Automatic diagnosis of malignant prostate cancer patients from mpMRI has been studied heavily in the past years. Model interpretation and domain drift have been the main road blocks for clinical utilization. As an extension from our previous work we trained on a public cohort with 201 patients and the cropped 2.5D slices of the prostate glands were used as the input, and the optimal model were searched in the model space using autoKeras. As an innovative move, peripheral zone (PZ) and central gland (CG) were trained and tested separately, the PZ detector and CG detector were demonstrated effective in highlighting the most suspicious slices out of a sequence, hopefully to greatly ease the workload for the physicians.

Gliomas are the most common primary malignant brain tumors. Accurate segmentation and quantitative analysis of brain tumor are critical for diagnosis and treatment planning. Automatically segmenting tumors and their subregions is a challenging task as demonstrated by the annual Multimodal Brain Tumor Segmentation Challenge (BraTS). In order to tackle this challenging task, we trained 2D non-local Mask R-CNN with 814 patients from the BraTS 2021 training dataset. Our performance on another 417 patients from the BraTS 2021 training dataset were as follows: DSC of 0.784, 0.851 and 0.817; sensitivity of 0.775, 0.844 and 0.825 for the enhancing tumor, whole tumor and tumor core, respectively. By applying the focal loss function, our method achieved a DSC of 0.775, 0.885 and 0.829, as well as sensitivity of 0.757, 0.877 and 0.801. We also experimented with data distillation to ensemble single model's predictions. Our refined results were DSC of 0.797, 0.884 and 0.833; sensitivity of 0.820, 0.855 and 0.820.

Abstract Gleason Grade Group Predictions from mp-MRI of Prostate Cancer Patients using Automated Deep Learning Though histopathology remains the gold standard, there have been significant interests in predicting Gleason Grade using noninvasive imaging such as mp-MRI. Most studies simplify the task into binary classification for the high uncertainty at each group. Handcrafted radiomic features were heavily investigated but prone to errors from the definition of region of interest, feature extraction variations, etc. We proposed an automated deep learning framework (auto-Keras) to predict the group directly based on the 3D data of the whole prostate gland. The training cohort A consisted of 96 PCa patients from SPIE-AAPM-NCI Challenge. The number of patients in each Group was 30, 35, 18, 7, and 6. The testing cohort B consisted of 34 PCa patients from our institute (10, 14, 4, 3, 3). We resampled and rigidly registered ADC and T2WI. N4-bias correction was applied to correct the non-uniformity. For each slice, we performed Gaussian blurring followed by prostate cropping from contour delineated by two clinicians.We tested five scenarios, including input of T2WI, ADC, both, two parallel inputs followed by feature concatenation, and prediction ensemble. The search space of augmentation included translation, flip, rotation, zooming, and contrast. The search space of the architectures had vanilla, ResNet, and Xception. With ADC alone, the model detected 75% of patients in Group 3. Using T2WI and ADC as input, 46% of Group 2 and 40% of Group 1 were identified. Since GG 2 is less aggressive and has a favorable outcome, we further studied the performance of classifying 1 VS. 2-5 and 1-2 VS. 3-5. The models' precision and recall were 91% and 72% for 1-2, 60% and 24% for 3-5. We separated 1 VS. 2-5, with a 96% precision and 73% recall for 2-5. The model had a better performance to predict lower GG when the input contained both T2WI and AD, and better at higher GG when the features were concatenated at the output level. Table 1.Performance of Precision and recall for Gleason Grade Group on the testing cohort.1 VS. 2 VS. 3 VS. 4 VS. 51-2 VS. 3-51 VS. 2-5ADC-onlyGroup 1Group 2Group 3Group 4Group 5Group 1-2Group 3-5Group 1Group 2-5Precision0.100.230.75000.300.500.300.50Recall0.250.380.14--0.580.240.580.24Input MergePrecision0.400.460.25000.910.200.400.61Recall0.310.380.25--0.720.500.310.70Feature MergePrecision0.200.080.250.3300.130.600.200.96Recall0.670.250.250.0600.430.230.670.73PredictionEnsemblePrecision0.200.080.750.3300.170.600.200.91Recall0.500.250.200.10-0.500.240.500.72 Citation Format: Weiwei Zong, Eric Carver, Aharon Feldman, Joon Lee, Zhen Sun, Lanyu Xu, Ali Dabaja, Ning Wen. Gleason grade group predictions from mp-MRI of prostate cancer patients using auto deep learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 186.

