Multiparametric MRI (mpMRI), including T2WI, DWI/ADC and DCE, is becoming a promising noninvasive tool for prostate cancer (PCa) detection, localization and stage. Although PI-RADS has provides recommendations for image reading and reporting, the interpretation of mpMRI is still challenging for clinical work, for poor interobserver agreement and strong experience dependence. We therefore developed a machine learning model that combines features derived from mpMRI for PCa detection and localization. The model predicted the transition zone (TZ) and peripheral zones (PZ) separately and compared with whole-mount step-section slide. The computer-aided diagnosis (CAD) achieved excellent performance both in PZ and TZ.