DEEP LEARNING FOR MRI-BASED DETECTION OF CLINICALLY SIGNIFICANT PROSTATE CANCER: A CRITICAL SYSTEMATIC REVIEW OF DIAGNOSTIC PERFORMANCE, VALIDATION QUALITY, AND CLINICAL TRANSLATION

Abstract

Prostate MRI is central to the diagnostic pathway for suspected organ-confined prostate cancer; however, interpretation remains reader-dependent and sensitive to variability in scanners, protocols, annotations, and reference standards. Deep learning models have therefore been proposed for lesion detection, segmentation, grading, biopsy support, and workflow optimization. This review critically evaluated MRI-based deep learning studies for clinically significant prostate cancer detection, focusing on diagnostic performance, validation quality, reference standards, risk of bias, explainability, and clinical readiness. A systematic search was conducted across PubMed, ScienceDirect, SpringerLink, the Cochrane Library, and Google Scholar for English-language literature published from 2012 to April 2026. Eligible studies used prostate MRI as input and reported artificial intelligence or deep learning methods for detection, segmentation, classification, risk prediction, or MRI workflow optimization. Reporting followed PRISMA 2020 and PRISMA-S. Risk of bias and applicability were assessed using domains adapted from QUADAS-2, PROBAST+AI, CLAIM 2024, and TRIPOD+AI. Thirty MRI-based model studies were included in the primary evidence synthesis. Dominant architectures included convolutional neural networks, 3D CNNs, U-Net/nnU-Net variants, Faster R-CNN-type detectors, ensemble models, and multimodal MRI-clinical systems. Reported AUROC values frequently ranged from approximately 0.80 to 0.95. However, evidence was limited by retrospective designs, inadequate external validation, inconsistent patient-level separation, heterogeneous reference standards, incomplete calibration reporting, and limited clinical-utility analyses. MRI-based deep learning is not yet suitable for autonomous diagnosis but shows value in radiologist-supported lesion localization, triage, targeted-biopsy planning, and quality assurance. Prospective multicentre validation, transparent reporting, pathology-linked endpoints, calibration, decision-curve analysis, failure-mode reporting, and domain-shift assessment remain essential.