Every year thousands of patients are diagnosed with a glioma, a type of malignant brain tumor. MRI plays an essential role in the diagnosis and treatment assessment of these patients. Neural networks show great potential to aid physicians in the medical image analysis. This study investigated the creation of synthetic brain T1-weighted (T1), post-contrast T1-weighted (T1CE), T2-weighted (T2), and T2 Fluid Attenuated Inversion Recovery (Flair) MR images. These synthetic MR (synMR) images were assessed quantitatively with four metrics. The synMR images were also assessed qualitatively by an authoring physician with notions that synMR possessed realism in its portrayal of structural boundaries but struggled to accurately depict tumor heterogeneity. Additionally, this study investigated the synMR images created by generative adversarial network (GAN) to overcome the lack of annotated medical image data in training U-Nets to segment enhancing tumor, whole tumor, and tumor core regions on gliomas. Multiple two-dimensional (2D) U-Nets were trained with original BraTS data and differing subsets of the synMR images. Dice similarity coefficient (DSC) was used as the loss function during training as well a quantitative metric. Additionally, Hausdorff Distance 95% CI (HD) was used to judge the quality of the contours created by these U-Nets. The model performance was improved in both DSC and HD when incorporating synMR in the training set. In summary, this study showed the ability to generate high quality Flair, T2, T1, and T1CE synMR images using GAN. Using synMR images showed encouraging results to improve the U-Net segmentation performance and shows potential to address the scarcity of annotated medical images.

Abstract Objective: Molecular subtypes have been found to be associated with prognosis and treatment response for prostate cancer (PCa) patients and have the potential to aid personalized treatment planning. We aim to stratify molecular subtypes by lesion characterization from multiparametric magnetic resonance images (mp-MRI) using convolutional neural networks (CNN) and knowledge transferred from lesion malignancy classification task. Methods: We identified 23 PCa patients with available mpMRI and molecular subtype information. Each patient may harbor multi-focal lesions of different molecular subtypes. Automated antibody based dual-color immunohistochemistry assays were developed for the simultaneous assessment of ERG-PTEN and ERG-SPINK1 status in PCa on the whole mount microscopic sections. In 7 patients 11 intraprostatic lesions (ILs) were identified as ERG+, 3 patients 3 ILs as SPINK1+, and 10 patients 11 ILs as triple negative (ERG-, SPINK1- and ETS-). We fed a CNN using intratumoral region of interest defined in a slice-by-slice manner from T2WI, ADC and DWIb50. The feature maps from the third convolutional layer were flattened into 7,744 dimensional vectors to represent each IL slice. To compensate for the small sample size, we utilized transfer learning, from the task of tumor malignancy stratification. The CNN model was pre-trained on mpMRI with 320 ILs from 201 patients for malignancy stratification using a different cohort we published previously, which was later fined tuned on this cohort on malignancy classification for domain adaptation purpose. Results: The clustering accuracy (Top 3 estimates) using cosine similarity metrics was shown in Table 1 for each molecular subtype category respectively. The preliminary results supported our hypothesis that task of lesion malignancy and molecular sub-types stratification were correlated in the imaging features derived from mpMRI. Table 1:Clustering accuracy for each molecular category. Each ILs input had 3 consecutive slices in a sequence, know as 2.5D, to incorporate lesion growth pattern.Accuracy (# of 2.5D image slices)Top 1 predictionsTop 2 predictionsTop 3 predictionsERG+ (21)0.670.900.95SPINK1+ (4)0.50.750.75Negative (17)0.650.760.76 Conclusions: This work showed the potential to classify the molecular subtypes of PCa from mp-MRI. The small sample size problem was tackled using transfer learning. Acknowledgement: The work was supported by a Research Scholar Grant: RSG-15-137-01-CCE from the American Cancer Society. Citation Format: Weiwei Zong, Eric N. Carver, Aharon Feldman, Nallasivam Palanisamy, Ning Wen. Molecular subtype stratification for prostate cancer from mpMRI and histopathology images using convolutional neural networks and transfer learning [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 5302.

