MRI‐based prostate cancer classification using 3D efficient capsule network

Abstract Background Prostate cancer (PCa) is the most common cancer in men and the second leading cause of male cancer‐related death. Gleason score (GS) is the primary driver of PCa risk‐stratification and medical decision‐making, but can only be assessed at present via biopsy under anesthesia. Magnetic resonance imaging (MRI) is a promising non‐invasive method to further characterize PCa, providing additional anatomical and functional information. Meanwhile, the diagnostic power of MRI is limited by qualitative or, at best, semi‐quantitative interpretation criteria, leading to inter‐reader variability. Purposes Computer‐aided diagnosis employing quantitative MRI analysis has yielded promising results in non‐invasive prediction of GS. However, convolutional neural networks (CNNs) do not implicitly impose a frame of reference to the objects. Thus, CNNs do not encode the positional information properly, limiting method robustness against simple image variations such as flipping, scaling, or rotation. Capsule network (CapsNet) has been proposed to address this limitation and achieves promising results in this domain. In this study, we develop a 3D Efficient CapsNet to stratify GS‐derived PCa risk using T2‐weighted (T2W) MRI images. Methods In our method, we used 3D CNN modules to extract spatial features and primary capsule layers to encode vector features. We then propose to integrate fully‐connected capsule layers (FC Caps) to create a deeper hierarchy for PCa grading prediction. FC Caps comprises a secondary capsule layer which routes active primary capsules and a final capsule layer which outputs PCa risk. To account for data imbalance, we propose a novel dynamic weighted margin loss. We evaluate our method on a public PCa T2W MRI dataset from the Cancer Imaging Archive containing data from 976 patients. Results Two groups of experiments were performed: (1) we first identified high‐risk disease by classifying low + medium risk versus high risk; (2) we then stratified disease in one‐versus‐one fashion: low versus high risk, medium versus high risk, and low versus medium risk. Five‐fold cross validation was performed. Our model achieved an area under receiver operating characteristic curve (AUC) of 0.83 and 0.64 F1‐score for low versus high grade, 0.79 AUC and 0.75 F1‐score for low + medium versus high grade, 0.75 AUC and 0.69 F1‐score for medium versus high grade and 0.59 AUC and 0.57 F1‐score for low versus medium grade. Our method outperformed state‐of‐the‐art radiomics‐based classification and deep learning methods with the highest metrics for each experiment. Our divide‐and‐conquer strategy achieved weighted Cohen's Kappa score of 0.41, suggesting moderate agreement with ground truth PCa risks. Conclusions In this study, we proposed a novel 3D Efficient CapsNet for PCa risk stratification and demonstrated its feasibility. This developed tool provided a non‐invasive approach to assess PCa risk from T2W MR images, which might have potential to personalize the treatment of PCa and reduce the number of unnecessary biopsies.