A Deep Reinforcement Learning Framework for Medical Image Analysis: Hybrid Rewards and Class-Specific Precision Histograms

This article proposes a deep reinforcement learning (DRL) framework for medical image analysis, combining ResNet-۱۰۱ and Proximal Policy Optimization (PPO) to address data scarcity, class imbalance, and interpretability. A hybrid reward system prioritizes clinical accuracy, while class-specific precision histograms expose performance disparities. Validated on histopathology, MRI, and X-ray datasets, the model achieves ۹۷.۱% accuracy in lung cancer classification and a Dice score of ۰.۹۴ for tumor segmentation. Full reproducibility is ensured via Google Colab code.

کلیدواژه ها:

image ، analysis ، segmentation ، classification ، optimization