Deep Learning Based Detection of Prostate Cancer in MRI Using Biopsy-Confirmed Ground Truth

Prostate cancer is one of the most common malignancies in men, and accurate lesion segmentation in magnetic resonance imaging (MRI) is essential for diagnosis, treatment planning, and disease monitoring. Manual delineation by radiologists is time-consuming and subject to interobserver variability. This study presents an automated, deep learning-based framework for 3D prostate lesion detection using modified U-Net architectures, guided by pathology-informed ground truth. The proposed approach leverages biopsy-verified lesion masks derived from the PROSTATEx and PROSTATEx2 datasets, ensuring biologically validated reference labels. Method 1 uses dice loss optimization to train a simplified 3D U-Net on full volume MRI data, while Method 2 uses a patch-based 3D U-Net with advanced preprocessing, extensive data augmentation, and a dice focal loss to reduce class imbalance and improve lesion localization. With a Dice similarity coefficient (DSC) of 92.3% and an intersection over union (IoU) of 87.8%, the quantitative data shows that the patch-oriented network performs better in segmentation. In contrast to models trained only on radiologist annotations, the work shows that pathology-informed learning improves lesion delineation accuracy, highlighting its potential for strong clinical translation in MRI-guided prostate cancer detection.