Springer Journals
Copyright
Copyright © 2015 by Khalvati et al.
Subject
Medicine & Public Health; Imaging / Radiology
eISSN
1471-2342
DOI
10.1186/s12880-015-0069-9
pmid
26242589
Publisher site
See Article on Publisher Site

Abstract

Background: Prostate cancer is the most common form of cancer and the second leading cause of cancer death in North America. Auto-detection of prostate cancer can play a major role in early detection of prostate cancer, which has a significant impact on patient survival rates. While multi-parametric magnetic resonance imaging (MP-MRI) has shown promise in diagnosis of prostate cancer, the existing auto-detection algorithms do not take advantage of abundance of data available in MP-MRI to improve detection accuracy. The goal of this research was to design a radiomics-based auto-detection method for prostate cancer via utilizing MP-MRI data. Methods: In this work, we present new MP-MRI texture feature models for radiomics-driven detection of prostate cancer. In addition to commonly used non-invasive imaging sequences in conventional MP-MRI, namely T2-weighted MRI (T2w) and diffusion-weighted imaging (DWI), our proposed MP-MRI texture feature models incorporate computed high-b DWI (CHB-DWI) and a new diffusion imaging modality called correlated diffusion imaging (CDI). Moreover, the proposed texture feature models incorporate features from individual b-value images. A comprehensive set of texture features was calculated for both the conventional MP-MRI and new MP-MRI texture feature models. We performed feature selection analysis for each individual modality and then combined best features from each modality to construct the optimized texture feature models. Results: The performance of the proposed MP-MRI texture feature models was evaluated via leave-one-patient-out cross-validation using a support vector machine (SVM) classifier trained on 40,975 cancerous and healthy tissue samples obtained from real clinical MP-MRI datasets. The proposed MP-MRI texture feature models outperformed the conventional model (i.e., T2w+DWI) with regard to cancer detection accuracy. Conclusions: Comprehensive texture feature models were developed for improved radiomics-driven detection of prostate cancer using MP-MRI. Using a comprehensive set of texture features and a feature selection method, optimal texture feature models were constructed that improved the prostate cancer auto-detection significantly compared to conventional MP-MRI texture feature models. Background cause of cancer death in men in the United States with Prostate cancer is the most common form of cancer diag- an estimated 29,480 deaths in 2014 [2]. Given that the nosed in North American men, with roughly 23,500 new median patient survival time for metastatic prostate can- cases in 2014 in Canada [1] and 233,000 new cases in 2014 cer ranges from 12.2 to 21.7 months [3], early diagnosis in the United States [2]. Furthermore, prostate cancer is of clinically significant prostate cancer would have signifi- the third leading cause of cancer death in Canadian men cant benefits to patient care. This is particularly true given with an estimated 4,000 deaths [1], and second leading that the five-year survival rate after diagnosis for patients with prostate cancer at the non-metastatic stage is 96 % in Canada [4]. *Correspondence: farzad.khalvati@sri.utoronto.ca Department of Medical Imaging, University of Toronto, Toronto, ON, Canada In the current clinical model, men with positive digital Physical Sciences, Sunnybrook Research Institute, Toronto, ON, Canada rectal exam (DRE) and elevated prostate-specific anti- Full list of author information is available at the end of the article gen (PSA) require multicore random biopsies for risk © 2015 Khalvati et al. Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 2 of 14 stratification. However, there is an ongoing controversy Radiologists’ interpretations of MP-MRI have shown to about the role of prostate PSA as a screening test in achieve good prostate cancer detection rates, reaching prostate cancer. Two recent major randomized clinical tri- accuracies of 80 % in the peripheral zone of the prostate als [5, 6] have demonstrated that PSA screening contains a gland [28]. Similarly, several algorithms have been pro- significant risk of overdiagnosis for prostate cancer where posed for auto-detection of prostate cancer using MP- it is estimated that 50 % of screened men are diagnosed MRI setting [13–15, 29–31]. These algorithms usually with prostate cancer. This leads to painful needle biopsies compute a set of low-level features from the MP-MRI data and subsequent potential overtreatment [5–8]. Moreover, to construct feature vectors. Next, a supervised classifier it has become increasingly clear that carrying out prostate is trained using the computed feature vectors from the biopsy procedures escalates hospital admission rates due training cases and their associated ‘ground-truth’ labels to infectious complications, regularly resulting in discom- (e.g., labeled healthy or cancerous). Finally, the trained fort and possible sexual dysfunction while with the chance classifier is used to classify new cases. The reported values of the needle missing cancerous tissue [9–11]. Neverthe- for accuracy of cancerous versus healthy tissue classifica- less, PSA testing has proven to reduce prostate cancer tion ranges from 64 % to 89 %, depending on the feature mortality by 20–30 % at long-term follow-ups [10]. There- sets and training/test data. fore, the PSA testing remains an important biomarker in Different texture features have been used in the litera- diagnosing prostate cancers that are clinically significant. ture for automatic detection and classification of prostate The remaining challenge is how to improve the prostate cancer. Most of the reviewed methods utilized texture cancer diagnosis to reduce the overdiagnosis of clinically features that are based on one or more of the follow- insignificant cancers. ing methods: First-Order statistical method, second-order Automatic detection of prostate cancer as part of a clin- statistical methods or Co-Occurrence Matrices [32, 33], ical decision support system can potentially help radiolo- steerable Gabor filter [34], Gradient based features (e.g., gists in interpreting images more accurately. Specifically, Kirsch [35]), fractal based features [36], run length matri- multi-parametric MR imaging (MP-MRI), which com- ces [37], and discrete cosine transform (DCT) [38]. Dif- bines two or more of T2-weighted MRI (T2w), diffusion- ferent classifiers are used for classification of pixels in weighted imaging (DWI), dynamic contrast enhanced prostate as cancerous and healthy among which support imaging (DCE), and spectroscopy has been investigated as vector machine (SVM), neural networks, naive Bayesian, a promising approach for prostate cancer diagnosis and and random forests are most frequently used. construction of detection algorithms [12–16]. By taking The usefulness of analysis of these texture features in advantage of the unique quantitative information pro- prostate cancer detection has been demonstrated in a vided by each individual imaging technique, MP-MRI can variety of applications. Madabhushi et al. [29] presented exploit the different characteristics of prostate tissue to the utility of combining multiple features in detecting improve differentiation between cancerous and surround- high likelihoods of prostatic adenocarcinoma from high- ing tissue. For example, cancerous tissue in the prostate resolution ex-vivo MRI (i.e., following radical prostate- gland may exhibit a moderate drop in signal in T2w [17] ctomy). In this method, the following feature sets were (which characterizes differences in transverse (spin-spin) extracted from 3D voxels to train an ensemble of classi- relaxation time of tissue), restricted diffusion in DWI [17] fiers: First- and second order statistical method, steerable (which characterizes diffusion of water in tissue), earlier Gabor filters, Gradient based features, and discrete cosine onset time, higher peak, and shorter peak time in DCE transform (DCT). The algorithm was applied to 5 MR [18] (which characterizes the concentration of an injected prostates and while specificity was high (e.g., 98 %), the gadolinium contrast agent over time as it passes into the sensitivity was reported to be low (36 % - 42 %). extracellular extravascular space of the tissue). Moreover, Duda et al. [39] defined multi-image texture analysis studies have demonstrated the ability of MP-MRI to direct (MITA) to characterize prostatic tissues in MR images. In biopsy with MRI/Ultrasound fusion techniques [19] and this method, multiple MRI acquisitions of corresponding to predict Gleason score [16] and tumour volume [20]. slices are performed to form a database of image n- Thepulse sequencethathas shownthe most promiseis tuples. This database includes T1-weighted (DCE), T2w, DWI in the peripheral zone and the combination of T2w and DWI of 19 patients. The process of validating any and DWI in the transition zone [21, 22]. The apparent MITA consisted of contouring Regions of Interest (ROI) diffusion coefficient (ADC) map in particular has shown by a clinician. Once all the ROIs were contoured, the cor- the most promise as a biomarker [16, 23–27]. Although responding slices from each sequence were combined to DCE is considered as part of MP-MRI, T2w+DWI is the form n-tuple images, from which texture features were most common MP-MRI because it has the most diagnos- extracted and concatenated in a feature vector used to tic value and does not require invasive contrast agent as train the classifiers. In addition to features used in [29], DCE does. fractal-based and run length features were also used. It Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 3 of 14 was shown that the MP-MRI performed better compared the other hand, improving the separability of cancerous to pairs of MRI modalities. Although accuracies of up to and healthy tissues in the images would have a signifi- 99 % was reported, the evaluation was only performed on cant impact on the performance of cancer auto-detection one slice (middle slice) for each modality. Moreover, the algorithms, potentially reducing the dependency on the ROIs used for classification and measuring the accuracy feature extraction methods. was considerably large (400 to 2,400 pixels). In this paper, we propose new MP-MRI texture fea- In another study, Litjens et al. [30] introduced the use of ture models that in addition to T2w and conventional a cascaded classifier in order to characterize benign con- DWI images, incorporate computed high-b diffusion- founders such as atrophy, inflammation, benign prostatic weighted imaging (CHB-DWI) [14] and the recently pro- hyperplasia (BPH), and prostatic intra-epithilial neoplasia posed correlated diffusion imaging (CDI) [42]. Compared (PIN) as the sources of challenge to diagnose malignant to DWI images, CHB-DWI and CDI have both shown prostate cancer. In this paper, the authors presented the initial promise to improve visual separability of cancer- biology behind the benign confounders and bridged it ous and healthy tissues in prostate, which can lead to with MRI sequences. The pathology annotations were improved performance of the proposed MP-MRI tex- propagated to MR images by registering the whole-mount ture feature models for detecting prostate cancer. One slides with MR images. Different features were extracted aspect of the proposed MP-MRI texture feature mod- from different MR images; second-order statistical and els is to use non-invasive modalities aiming for higher GaborfeaturesfromT2w,multi-scaleblobnessfilterfrom usability in the clinical practice. Hence, we did not use ADC images, and curve fitting and pharmacokinetic fea- DCE images in our models. Moreover, while most can- tures from DCE images [31]. The maximum relevance, cer detection algorithms use combined b-value images in minimum redundancy (mRMR) feature selection tech- the form of apparent diffusion coefficient (ADC) map, nique [40] was used to determine best features for sep- our proposed texture feature models utilize the individ- arating of cancer from three non-cancer classes (BPH, ual b-value images of DWI to extract additional sets of inflammation, and atrophy). A cascaded classifier was features leading to improved accuracies. For each modal- used to gradually determine whether the sample is can- ity, the best feature subsets are selected based on different cerous. MRI data of 31 patients with 44 corresponding performance evaluation criteria (sensitivity or specificity). histological H&E stained slides were used to evaluate the These best feature subsets are then combined to construct detection algorithm; a maximum accuracy of 76.4 % was the comprehensive feature set from which the final best achieved. feature subset is selected to be used by the classifier. To Tiwari et al. [13] proposed a method that combines the authors’ best knowledge, the proposed comprehensive structural and metabolic imaging data for separating texture feature models are the first that utilize all of the benign versus cancerous and high Gleason score ver- above-mentioned MP-MRI modalities and combine them sus low Gleason grade regions in MP-MRI that includes using best feature subsets to construct an optimal texture T2w and magnetic resonance spectroscopy (MRS). Sim- feature model. ilar set of features used in [29] was used with a random The proposed MP-MRI texture feature models are the forest classifier for detection where the evaluation was first attempt in designing comprehensive quantitative fea- performed on 29 patient studies; accuracy of 86 % was ture sequences or radiomics as a high dimensional mine- achieved. able feature space that can be used as both detection and Ozer et al. [15] extracted second-order statistical and prognostic tools for prostate cancer [43]. The proposed DCT features from 19 patients’ MP-MRI (T2w, DWI, and radiomics-driven models in this paper have been used DCE) and used two classifiers (SVM and Relevance Vector for prostate cancer detection and they can be augmented Machine (RVM) [41]) to autodetect prostate cancer. The for prognostic of prostate cancer as well. Studies on lung best achieved sensitivity and specificity were 78 % and and head-and-neck cancer patients have confirmed the 79 %, respectively. Glaister et al. [14] studied computed prognostic power of radiomics features when it comes high b-value DWI for localization of prostate cancer and to patient outcome prediction for personalized medicine it was found that using ultra-high b-value (≥ 2000s/mm ) [44, 45]. However, the prognostic capability of radiomics improves the separability of cancerous and healthy tissues features has not been fully investigated for prostate can- significantly. cer and this is a novel approach for identifying prostate The underlying challenge in all of these auto-detection tumours phenotypes. algorithms is whether there is enough separability In a previous work [46], the preliminary results for the between the cancerous and healthy tissues