PURPOSE:Deep learning models have had a great success in disease classifications using large data pools of skin cancer images or lung X-rays. However, data scarcity has been the roadblock of applying deep learning models directly on prostate multiparametric MRI (mpMRI). Although model interpretation has been heavily studied for natural images for the past few years, there has been a lack of interpretation of deep learning models trained on medical images. In this paper, an efficient convolutional neural network (CNN) was developed and the model interpretation at various convolutional layers was systematically analyzed to improve the understanding of how CNN interprets multimodality medical images and the predictive powers of features at each layer. The problem of small sample size was addressed by feeding the intermediate features into a traditional classification algorithm known as weighted extreme learning machine (wELM), with imbalanced distribution among output categories taken into consideration.METHODS:The training data collection used a retrospective set of prostate MR studies, from SPIE-AAPM-NCI PROSTATEx Challenges held in 2017. Three hundred twenty biopsy samples of lesions from 201 prostate cancer patients were diagnosed and identified as clinically significant (malignant) or not significant (benign). All studies included T2-weighted (T2W), proton density-weighted (PD-W), dynamic contrast enhanced (DCE) and diffusion-weighted (DW) imaging. After registration and lesion-based normalization, a CNN with four convolutional layers were developed and trained on tenfold cross validation. The features from intermediate layers were then extracted as input to wELM to test the discriminative power of each individual layer. The best performing model from the tenfolds was chosen to be tested on the holdout cohort from two sources. Feature maps after each convolutional layer were then visualized to monitor the trend, as the layer propagated. Scatter plotting was used to visualize the transformation of data distribution. Finally, a class activation map was generated to highlight the region of interest based on the model perspective.RESULTS:Experimental trials indicated that the best input for CNN was a modality combination of T2W, apparent diffusion coefficient (ADC) and DWIb50 . The convolutional features from CNN paired with a weighted extreme learning classifier showed substantial performance compared to a CNN end-to-end training model. The feature map visualization reveals similar findings on natural images where lower layers tend to learn lower level features such as edges, intensity changes, etc, while higher layers learn more abstract and task-related concept such as the lesion region. The generated saliency map revealed that the model was able to focus on the region of interest where the lesion resided and filter out background information, including prostate boundary, rectum, etc. CONCLUSIONS: This work designs a customized workflow for the small and imbalanced dataset of prostate mpMRI where features were extracted from a deep learning model and then analyzed by a traditional machine learning classifier. In addition, this work contributes to revealing how deep learning models interpret mpMRI for prostate cancer patient stratification.

PURPOSE:Accurate delineation of the prostate gland and intraprostatic lesions (ILs) is essential for prostate cancer dose-escalated radiation therapy. The aim of this study was to develop a sophisticated deep neural network approach to magnetic resonance image analysis that will help IL detection and delineation for clinicians.METHODS AND MATERIALS:We trained and evaluated mask region-based convolutional neural networks to perform the prostate gland and IL segmentation. There were 2 cohorts in this study: 78 public patients (cohort 1) and 42 private patients from our institution (cohort 2). Prostate gland segmentation was performed using T2-weighted images (T2WIs), although IL segmentation was performed using T2WIs and coregistered apparent diffusion coefficient maps with prostate patches cropped out. The IL segmentation model was extended to select 5 highly suspicious volumetric lesions within the entire prostate.RESULTS:The mask region-based convolutional neural networks model was able to segment the prostate with dice similarity coefficient (DSC) of 0.88 ± 0.04, 0.86 ± 0.04, and 0.82 ± 0.05; sensitivity (Sens.) of 0.93, 0.95, and 0.95; and specificity (Spec.) of 0.98, 0.85, and 0.90. However, ILs were segmented with DSC of 0.62 ± 0.17, 0.59 ± 0.14, and 0.38 ± 0.19; Sens. of 0.55 ± 0.30, 0.63 ± 0.28, and 0.22 ± 0.24; and Spec. of 0.974 ± 0.010, 0.964 ± 0.015, and 0.972 ± 0.015 in public validation/public testing/private testing patients when trained with patients from cohort 1 only. When trained with patients from both cohorts, the values were as follows: DSC of 0.64 ± 0.11, 0.56 ± 0.15, and 0.46 ± 0.15; Sens. of 0.57 ± 0.23, 0.50 ± 0.28, and 0.33 ± 0.17; and Spec. of 0.980 ± 0.009, 0.969 ± 0.016, and 0.977 ± 0.013.CONCLUSIONS:Our research framework is able to perform as an end-to-end system that automatically segmented the prostate gland and identified and delineated highly suspicious ILs within the entire prostate. Therefore, this system demonstrated the potential for assisting the clinicians in tumor delineation.

