Gradient-guided Mechanism for Long-Tailed Medical Image Classification

Due to the difficulty of collecting data samples and the cost of annotation, medical image datasets often exhibit a long-tailed distribution. In the long-tailed distribution, the head classes have a large amount of data, which can dominate the model performance in the training process, leading the model biased towards the majority classes. We find such performance bottleneck is mainly caused by the imbalanced gradients and the positive negative gradients ratio can be seen as a balance indicator in the training process. In this study, we propose a novel loss function considering the accumulative positive-to-negative gradient ratio to reweight the contributions of different classes to solve the long-tailed problems. We conduct experiments on two datasets with different class imbalance ratios using the mainstream classification models with common loss functions and the proposed loss. The results demonstrate that our approach achieves competitive or even superior performance compared to the state-of-the-art (SOTA) techniques.