A Deep Reinforcement Learning Control Framework for Medical Image Augmentation

Data augmentation is essential for handling limited and unbalanced medical datasets in classification tasks. Traditional methods, such as noise injection, random transformations, and generative adversarial networks enhance classifier accuracy, but lack intrinsic optimization for classification metrics. This work proposes a Deep Reinforcement Learning-based framework that autonomously selects and optimizes geometric/color transformations for multi-class medical image classification. Experiments on three medical datasets show that the proposed approach outperforms state-of-the-art augmentation techniques, improving classification accuracy on unseen images by over 13%.