in a given proposed approach was reported. This paper is signifi- image. This means if the separability is poor, even sophis- cantly different than the initial work as follows. First, the ticated feature extraction algorithms may not have a sig- previous work only used T2w, ADC, CHB-DWI and CDI nificant effect on the accuracy of cancer detection. On whereas the current approach also utilizes four b-value Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 4 of 14 images. Second, in [46], only 19 features were used com- Diffusion-weighted imaging (DWI) pared to 96 features used in this work. As it will be seen in DWI is a promising imaging modality in which the sen- the “Results” section, using more data (more images using sitivity of tissue to Brownian motion of water molecules. b-value images and more features) makes the texture fea- The signal intensity is measured by applying pairs of ture model more accurate, in terms of the cancer detection opposing magnetic field gradient pulses, also known as accuracy. Third, in the previous work, feature selection lobe gradients [49]. The radio-frequency is excited by was not used while here we use feature selection for each applying a 180 degree pulse on the phase of all the spins. modality and also for combination of different modalities. The first gradient lobe, in turn, introduces a signal diphase Feature selection allows to build a more optimal texture in all the spins proportional to the gradient lobe area. feature model leading to more accurate results. Finally, The spins, then, evolve freely, divided into static spins only five patients datasets were used in the previous work and spins that move with respect to their relative posi- whereas 20 patients datasets have been used in this paper tion. The same intensity and polarity of the first gradient (6,535 cancerous and healthy tissue samples versus 40,975 lobe is used again for a second gradient lobe, where all the samples) allowing for a better validation of the proposed static spins align to the 90 degree pulse and the moving texture feature models. spins never recovering the phase. The moving spins create higher diphase among the spins, acquiring less signal than that of the static spins. The diffusion-weighted signal, S is Methods and materials formulated as: We propose MP-MRI texture feature models for prostate −bD cancer detection which take advantage of abundance of S = S e (1) data from different MR modalities to compute features used by the classifier. The goal is to combine features from where S is the signal intensity without the diffusion each imaging modality that best separates cancerous pix- weighting. The signal loss due to spins diphase, accord- els from healthy ones. In the following, we present the ing to Stejskal-Tanner sequence, can be controlled by b, imaging methods used in the proposed model, the feature which consists of amplitude and duration of the diffu- sets, and the proposed texture feature models. In addi- sion pulses, gradient intensity and the time between the tion, details about the image acquisition protocols and the two pulses as well as the gyromagnetic ratio, and D repre- performance measures are presented. sents the strength of the diffusion. The diffusion-weighted image (S) is usually generated with different b values which can be used to estimate apparent diffusion coef- Imaging methods ficient map (ADC) using the least-squares or maximum The main criteria for choosing imaging modalities used likelihood strategies [49]. The cancerous tissue in ADC is in the proposed texture feature models are twofold. First, usually represented by a darker intensity compared to the images that are part of well-known radiology reporting surrounding tissue. system. Second, they are acquired non-invasively, with no need for contrast agents, and can be collected in a single Computed high-b diffusion-weighted imaging (CHB-DWI) imaging session. Recently, a structured 5-scale reporting Previous research has shown that high b-value DWI system, PI-RADS, was proposed for consistent prostate 2 images (e.g., b-values greater than 1,000 s/mm ) allow MP-MRI reading [47] with subsequent studies confirm- for increased delineation between tumours and healthy ing its effectiveness with respect to biopsy results [48]. tissues [14, 50] which makes the prostate cancer detec- PI-RADS consisst of T2w, DWI (ADC) as well as DCE tion more robust. Nevertheless, due to hardware limita- images. Instead of using DCE which requires contrast tions, most MRI machines in practice do not produce agent, the proposed texture feature models use additional 2 DWI with b-values higher than 1,500 s/mm for prostate information available by DWI images which includes imaging. CHB-DWI is an alternative approach to obtain computed high b-value image, individual b-value images, high-b DWI in which a computational model is used to and correlated diffusion images. This subsection summa- reconstruct DWI at high b-values using low b-value DWI rizes the imaging methods used in the proposed MP-MRI acquisitions [14, 51]. For our experiments, we constructed feature models. 2 CHB-DWI with b-value at 2000s/mm using a Bayesian model with the same least squares estimation technique T2-weighted imaging (T2w) 2 used for ADC, extrapolating to the b-value of 2000s/mm . T2w is a MR imaging modality in which the sensitivity of tissue is characterized by measuring the relaxation time Correlated diffusion imaging (CDI) (spin-spin) of the applied magnetic field. The T2w image CDI [42] is a new diffusion magnetic resonance imaging of prostate usually shows a small reduction in signal in the modality, which takes advantage of the joint correlation in cancerous tissue [17]. signal attenuation across multiple gradient pulse strengths Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 5 of 14 Table 1 Summary of textural features used in the feature model and timings to not only reduce the dependency on the way diffusion gradient pulses are applied, but also improve Feature class Feature delineation between cancerous and healthy tissue. The First-order statistical features Mean, Standard deviation effectiveness of the delineation process depends on the Skewness models of the different types of tissue, since tumorous tis- Kurtosis sue has been empirically demonstrated to generate higher Energy, Contrast greyscale intensities at higher b-values. As such, in con- structing CDI, these properties are exploited where the Correlation, Variance utilized b-values are adjusted for a given application. The Inverse difference moment local correlation of signal attenuation across all b-values Sum average, Sum variance within a local sub-volume is calculated to better represent Sum entropy, Entropy the overall characterization of the water diffusion prop- Second-order statistical Difference variance erties of the tissue. The CDI signal is obtained via signal features (Haralick) Difference entropy mixing as follows [42]: Information measure of correlation n Homogeneity, Autocorrelation CDI(x) = ... S (x)...S (x)P(S (x), ... , S (x)| 0 n 0 n Dissimilarity, Cluster shade Cluster prominence V (x)) × dS (x)...dS (x) (2) 0 n Maximum probability Gabor filters 3 scales and 4 orientations where x denotes spatial location, b represents b values, S denotes the acquired signal, P denotes the conditional Kirsch filters 8 directions joint probability density function, and V (x) denotes the local subvolume around x. from different sets of images to construct radiomics fea- Feature extraction tures; a high-dimensional feature space that can be mined In