Automatic delineation of Glioblastoma (GBM) plays an important role in radiation therapy. Recently, segmentation algorithms using supervised deep neural networks (DNN) have shown promising results, but small volumes of annotated data pose challenges on powering them. Current collection of dataset relies on radiologists’ contour as ground truth and is expensive and time-consuming. One possible solution to overcome the limitation of small dataset is to generate synthetic MR images representing different clinical scenario. The aim of this study is to apply a generative adversarial network (GAN) to synthesize highly realistic MR images from manipulated annotations that are able to feed as new training samples for DNNs. Data was obtained from the BraTS multimodal Brain Tumor Segmentation Challenge 2018. 19 different institutions provided a total of 210 patients. T1WI, T1CE, T2WI, and FLAIR were provided for each patient. 82 patients were used for training and 128 patients for validation. The network consisted of a generator and two discriminators. Image per-pixel loss, perceptual loss and adversarial loss were used. By manipulating on annotations from radiologists, the generator was able to output new synthetic MR images, and boost the size of dataset. The realism of synthetic images was evaluated both quantitatively and qualitatively. Synthetic image generated from non-manipulated annotation was compared with its corresponding real image. Mean Square Error (MSE), Mean Absolute Error (MAE), Peak Signal to Noise Ratio (PSNR), and Structural Similarity Index (SSIM) for synthetic MR images were 19.246±0.308, 23.375±0.586, 43.068±0.443 and 0.788 ± 0.002; 19.249±0.274, 22.805±0.583, 43.054±0.437 and 0.789 ± 0.004; 19.246±0.290, 23.391±0.400, 43.102±0.45 and 0.784 ± 0.003; 18.930±0.40, 24.119±1.48, 43.126±0.46 and 0.794 ± 0.005, respectively, for T1, T1CE, T2 and Flair. A subset of 9 real and 10 generated patients were assessed by a physician. 8.3%, 41.7%, 50% of real images and 22.5%, 47.5%, 30% of synthetic images were commented as poor, marginal and good quality. The misclassified rate were 26.3%, 10.5%, 26.3% and 26.3% for T1, T1CE, T2 and Flair. We proposed to apply GAN to synthesize GBM MR images from manipulated annotations to increase the dataset size to train deep learning segmentation models. The evaluation results showed synthetic MRIs had comparable image quality to real MRIs that had potential to be used for DNN training.

Extraction of multiscale radiomic features from preoperative MRI scans provides an opportunity for quantitative, non-invasive, image-based phenotyping of glioblastoma (GBM). Upon obtaining tumor tissue, genomic sequencing can further enhance predictive value. This study aims to predict overall survival (OS) following gross total resection of primary glioblastoma using a combination of MRI radiomics, tumor genomics, and patient clinical factors. In this retrospective study, preoperative image data from 61 patients from the 2018 Multimodal Brain Tumor Image Segmentation (BraTS) Competition publicly-available dataset were used to generate a radiomic signature. A total of 968 radiomic features were extracted from three manually delineated structures (Enhancing Tumor, Tumor Core, and Tumor Edema) on the T1, T2, T1-contrast, and FLAIR sequences. The features associated with overall survival were selected using the univariate Cox Proportional Hazards (CPH) model (p-value<0.05) and 10 cross-fold SVM recursive feature elimination (SVM-RFE). These features were then tested on a validation set of patients with GBM treated at our institution over five years (2013-2018). A total of 152 patients were identified, of which 20 had radiologically confirmed gross total resection on postoperative MRI as well as genomic sequencing information available. Institutional preoperative MRI images—T1, T2, T1-contrast, and FLAIR sequences—were manually contoured according to established BraTS benchmarks. Genomic features included expression levels of nine genes involved in neuroactive ligand-receptor interaction, cysteine metabolism, and ephrin A reverse signaling. Clinical variables included in the analysis were age, sex, tumor location, and Karnofsky Performance Score (KPS). A multivariate Cox proportional hazards analysis was performed to assess the association between OS and the radiomic, genomic, and clinical features. Ten radiomic features were found to be statistically relevant with OS from a training set of 61 patients. Features associated with OS on the univariate analysis included age, tumor location, KPS, LDHA gene expression and EPHA5 gene expression. In the multivariate analysis, the features significantly associated with OS included tumor location (p=0.03, Hazard ratio (HR) = 1.09), LDHA (p<0.005, HR=0.14), and two radiomics features from gray-level size-zone matrix: minimum value of large-zone-low-gray-level emphasis from T2 (p=0.04, HR=0.49) and kurtosis of small-zone-low-grey-level-emphasis from FLAIR (p=0.01, HR=2.02). A more robust validation set of patients is needed to draw more meaningful conclusions, but there exist potential imaging features that can provide clinical prognostic information. One aspect that necessitates further investigation is the failure of well-established clinical and genomics factors to reach significance in the present study.

