An Explainable Deep Reinforcement Learning Approach for Clinically Significant Prostate Cancer Detection Using Multiparametric MRI

Abstract

The accurate detection of clinically significant prostate cancer (csPCa) using multipara metric MRI (mpMRI) is critical for guiding biopsy and treatment decisions. While reinforcement learning offers promising avenues for automation, developing high-performance models is often constrained by the limited size of annotated medical imaging datasets. The proposed algorithm aims to develop and validate a novel deep learning algorithm that leverages transfer reinforcement learning to enhance the detection and classification of csPCa from mpMRI sequences, ultimately aiming to improve diagnostic accuracy and reduce unnecessary biopsies. The transfer reinforcement learning-based model achieved a high AUC of 0.96 (95% CI: 0.91-0.96) for classifying csPCa on the internal test set, significantly outperforming a model trained from scratch (AUC: 0.85). The algorithm demonstrated robust performance on the external validation set (AUC: 0.91), confirming its generalizability. It showed a sensitivity of 94%, a specificity of 90.8%, and an NPV of 96% at the optimal operating threshold.