order to separate the cancerous tissue from the healthy for different purposes such as detection or prognosis of one, a set of features is calculated on a given MR imag- cancer. Similar to conventional MP-MRI, the proposed ing modality (i.e., T2w, DWI, CHB-DWI, CDI, and indi- feature models include T2w and ADC modalities. They vidual b-value images). We incorporate four well-known also incorporate CHB-DWI, which has been shown to classes of texture features used in different studies to sep- increase separation between healthy and cancerous tis- arate cancerous and healthy tissues in prostate. These sue. As discussed in Section “Correlated diffusion imaging features include first- and second-order statistical features (CDI)”, as a new diffusion magnetic resonance modality, (Haralick [32, 33]), steerable Gabor filter features [34], and CDI has shown promise in separating healthy tissue from Kirsch filter features [35]. The first-order statistical fea- cancerous one. Although ADC incorporates all b-value tures include mean and standard deviation of grey-level images implicitly, individual b-value images may contain intensity, skewness, and kurtosis. Second-order statisti- information to help further distinguish healthy tissues cal features such as entropy and contrast are extracted from cancerous tissues. Therefore, we also incorporate from the gray-level co-occurrence matrix (GLCM) in four four b-value images into our proposed texture feature directions: 0 °, 45 °, 90 °, and 135 °. These texture features models. The following lists all the imaging modalities used include 18 features in each direction generating a total of by the proposed texture feature models for prostate cancer 72 features. Gabor features includes 12 features from three detection: scales and four orientations and Kirsch features include the maximum gradient in eight directions. As a result, I =T2w the proposed MP-MRI texture feature models consist of a I =ADC total of 96 features for each imaging modality: four from I = CHB-DWI: b-value at 2000s/mm first-order and 72 from second-order statistical features, I =CDI eight from Kirsch, and 12 from Gabor filters. Table 1 sum- I = b : b-value at 0s/mm 5 1 marizes all features used in the proposed texture feature I = b : b-value at 100s/mm 6 2 models. I = b : b-value at 400s/mm 7 3 I = b : b-value at 1000s/mm 8 4 Texture feature model Figure 1 shows the block diagram of the proposed texture For each modality, I , from the list above, the features feature models. The goal is to incorporate information described in Table 1 are calculated for a local window (e.g., Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 6 of 14 Fig. 1 Block diagram of the proposed texture feature models 3 × 3 pixels) sliding on the prostate gland. Each window on performance criteria used, the texture feature models is labeled either a tumour or non-tumour voxel. For each produce different results. imaging modality I , this gives a feature vector F . Once the best feature subsets for each imaging modality i i For each voxel in each image, the feature extraction was determined, the next step is to combine them to build function produces 96 features. A feature selection algo- different texture feature models (TFM) as follows: rithm determines a subset of features that contribute the TFM = T2w+ADC most to the separability of classes (e.g., cancerous vs. non- TFM = T2w+ADC+CHB-HBV cancerous tissues). This allows us to use the best features TFM = T2+CDI for each imaging modality when building the final texture TFM = T2w+ADC+CDI feature models. The feature selection algorithms usually TFM = T2w+ADC+HBV+CDI require the number of best features to be selected. For a TFM = T2w+ADC+HBV+CDI+b +b +b +b 6 1 2 3 4 given imaging modality I , to determine the optimal num- ber of features m , we perform an exhaustive search over The feature selection method is applied to each texture thefeature spacetoevaluatethe performanceofany num- feature model to build the final models. At this stage, the ber of features. This allows us to select m features as two performance criteria (sensitivity and specificity) are the feature vector F that produces the best results for a used to select the final best feature subsets for each tex- given imaging modality I . ture feature model. Algorithm 1 summarizes the texture To evaluate the performance of a given number of fea- feature model construction steps. tures, the accuracy or area under curve (AUC) for receiver operating characteristic (ROC) curve of the classification is usually used. Cancer cells in prostate usually constitute Algorithm 1 Texture Feature Model Construction a small fraction of the entire prostate gland (i.e., around 1: For each imaging modality I in training the set T = 1 %). This means that an accuracy of an algorithm may {I , I , ... , I }, apply feature extraction function: F = 1 2 n i be very high (e.g., 0.90) while it is unable to correctly Features(I ). locate the cancerous cells (i.e., low sensitivity). On the 2: For the feature set of each imaging modality in train- other hand, depending on the clinical procedures, differ- i ing set F , apply feature selection function: F = ent levels of sensitivity or specificity may be required. For F_Selection(F ) where m is thenumberofthe best i i example, for cancer screening programs, high sensitivity selected features for imaging modality I . (e.g., 0.90) is required where a moderate specificity (e.g., 3: Construct different combinations of selected features 0.60) is deemed to be adequate. On the other hand, for m 1 from different imaging modalities: F = F ∪ j i a procedure such as radical prostactomy, a high speci- m m 2 in F ... ∪ F where m = m + m + ... + m . 1 2 n i i 2 n ficity (e.g., 0.99) with moderate sensitivity (e.g., 0.60) is 4: Apply feature selection function to the constructed necessary to avoid unnecessary surgery. As a result, when m 0 m feature set F : F = F_Selection(F ) where m is j j j choosing the best feature subset, it is important to con- thenumberofthe finalbestselectedfeatures. sider different clinical scenarios by considering different 5: Apply classification to F . performance evaluation criteria for feature selection. To determine the best feature subsets, we examine two scenarios where in each scenario, it is assumed that either sensitivity or specificity has a higher priority in the per- For feature extraction function, we used the maximum formance evaluation of the proposed texture feature mod- relevance, minimum redundancy (mRMR) technique [40], els. As it will be seen in the results section, depending which is based on maximum relevance and minimum Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 7 of 14 redundancy of features. In this method, the feature subset ROI-based. In pixel-based approach [53], small neighbor- F is selected to satisfy the following criteria: hoods of pixels (e.g., 3 × 3) are considered to distinguish cancerous tissues from healthy ones. In other words, accu- racy determines the percentages of these neighborhood max D(F , c), D = MI(f ; c) (3) |F | m that were correctly labeled as cancerous or healthy. ROI- f ∈F based approach [29, 39, 54] is similar to pixel-based with the difference that it uses larger neighborhoods of pixels min R(F , c), R = MI(f ; f ) (4) i j m 2 |F | (e.g., 50 × 50) for calculating accuracy measures. f ,f ∈F i j In evaluating the performance of the proposed tex- where F is the best feature subset that we would like ture feature models in this paper, we use the pixel-based to find, c is the target class, f is a feature and