Reinforcement Learning for Medical Image Classification Based on Deep Deterministic Policy Gradients

When it comes to interpreting medical images, traditional neural network training techniques can lead to undesirable outcomes such as offset and noise in image analysis due to problems with parameter (or hyperparameters) selection. We propose a reinforcement learning framework based on Deep Deterministic Policy Gradients (DDPG) to avoid this problem and improve the model’s stability. The goal of this framework is to automatically optimize data augmentation parameters, thus improving the performance of image classification models. Compared to conventional neural networks, our proposed approach for the classification of brain tumor images outperforms traditional data augmentation techniques with improved accuracy (≈ 13%) and loss (≈ 0.15). We also looked at interpretability methods to verify model attention focus and accuracy of prediction, which demonstrate that our model provides better interpretability in accurately distinguishing lesion features. Our findings show that our method based on deep learning models can be advantageous when used in medical image analysis. The code is available at https://github.com/AIPMLab/RL.