Sorting out PI-RADS 3: A radiologist-assist tool using representation learning to avoid unnecessary biopsies

Motivation: Although MRI has a high negative predictive value for low-risk prostate cancers (PCa), it suffers from substantially high-number of false-positives for intermediate-risk cases, resulting in unnecessary biopsies. Goal(s): To develop an AI-based framework to detect clinically significant PCa (csPCa) in patients diagnosed as intermediate-risk by radiologists. Approach: We use PI-RADS-guided contrastive learning to generate latent representations from MR images. These serve as a guide to disambiguate PI-RADS3 and provide a tool for identifying potential negative biopsies. Results: We observe performance comparable to radiologists in identifying csPCa. Improved performance is noted in PI-RADS3 assessments where 10-28% negative biopsies can be avoided with appropriate risk-thresholds. Impact: Powered with PI-RADS guided representational learning, deep learning models can provide radiologists with additional information to disambiguate intermediate risk PI-RADS3 assessments, avoiding unnecessary biopsies, and potentially helping patient retention in active surveillance protocols.