MI is the approach so that the accuracy measurements are calcu- mutual information function. D and R are the relevance lated more precisely. As ground-truth, all MP-MR images and redundancy of features, respectively. Maximum rele- were reviewed and marked as healthy and cancerous vance guarantees that the selected features have the high- tissue by a radiologist with 18 and 13 years of experi- est shared information with the target class and minimum ence interpreting body and prostate MRI, respectively. In redundancy ensures that the redundant features are elim- addition, for cases with cancer, the MP-MRI images and inated. For the classifier, we used the SVM implemented expert annotations were compared to the corresponding in [52]. histopathology data, obtained through radical prostate- The proposed radiomics-driven cancer detection mod- ctomy with Gleason score seven and above, as ground- els combine a plethora of data from different imaging truth to confirm the accuracy of the MP-MRI markings. modalities of MP-MRI to construct comprehensive tex- ture feature models which can be used for both detection Results and prognosis purposes in prostate cancer. Figure 2 shows sensitivity and specificity for all 8 MP-MRI modalities using different number of best features (e.g., 10 Image data features to 96 features). For each modality, 40,975 sam- MRIdataof20patients(17 with cancer andthree ples (40,369 healthy and 606 cancerous samples confirmed without cancer) were acquired using a Philips Achieva by the radiologist) was used for the leave-one-patient-out 3.0T machine at Sunnybrook Health Sciences Centre, cross-validation. Toronto, Ontario, Canada. All data was obtained retro- Tables 3 and 4 show the quantitative results for dif- spectively under the local institutional research ethics ferent modalities and combinations of modalities. Using board (Research Ethics Board of Sunnybrook Health Sci- sensitivity as feature selection criteria (Table 3), the sen- ences Centre). For each patient, the following MP-MRI sitivity of the texture feature models reaches 0.86 using modalities were obtained (Table 2): T2w, DWI, and CDI. CDI alone. It is interesting to observe that CDI also out- The patients’ age ranged from 53 to 83. Table 2 sum- 1 performs the conventional MP-MRI (i.e., TFM )and marizes the information about the 20 patients’ datasets combination of conventional MP-MRI and CHB-DWI used in this research, which includes displayed field of (i.e., TFM ) (0.86 vs. 0.77 and 0.86 vs 0.69, respectively). view (DFOV), resolution, echo time (TE), and repeti- Although CDI alone gives the best results for sensitivity tion time (TR). Images were processed in the ProCanVAS (0.86), the full feature sets model (i.e., TFM ) produces the (Prostate Cancer Visual Analysis System) platform devel- best results when considering specificity, accuracy, and oped at Sunnybrook Research Institute, Toronto, ON, AUC as well (0.82, 0.82, and 0.86, respectively). Compar- Canada. Each modality (e.g., CDI) provided 40,975 sam- ing TFM to all other models in Table 3, at least 2 metrics ples used for the leave-one-patient-out cross-validation of out of 4 are significantly different than each of other mod- the algorithms. els. For example, comparing TFM to TFME ,the P values 6 5 for specificity and accuracy via Wilcoxon signed-rank test Evaluation metrics are 0.006 and 0.01, respectively. To evaluate the performance of cancer detection algo- Table 4 shows the performance results for using speci- rithms, two approaches may be used: pixel-based and ficity as performance evaluation criteria for feature selec- tion. It is observed that compared to the previous Table 2 Description of the prostate T2w, DWI, and CDI images approach (Table 3), the full feature sets model (TFM ) 2 3 Modality DFOV (cm)Resolution(mm ) TE (ms) TR (ms) improves the specificity by 0.06 (0.88). This was expected since the performance evaluation criteria used for fea- T2w 22 × 22 0.49 × 0.49 × 3 110 4,687 ture selection affects the final results. Thus, as discussed DWI 20 × 20 1.56 × 1.56 × 3 61 6,178 in Section “Texture feature model”, depending on the CDI 20 × 20 1.56 × 1.56 × 3 61 6,178 clinical scenario, one can choose different performance Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 8 of 14 Fig. 2 Performance results for different modalities (T2w, ADC, CHB-DWI, CDI, and 4 DWI images at different b values) across all features Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 9 of 14 Table 3 Evaluation results for prostate cancer detection: Feature selection based on Sensitivity (Results are shown with 95 % confidence interval) Imaging Number of Sensitivity Specificity Accuracy AUC modality features T2w 96 0.71 [0.54 0.89] 0.44 [0.39 0.49] 0.45 [0.40 0.50] 0.58 [0.48 0.68] CHB-DWI 90 0.73 [0.58 0.88] 0.78 [0.71 0.85] 0.77 [0.71 0.84] 0.79 [0.73 0.85] ADC 20 0.76 [0.64 0.88] 0.59 [0.51 0.67] 0.60 [0.52 0.67] 0.68 [0.63 0.74] CDI 96 0.86 [0.76 0.97] 0.80 [0.75 0.85] 0.79 [0.74 0.84] 0.85 [0.81 0.90] TFM = T2w+ADC 20 0.77 [0.64 0.91] 0.57 [0.49 0.65] 0.59 [0.51 0.66] 0.68 [0.62 0.74] TFM =T2w+ADC+CHB-DWI 208 0.69 [0.54 0.84] 0.79 [0.73 0.84] 0.78 [0.73 0.84] 0.78 [0.72 0.85] TFM =T2w+CDI 196 0.85 [0.75 0.96] 0.81 [0.76 0.86] 0.80 [0.76 0.85] 0.85 [0.81 0.90] TFM =T2w+ADC+CDI 216 0.86 [0.76 0.96] 0.81 [0.76 0.86] 0.80 [0.76 0.85] 0.85 [0.81 0.90] TFM =T2w+ADC 300 0.86 [0.75 0.96] 0.81 [0.77 0.86] 0.81 [0.77 0.85] 0.86 [0.83 0.90] +CHB-DWI+CDI TFM = T2w+ADC +CHB-DWI+CDI 416 0.86 [0.75 0.97] 0.82 [0.78 0.87] 0.82 [0.78 0.86] 0.86 [0.81 0.91] +b +b +b +b 1 2 3 4 evaluation criteria to better suit the clinical procedure sensitivity (0.90 vs. 0.87). Figure 4 shows the ROC curves requirements. Comparing TFM to all other models in for all six models as well as individual imaging modali- Table 4 (except for TFM ), at least 2 metrics out of four ties discussed in Section “Texture feature model”. It is seen are significantly different than each of other models. For that the combination of all imaging modalities, TFM , example, comparing TFM to TFM ,the P values for gives the best results in terms of AUC (0.90). This result 6 5 specificity and accuracy via Wilcoxon signed-rank test is significantly different with respect to any other imaging are 0.01. Comparing TFM to TFM , the two models are modality or texture feature model where P < 0.009. 6 3 significantly different with respect to AUC (P = 0.01). Table 5 shows the optimal results with the target of max- Tables 3 and 4 show the result when the goal was to imizing sensitivity, specificity, or AUC. As it can be seen, maximize sensitivity (Table 3) or specificity (Table 4). choosing a target yields the best result for the selected tar- Figure 3 shows the combinations of all eight imaging get. Setting AUC as the target maximizes the AUC (0.90) modalities (TFM )withbestfeature subsetsbased on sen- and at the same time generates more balanced results with sitivity and specificity with the objective of maximizing for respect to sensitivity and specificity (0.84 and 0.86). Using AUC. It can be seen that using specificity as performance sensitivity as the performance evaluation criteria maxi- evaluation criteria gives a higher best AUC compared to mizes the result for sensitivity (0.86). Using specificity as Table 4 Evaluation results for prostate cancer detection: Feature selection based on specificity (results are shown with 95 % confidence interval) Imaging Number of Sensitivity Specificity Accuracy AUC modality features T2w 10 0.66 [0.50 0.81] 0.47 [0.42 0.53] 0.48 [0.43 0.53] 0.57 [0.48 0.66] CHB-DWI 