Ensemble Deep Learning Models for Segmentation of Prostate Zonal Anatomy and Pathologically Suspicious Areas

Prostate cancer is the most frequently occurring type of malignant tumor and one of the leading causes of cancer death in men. This article presents the development and evaluation of ensemble models for the segmentation of prostate zones (the anatomical model) and pathologically suspicious areas (the detection model). The novelty of the solution lies in its incorporation of pre-training into the standard nnU-Net architecture and its use of a modified loss function, which considers the subjective variability of annotations between experts. The anatomical segmentation model achieved a dice score (DSC) of 0.915 for the prostate gland, 0.865 for the transition zone, and 0.736 for the peripheral zone. The lesion detection models delivered promising performance, with an area under the receiver operating characteristic curve (AUROC) of 0.9692/0.8727 and an average precision (AP) of 0.798/0.68 for radiological and clinical settings, respectively. The authors highlight the importance of incorporating domain knowledge, customised architectures, and uncertainty estimation to develop robust and clinically relevant AI-based solutions for this complex medical imaging task.