Purpose: This study was designed to evaluate the ability of a U-net neural net-work to properly identify three regions of a brain tumor and an ELM for the prediction of patient overall survival after gross tumor resection using preoperative MR images. Methods: 210 GBM patients were used for training, while 66 LGG and GBM patients were used for validation. Multiple preprocessing steps were performed on each patient’s data before loading them into the model. The segmentation model consists of three different U-nets, one for each region of interest. These created segmentations were then analyzed by use of common quantitative metrics with respect to physician created contours. Regarding the patient overall survival prediction, 59 high grade glioma patients with gross total resection (GTR) were provided for training. 28 patients with GTR were used to validate the algorithm. Results: The average [s.d] DSC for the whole tumor, enhanced tumor, and tumor core contours were 0.882 [0.080], 0.712 [0.294], and 0.769 [0.263], respectively. The average [s.d.] Hausdorff distance were 7.09 [11.57], 4.46 [8.32], and 9.57 [14.08], respectively. The average [s.d.] sensitivity for the whole tumor, enhanced tumor, and tumor core contours were 0.887 [0.126], 0.770 [0.245], and 0.750 [0.293], respectively. The average [s.d.] specificity was 0. 993 [0.005], 0.998 [0.003], 0.998 [0.002], respectively. The predictive power of patient overall survival is 0.607 using an extreme learning machine algorithm. Conclusion: The U-Net model was very effective in determining accurate location of the whole tumor and segmenting the whole tumor, enhancing tumor and tumor core. The most predictive features of patient overall survival are both age and location of the tumor when all 163 validation cases were utilized.

Abstract INTRODUCTION Pre-operative differentiation of IDH mutant gliomas from similar appearing pathologies on imaging prior to definitive surgical diagnosis may aid treatment navigation, maximize the surgical approach, and provide diagnostic support for inoperable tumors. Quantitative image feature analysis offers a potential non-invasive method to identify diagnostic, prognostic, and predictive imaging biomarkers. We investigated the use of radiomic MR imaging features to classify tumors based on IDH mutation status. METHOD Pre-operative T1-weighted (T1W), T2-weighted (T2W), T1-contrast enhanced (T1CE), and fluid attenuated inversion recovery (FLAIR) MR brain images, along with patient IDH mutation status (mutant/wildtype) were obtained for 128 glioma patients from The Cancer Genome Atlas (TCGA). Enhancing tumor was delineated by GLISTRboost. GlistrBoost is a hybrid-discriminative model that segments tumors based on an expectation-maximization framework with a classification scheme and uses a probabilistic Bayesian strategy for segmentation refinement. MR studies for 78 glioma patients from six institutions were used for training and 50 glioma patients from a different institution were used for validation. Pre-processing included registration, resampling, and normalization. Cancer Imaging Phenomics Toolkit (CaPTK) extracted 938 radiomic image features per sequence for the enhancing tumor contour. Relevance of each individual feature was determined by the least absolute shrinkage and selection operator (LASSO). The ability of relevant radiomic image features to identify mutation status of IDH was assessed by logistic regression. RESULTS LASSO identified one highly informative radiomic imaging feature, the minimum of the mean absolute histogram deviation on T1 MR images, which was able to predict IDH mutation status with an accuracy of 0.74, precision of 1.0, and recall of 0.32. CONCLUSION Non-invasive prediction of IDH mutation status from pre-surgical MR images offers potential diagnostic, therapeutic, and prognostic benefits for glioma patients. Quantitative image feature analysis is a feasible method for identifying potential radiomic imaging features.