10 0.69 [0.52 0.86] 0.82 [0.75 0.88] 0.81 [0.75 0.87] 0.76 [0.68 0.84] ADC 96 0.73 [0.60 0.85] 0.62 [0.55 0.70] 0.63 [0.56 0.71] 0.70. [0.64 0.76] CDI 10 0.82 [0.69 0.94] 0.85 [0.80 0.89] 0.84 [0.80 0.88] 0.84 [0.78 0.89] TFM = T2w+ADC 110 0.72 [0.59 0.86] 0.63 [0.55 0.70] 0.64 [0.56 0.71] 0.69 [0.63 0.75] TFM =T2w+ADC 40 0.66 [0.50 0.82] 0.77 [0.71 0.83] 0.77 [0.71 0.82] 0.73 [0.65 0.81] +CHB-DWI 20 0.78 [0.65 0.91] 0.86 [0.82 0.90] 0.86 [0.82 0.89] 0.84 [0.78 0.90] TFM =T2w+CDI 40 0.77 [0.63 0.90] 0.86 [0.82 0.90] 0.85 [0.81 0.89] 0.84 [0.79 0.89] TFM =T2w+ADC+CDI TFM =T2w+ADC 50 0.78 [0.64 0.91] 0.86 [0.82 0.90] 0.85 [0.82 0.89] 0.84 [0.78 0.90] +CHB-DWI+CDI TFM =T2w+ADC +CHB-DWI+CDI 130 0.80 [0.69 0.91] 0.88 [0.85 0.92] 0.88 [0.84 0.91] 0.88 [0.83 0.93] +b +b +b +b 1 2 3 4 Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 10 of 14 the performance evaluation criteria maximizes the result for specificity (0.88) and AUC (0.90), depending on the selected target. Figure 5 shows an example for all four modalities which include T2w, ADC, CHB-DWI, and CDI. As it can be seen, CDI (Fig. 5d) is the only modality that clearly shows a bright nodule where a tumour is located (confirmed by histopathology data - Fig. 6). Discussion Automated prostate cancer detection has been investi- gated by different research groups in the field. The under- lying building blocks of these algorithms consist of feature extraction and classification applied to local windows of pixels in the images. Most algorithms in the literature usually incorporate few imaging sequences into their pro- posed feature models. For example, the work presented in Fig. 4 ROC for different texture feature models [39] used three sequences (T1w, T2w, and DCE) to build the texture feature model. In contrast, in this paper, we have taken one step further by incorporating information from all available MR imaging data which includes T2w, One important aspect in the clinical workflow for ADC, and different b-value images of DWI (i.e., b-values prostate cancer detection is the targeted clinical proce- at 0, 100, 400, and 1000s/mm ). Moreover, we incorpo- dure. For example, cancer screening programs impose rated computed high-b DWI (CHB-DWI) [14] as well as different performance requirements compared to proce- correlated diffusion imaging (CDI) [42] into our model. dures such as radical prostatectomy. We designed the Adding these two extra imaging modalities enriched our proposed texture feature model accounting for such texture feature models in terms of the diversity of data requirements where the performance of the model can where 6 different models were developed and evalu- be optimized for sensitivity, specificity, or the area under ated (Sections “Texture feature model” and “Results”). the ROC curve. For example, to use the proposed tex- As a quantitative radiomics approach for prostate can- ture feature model for cancer screening, sensitivity can be cer detection, we used a comprehensive texture feature used as the performance evaluation criteria to steer the model which incorporated eight different imaging modal- feature selection process which would lead to best result ities where each modality contributed with its best feature for sensitivity (0.86) with reasonable results for specificity subset to the ultimate texture feature model in which all (0.82). For cases where higher specificity is required, one modalities were combined. can use specificity as the performance evaluation criteria to optimize the results for specificity (0.88) with accept- able sensitivity (0.80). Our experiments showed that using specificity as the performance evaluation criteria can also maximize the results for AUC (0.90) which leads to a bal- anced results for sensitivity and specificity; 0.84 and 0.86, respectively. The fact that the proposed model is flexible in terms of optimizing the results for the procedure it is used for makes it more practical. This is another novel aspect of the proposed model in this paper which to the authors’ best knowledge has not been fully explored in the literature. The limitations of our research include a relatively lim- ited number of datasets (20 patients) and targeting only Gleason score of seven and above. Evaluating the pro- posed model using a larger dataset and considering lower Gleason scores (e.g., six) will add more confidence to the reliability of the model which will be done as future Fig. 3 AUC based on using sensitivity and specificity as performance work. Other limitation is that the proposed model was not evaluation criteria assessed by clinicians to investigate whether it improves Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 11 of 14 Table 5 Evaluation results for prostate cancer detection: Feature selection based on Sensitivity and Specificity (Results are shown with 95 % confidence interval) Target Performance evaluation criteria Sensitivity Specificity AUC Sensitivity Sensitivity 0.86 [0.75 0.97] 0.82 [0.78 0.87] 0.86 [0.81 0.91] Specificity Specificity 0.80 [0.69 0.91] 0.88 [0.85 0.92] 0.88 [0.83 0.93] AUC Specificity 0.84 [0.76 0.91] 0.86 [0.82 0.91] 0.90 [0.88 0.93] the clinical readings by radiologists. Similar to the work Section “Results”, boosted the results significantly. We reported in [55], clinical assessment of the proposed auto- have developed an enhanced version of CDI, called dual- detection model will be performed to evaluate its effect stage correlated diffusion imaging (D-CDI) which has on the clinicians’ performance. Finally, given the fact that shown promise in enhancing separability of cancerous and cancerous pixels are a small fraction of the entire prostate healthy tissue in prostate MRI compared to CDI [56]. gland, it is possible that the reported specificity results are As future work, we will incorporate D-CDI to the pro- an overestimation. A larger and more diversified dataset posed texture feature model to investigate the effect on will help to investigate this more thoroughly. performance. We will also investigate developing a hybrid Our proposed texture feature model incorporated CDI morphological-textural feature model for prostate cancer as one of the imaging modalities which as shown in where in addition to texture analysis, the morphological Fig. 5 a T2w does not clearly show a tumour although there is mild signal alteration in the left peripheral zone (arrow). b ADC does not clearly show a tumour (arrow). c CHB-DWI of 2000 s/mm shows no tumour (arrow). d CDI clearly shows a bright nodule (arrow) corresponding to tumour Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 12 of 14 Fig. 6 Corresponding axial hematoxylin and eosin stained tissue showing a Gleason 7 (4+3) tumor circled in red corresponding to the lesion identified best on the CDI images in Fig. 5-d characteristics (e.g., shape) of candidate regions are taken optimal texture feature models. A SVM classifier was into account to detect cancer. A preliminary work on trained via leave-one-patient-out setting to classify the morphological feature model has been presented in [57] new cases. The proposed MP-MRI texture feature models upon which we will extend and build the hybrid model. showed promise in accurate detection of prostate cancer. The b-value images of DWI are usually distorted due to Endnote patient movement during the image acquisition which The conventional MP-MRI refers to the combination may reduce cancer separability. We have presented pre- of T2w and DWI, which is represented as ADC. Thus, liminary results for co-registering the b-value images to throughout the paper, T2w+DWI