Reinforcement Learning for Breast Cancer Diagnosis with Visual Interpretability

This paper presents a reinforcement learning (RL) based approach for breast cancer diagnosis using mammographic images. The diagnosis task is formulated as a binary classification problem between malignant and benign findings. We adopt the Proximal Policy Optimization (PPO) algorithm combined with a custom convolutional neural network (CNN) feature extractor and Grad-CAM-based visual explanations. To emulate a diagnostic workflow, we design a custom multi-step environment where the RL agent processes sequences of mammograms. We report training performance achieving a maximum validation accuracy of 77.1% with mean performance of 72.2% ± 4.5% across five independent runs, classification metrics, and qualitative visualizations of attention maps to demonstrate the model’s interpretability. Our approach bridges the gap between traditional static classifiers and sequential decision-making processes that better reflect clinical diagnostic workflows