Introduction: Multiparametric MR imaging (mpMRI) has shown promising results in the diagnosis and localization of prostate cancer. Furthermore, mpMRI may play an important role in identifying the dominant intraprostatic lesion (DIL) for radiotherapy boost. We sought to investigate the level of correlation between dominant tumor foci contoured on various mpMRI sequences.Methods: mpMRI data from 90 patients with MR-guided biopsy-proven prostate cancer were obtained from the SPIE-AAPM-NCI Prostate MR Classification Challenge. Each case consisted of T2-weighted (T2W), apparent diffusion coefficient (ADC), and Ktrans images computed from dynamic contrast-enhanced sequences. All image sets were rigidly co-registered, and the dominant tumor foci were identified and contoured for each MRI sequence. Hausdorff distance (HD), mean distance to agreement (MDA), and Dice and Jaccard coefficients were calculated between the contours for each pair of MRI sequences (i.e., T2 vs. ADC, T2 vs. Ktrans, and ADC vs. Ktrans). The voxel wise spearman correlation was also obtained between these image pairs.Results: The DILs were located in the anterior fibromuscular stroma, central zone, peripheral zone, and transition zone in 35.2, 5.6, 32.4, and 25.4% of patients, respectively. Gleason grade groups 1–5 represented 29.6, 40.8, 15.5, and 14.1% of the study population, respectively (with group grades 4 and 5 analyzed together). The mean contour volumes for the T2W images, and the ADC and Ktrans maps were 2.14 ± 2.1, 2.22 ± 2.2, and 1.84 ± 1.5 mL, respectively. Ktrans values were indistinguishable between cancerous regions and the rest of prostatic regions for 19 patients. The Dice coefficient and Jaccard index were 0.74 ± 0.13, 0.60 ± 0.15 for T2W-ADC and 0.61 ± 0.16, 0.46 ± 0.16 for T2W-Ktrans. The voxel-based Spearman correlations were 0.20 ± 0.20 for T2W-ADC and 0.13 ± 0.25 for T2W-Ktrans.Conclusions: The DIL contoured on T2W images had a high level of agreement with those contoured on ADC maps, but there was little to no quantitative correlation of these results with tumor location and Gleason grade group. Technical hurdles are yet to be solved for precision radiotherapy to target the DILs based on physiological imaging. A Boolean sum volume (BSV) incorporating all available MR sequences may be reasonable in delineating the DIL boost volume.

Mask-RCNN is a deep structural learning algorithm that has been investigated in other industries for structure mapping and recognition. We attempted to use this platform for normal prostate segmentation and dominant intraprostatic lesion (DIL) segmentation and localization on multi-parametric MRI (Mp-MRI). This can potentially aid in diagnosis and therapeutic planning, as precise localization of the site of disease can guide targeted surgical resection or radiation boost volumes. A total of 78 patients with biopsy proven prostate adenocarcinoma and MRI imaging were reviewed. Of these, 54 patients were used as a training set. The remaining 24 patients were selected for normal prostate segmentation. This cohort was then evenly divided into validation and testing sets. For the DIL localization and segmentation review, 57 patients were used as a training set. Of the remaining 21 patients, 11 were included in the validation set and 10 were used for the testing set. Mp-MRI images, including T2-weighted (T2W) and apparent diffusion coefficient (ADC) images were adopted for our experiment. T2W images alone were used in the normal prostate segmentation task and co-registered ADC maps to T2W images with the prostate patch cropped out were used in the DIL detection and segmentation task. All images were normalized, the histogram was equalized, and images were resized and padded slice by slice to a fixed size of 384 × 384 pixels prior to the training, validation and testing phase. The normal prostate and DILs were contoured by a radiation oncologist from our institution. The DIL was separately contoured for 20 patients by a second radiation oncologist to assess for interobserver variability. Sensitivity and specificity for prostate segmentation were evaluated for all slices of patient imaging and was calculated to be 0.94 and 0.92, respectively. The dice similarity coefficient (DSC) was 0.87±0.04. The 95% hausdorff distance was 6.12±2.39 mm. To evaluate the performance of Mask-RCNN, we used 2D and 3D Unet benchmarks, which had the DSC of 0.85±0.03 and 0.83±0.07, respectively. For the DIL segmentation, our overall detection rate was 81%, and for those detected lesions, the DSC was 0.64±0.14. The DSC of the contours between two clinicians was 0.67±0.21. Our results demonstrate that the Mask-RCNN model can produce high quality segmentation results for prostate recognition on standard MRI images and can potentially aid in detecting and segmenting prostate cancer on Mp-MRI images. There was also similar interobserver variability on the DIL segmentation using Mask-RCNN. These results can potentially influence therapeutic options, such as targeted resection or intraprostatic boost volumes.