and T2w+ADC are compensate for patient movement [58]. As future work, used interchangeably. we will incorporate this co-registration algorithm into our proposed texture feature models to investigate the effect Competing interests on the accuracy of cancer detection. Finally, normalized The authors have declared that no competing interests exist. entropy has been shown to be a strong predictor of patient survival rate for lung and renal cell cancers in CT images Authors’ contributions Conceived and designed the methodology: FKh, AW, MAH. Performed the [59]. The future work will also involve investigating the experiments: FKh, AW. Analyzed the data: FKh, AW, MAH. Wrote the paper: efficacy of the proposed texture feature model in this FKh, AW, MAH. All authors read and approved the final manuscript. paper for normalized MP-MRI entropy characterization Acknowledgements of prostate cancer. This research has been supported by the Ontario Institute of Cancer Research (OICR), Canada Research Chairs programs, Natural Sciences and Engineering Conclusion Research Council of Canada (NSERC), and the Ministry of Research and Innovation of Ontario. The authors would like to thank Prof. Rajiv Chopra, In this paper, we introduced new multi-parametric MRI Department of Radiology, UT Southwestern Medical Center, for providing texture feature models for prostate cancer detection. Our histology images. new MP-MRI texture feature models add two new imag- Author details ing modalities, computed high-b DWI and correlated Department of Medical Imaging, University of Toronto, Toronto, ON, Canada. diffusion imaging, to the most commonly used MP-MRI, Physical Sciences, Sunnybrook Research Institute, Toronto, ON, Canada. T2w+ADC. As a quantitative radiomics approach for Department of Systems Design Engineering, University of Waterloo, Waterloo, ON, Canada. automatic detection of prostate cancer, a comprehensive set of texture features were calculated for the conventional Received: 2 December 2014 Accepted: 9 July 2015 MP-MRI and new MP-MRI texture feature models. A fea- ture selection method was used to select the optimal fea- References tures for each modality. Two different performance eval- 1. Canadian Cancer Society. Canadian Cancer Statistics. 2014. http://www. uation criteria were used for feature selection to reflect cancer.ca/~/media/cancer.ca/CW/cancer%20information/cancer different clinical workflows in prostate cancer diagnosis. %20101/Canadian%20cancer%20statistics/Canadian-Cancer-Statistics- 2014--EN.pdf. Thebestfeature subsetswerethencombinedtoconstruct Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 13 of 14 2. American Cancer Society. Cancer Facts and Figures. 2014. http://www. 22. Oto A, Kayhan A, Jiang Y, Tretiakova M, Yang C, Antic T, et al. Prostate cancer.org/research/cancerfactsstatistics/cancerfactsfigures2014/. cancer: differentiation of central gland cancer from benign prostatic 3. Jemal A, Bray F, Center MM, Ferlay J, Ward E, Forman D. Global cancer hyperplasia by using diffusion-weighted and dynamic contrast-enhanced statistics. CA: A Cancer J Clin. 2011;61(2):69–90. MR imaging. Radiology. 2010;257:715–23. 4. Canadian Cancer Society. Canadian Cancer Statistics. 2011. http://www. 23. Shigemura K, Yamanaka N, Yamashita M. Can diffusion-weighted cancer.ca/~/media/cancer.ca/CW/cancer%20information/cancer magnetic resonance imaging predict a high Gleason score of prostate %20101/Canadian%20cancer%20statistics/Canadian-Cancer-Statistics- cancer? Korean J Urol. 2013;54:234–38. 2011-EN.pdf. 24. Caivano R, Rabasco P, Lotumolo A, Cirillo P, D’Antuono F, Zandolino A, 5. Andriole GL, Crawford D, Grubb RL, Buys SS, Chia D, Church TR, et al. et al. Comparison between Gleason score and apparent diffusion PLCO Project Team. Mortality results from a randomized prostate-cancer coefficient obtained from diffusion-weighted imaging of prostate cancer screening trial. N Engl J Med. 2009;360:1310–9. patients. Cancer Invest. 2013;31:625–9. 6. Schroder FH, Hugosson J, Roobol MJ, Tammela TL, Ciatto S, Nelen V, 25. Bae H, Yoshida S, Matsuoka Y, Nakajima H, Ito E, Tanaka H, et al. Apparent et al. ERSPC Investigators. Screening and prostate-cancer mortality in a diffusion coefficient value as a biomarker reflecting morphological and randomized European study. N Engl J Med. 2009;360:1320–8. biological features of prostate cancer. Int Urol Nephrol. 2014;46(3):555–61. 7. Greene KL, Albertsen PC, Babaian RJ, Albertsen PC, Carter HB, Gann PH, 26. Nagarajan R, Margolis D, Raman S, Sheng K, King C, Reiter R, et al. et al. Prostate Specific Antigen Best Practice Statement 2009 Update. J Correlation of Gleason scores with diffusion-weighted imaging findings Urol. 2009;182(5):2232–41. of prostate cancer. Adv Urol. 2012;2012:374805. 8. Vickers AJ, Sjoberg DD, Ulmert D, Vertosick E, Roobol MJ, Thompson I, 27. Bittencourt LK, Barentsz JO, de Miranda LC, Gasparetto EL. Prostate MRI: et al. Empirical estimates of prostate cancer overdiagnosis by age and diffusion-weighted imaging at 1.5T correlates better with prostatectomy prostate-specific antigen. BMC Medicine. 2014;12:26. Gleason Grades than TRUS-guided biopsies in peripheral zone tumours. 9. Loeb S, Vellekoop A, Ahmed HU, Catto J, Emberton M, Nam R, et al. Eur Radiol. 2012;22:468–75. Systematic review of complications of prostate biopsy. Eur Urol. 28. Langer DL, van der Kwast TH, Evans AJ, Sun L, Yaffe MJ, Trachtenberg J, 2013;64(6):876–92. et al. Intermixed normal tissue within prostate cancer: effect on MR 10. Schroder FH, Hugosson J, Roobol MJ, Tammela TLJ, Ciatto S, Nelen V, imaging measurements of apparent diffusion coefficient and T2–sparse et al. ERSPC Investigators. “Prostate-cancer mortality at 11 years of versus dense cancers. Radiology. 2008;249:900–8. follow-up. N Engl J Med. 2012;366(11):981–90. 29. Madabhushi A, Feldman MD, Metaxas DN, Tomaszeweski J, Chute D. 11. Nam RK, Saskin R, Lee Y, Liu Y, Law C, Klotz LH, et al. Increasing hospital Automated detection of prostatic adenocarcinoma from high-resolution admission rates for urological complications after transrectal ultrasound ex vivo MRI. IEEE Trans Med Img. 2005;24(12):1611–25. guided prostate biopsy. Division of Urology, Sunnybrook Research 30. Litjensa GJS, Elliottb R, Shihc N, Feldmanc M, Barentsza JO, Institute. 2013. doi: 10.1016/j.juro.2009.11.043. Hulsbergen - van de Kaaa CA, et al. Distinguishing prostate cancer from 12. Arumainayagam N, Ahmed HU, Moore CM, Freeman A, Allen C, benign confounders via a cascaded classifier on multi-parametric MRI. Sohaib SA, et al. Multiparametric MR Imaging for Detection of Clinically Medical Imaging. 2014. doi:10.1117/12.2043751. Significant Prostate Cancer: A Validation Cohort Study with Transperineal 31. Vos PC, Barentsz JO, Karssemeijer N, Huisman HJ. Automatic Template Prostate Mapping as the Reference Standard. Radiology. computer-aided detection of prostate cancer based on multiparametric 2013;268(3):761–769. magnetic resonance image analysis. Phys Med Biol. 2012;57:1527–42. 13. Tiwari P, Kurhanewicz J, Madabhushi A. Multi-kernel graph embedding 32. Haralick RM, Shanmugan K, Dinstein I. Textural features for image for detection, gleason grading of prostate cancer via MRI-mrs. Med Image classification. IEEE Trans Syst Man Cybern. 1973;SMC-3:610–21. Anal. 2013;17:219–35. 33. Soh L, Tsatsoulis C. Texture Analysis of SAR Sea Ice Imagery Using Gray 14. Glaister J, Cameron A, Wong A, Haider MA. Quantitative investigative Level Co-Occurrence Matrices. IEEE Trans Geosci Remote Sensing. analysis of tumour separability in the prostate gland using ultra-high 1999;37(2):. doi:10.1109/36.752194. b-value computed diffusion imaging. In: EMBC’2012; 2012. p. 420–423. 