Enhancing matRadiomics with a Cascade of 3D U-Nets for Prostate Cancer Segmentation and Digital Biopsy Applications

Aim of this study is to develop an end-to-end deep learning pipeline for the segmentation of prostate cancer to enhance matRadiomics segmentation capabilities. Therefore, two 3D U-Net architectures assembled in cascade were developed, the first one for whole prostate gland segmentation takes as input T2w sequences, while the second one for lesion segmentation takes as input T2w, Apparent Diffusion Coefficient (ADC) images, and high b-value (HBV) Diffusion Weighted Images (DWI), each reduced to the prostate gland volume of interest. On average a whole prostate gland segmentation Dice of 0.87 and a lesion segmentation Dice of 0.53 are reached. Moreover, cancer detection at patient level, measured through the Area Under Curve (AUC), reached a value of 0.75. With this study, we plan to extend the functionalities of matRadiomics and increase its potential for digital biopsy applications.