Prostate cancer risk assessment models for bi-parametric MRI in a large patient population

Motivation: Although the rate of prostate cancer (PCa) growth varies across individuals, the current active surveillance (AS) strategy for PCa follows a one-size-fits-all approach. Goal(s): To develop an AI-based risk assessment framework to provide physicians with tools to make personalized and objective decisions for AS in prostate. Approach: We train models with representational learning approaches to predict patient-specific risk of prostate cancer (current and future) using biparametric MR images of the prostate acquired on a large patient population (n=28,342). Results: The risk assessment framework demonstrated AUCs of 0.88, 0.84, and 0.82 for predicting current, 2-year, and 5-year risk of prostate cancer. Impact: Automated AI-based risk-assessment frameworks can aid personalized and objective decisions for AS. Patients with higher risks can be managed more aggressively with imaging and biopsy compared to those with lower risks, potentially avoiding overtreatment and overdiagnosis of prostate cancer.