34. Gabor D. Theory of communication. J Inst Electr Eng. 1946;93:429–41. doi:10.1109/EMBC.2012.6345957. 35. Kirsch RA. Computer determination of the constituent structure of 15. Ozer S, Haider MA, Langer DL, van der Kwast TH, Evans AJ, Wernick MN, biological images. Comput Biomed Res. 1971;4:315–28. et al. Prostate cancer localization with multispectral MRI based on 36. Chen C, Daponte JS, Fox MD. Fractal feature analysis and classification in relevance vector machines. In: 2009 IEEE International Symposium on medical imaging. IEEE Trans Med Imag. 1989;8(2):133–42. Biomedical Imaging From Nano to Macro; 2009. p. 73–76. 37. Galloway MM. Texture analysis using gray level run lengths. Comp Graph doi:10.1109/ISBI.2009.5192986. Im Proc. 1975;4(2):172–9. 16. Turkbey B, Mani H, Shah V, Rastinehad AR, Bernardo M, Pohida T, et al. 38. Khayam SA. The Discrete, Cosine Transform (DCT): Theory and Multiparametric 3T prostate magnetic resonance imaging to detect Application: Department of Electrical & Computer Engineering, Michigan cancer: histopathological correlation using prostatectomy specimens State University; 2003. processed in customized magnetic resonance imaging based molds. J 39. Duda D, Kretowski M, Mathieu R, Crevoisier RD, Bezy-Wendling J. Urol. 2011;186:1818–24. Multi-Image Texture Analysis in Classification of Prostatic Tissues from 17. Haider MA, van der Kwast TH, Tanguay J, Evans AJ, Hashmi A, MRI. Preliminary Results. Information Technologies in Biomedicine, 139 Lockwood G, et al. Combined T2-weighted and diffusion-weighted MRI Volume 3, Advances in Intelligent Systems and Computing 283, 2014. for localization of prostate cancer. AJR Am J Roentgenol. 2007;189(2): doi:10.1007/978-3-319-06593-9_13. 323–8. 40. Peng H, Long F, Ding C. Feature selection based on mutual information 18. Langer DL, van der Kwast TH, Evans AJ, Plotkin A, Trachtenberg J, criteria of maxdependency, max-relevance, and min-redundancy. IEEE Wilson BC, et al. Prostate Tissue Composition and MR Measurements: Trans Pattern Anal Mach Intell. 2005;27:1226–38. Investigating the Relationships between ADC, T2, K-trans, v-e, and 41. Tipping ME. Sparse Bayesian learning and the relevance vector machine. Corresponding Histologic Features. Radiology. 2010;255(2):485–94. J Mach Learn Res. 2001;1:211–244. 19. Vourganti S, Rastinehad A, Yerram NK, Nix J, Volkin D, Hoang A, et al. 42. Wong A, Glaister J, Cameron A, Haider M. Correlated Diffusion Imaging. Multiparametric magnetic resonance imaging and ultrasound fusion BMC Medical Imaging. 201313–26. doi:10.1186/1471-2342-13-26. biopsy detect prostate cancer in patients with prior negative transrectal 43. Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RG, ultrasound biopsies. J Urol. 2012;188:2152–7. Granton P, et al. Radiomics: extracting more information from medical 20. Turkbey B, Mani H, Aras O, Rastinehad AR, Shah V, Bernardo M, et al. images using advanced feature analysis. Eur J Cancer. 2012;48(4):441–6. Correlation of magnetic resonance imaging tumor volume with doi:10.1016/j.ejca.2011.11.036. histopathology. J Urol. 2012;188: 44. Aerts HJ, Velazquez ER, Leijenaar RT, Parmar C, Grossmann P, 1157–63. Carvalho S, et al. Decoding tumour phenotype by noninvasive imaging 21. Tan CH, Wei W, Johnson V, Kundra V. Diffusion-weighted MRI in the using a quantitative radiomics approach. Nat Commun. 2014;5:4006. detection of prostate cancer: meta-analysis. AJR Am J Roentgenol. 45. Gevaert O, Xu J, Hoang CD, Leung AN, Xu Y, Quon A, et al. Non-small 2012;199:822–829. cell lung cancer: identifying prognostic imaging biomarkers by leveraging Khalvati et al. BMC Medical Imaging (2015) 15:27 Page 14 of 14 public gene expression microarray data–methods and preliminary results. Radiology. 2012;264(2):387–96. doi:10.1148/radiol.12111607, 2012. 46. Khalvati F, Modhafar A, Cameron A, Wong A, Haider MA. Multi- Parametric Diffusion Magnetic Resonance Imaging Texture Feature Model for Prostate Cancer Analysis. International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), Computational Diffusion MRI Workshop. Boston, Mass., USA; 2014. p. 79–88. 47. Barentsz JO, Richenberg J, Clements R, Choyke P, Verma S, Villeirs G, et al. ESUR prostate MR guidelines 2012. Eur Radiol. 2012;22(4):746–57 22:746–757, 2012. 48. Portalez D, Mozer P, Cornud F, Renard-Penna R, Misrai V, Thoulouzan M, et al. “Validation of the European Society of Urogenital Radiology scoring system for prostate cancer diagnosis on multiparametric magnetic resonance imaging in a cohort of repeat biopsy patients. Eur Urol. 2012;62(6):986–96. doi:10.1016/j.eururo.2012.06.044. Epub. 49. Koh DM, Padhani AR. Diffusion-weighted MRI: a new functional clinical technique for tumour imaging. Br J Radiol. 2006;79:633–5. 50. Rosenkrantz AB, Taneja SS. Radiologist, be aware: ten pitfalls that confound the interpretation of multiparametric prostate MRI. AJR Am J Roentgenol. 2014;202:109–20. 51. Shafiee MJ, Haider SA, Wong A, Lui D, Cameron A, Modhafar A, et al. Apparent Ultra-High b-value Diffusion-Weighted Image Reconstruction via Hidden Conditional Random Fields. IEEE Trans Med Imaging. 2015;34(5):1111–24. 52. Chang C, Lin C. LIBSVM: A library for support vector machines. ACM Trans Intell Syst Technol. 2011;2(3):27:1–27:27. 53. Ozer S, Langer DL, Liu X, Haider MA. Supervised and unsupervised methods for prostate cancer segmentation with multispectral MRI. Med Phys. 2010;37(3):1873–83. 54. Rezaeilouyeh H, Mahoor MH, Zhang J, La Rosa FG, Chang S, Werahera PN. Diagnosis of Prostatic Carcinoma on Multiparametric Magnetic Resonance Imaging Using Shearlet Transform. In: 36 th Intenational Conference of the IEEE Engineering In Medicine and Biology Society (EMBC); 2014. p. 6442–6445. doi:10.1109/EMBC.2014.6945103. 55. Hambrock T, Vos PC, Hulsbergen-van de Kaa CA, Barentsz JO, Huisman HJ. Prostate Cancer: Computer-aided Diagnosis with Multiparametric 3-T MR Imaging–Effect on Observer Performance. Radiology. 2013;266:521–30. 56. Wong A, Khalvati F, Haider MA. Dual-Stage Correlated, Diffusion Imaging. International Symposium on Biomedical Imaging. Brooklyn, NY, USA; 2015, pp. 75–78. 57. Cameron A, Modhafar A, Khalvati F, Lui D, Shafiee MJ, Wong A, et al. Multi-Parametric MRI Prostate Cancer Analysis via a Hybrid Morphological-Textural Model. In: International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). Chicago, Ill., USA; 2014. p. 3357–3360. 58. Cho DS, Khalvati F, Wong A, Haider MA. Prostate DWI Co-Registration via Maximization of Hybrid Statistical Likelihood and Cross-Correlation for Improved ADC and Computed Ultra-High b-value DWI Calculation. In: Proceedings of International Society for Magnetic Resonance in Medicine (ISMRM). Toronto, ON, Canada; 2015. p. 3748. 59. Khalvati F, Wong A, Bjarnason GA, Haider MA. Semi-Automatic Normalized Entropy Characterization of Metastatic Renal Cell Cancer via Spatio-Textural Tumour Classification. In: International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). Chicago, Ill., USA; 2014. Submit your next manuscript to BioMed Central and take full advantage of: • Convenient online submission • Thorough peer review • No space constraints or color figure charges • Immediate publication on acceptance • Inclusion in PubMed, CAS, Scopus and Google Scholar • Research which is freely available for redistribution Submit your manuscript at www.biomedcentral.com/submit

Journal

BMC Medical ImagingSpringer Journals

Published: Aug 5, 2015

There are no references for this article.