Improving imbalanced medical image classification through GAN-based data augmentation methods

In the medical field, there exists a prevalent issue of data imbalance, severely impacting the performance of machine learning. Traditional data augmentation methods struggle to effectively generate augmented samples with strong diversity. Generative Adversarial Networks (GANs) can produce more effective new samples by learning the global distribution of samples. Although existing GAN models can balance inter-class distributions, the presence of sparse samples within classes can lead to intra-class mode collapse, rendering them unable to effectively fit the sparse region distribution. Based on this, our study proposes a two-step solution. Firstly, we employ a Cluster-Based Local Outlier Factor (CBLOF) algorithm to identify sparse and dense samples intra-class. Then, using these sparse and dense samples as conditions, we train the GAN model to better focus on fitting sparse samples intra-class. Finally, after training the GAN model, we propose using the One-Class SVM (OCS) algorithm as a noise filter to obtain pure augmented samples. We conducted extensive validation experiments on four medical datasets: BloodMNIST, OrganCMNIST, PathMNIST, and PneumoniaMNIST. The experimental results indicate that the method proposed in this study can generate samples with greater diversity and higher quality. Furthermore, by incorporating augmented samples, the accuracy improved by approximately 3% across four datasets.