LEGAN: Addressing Intra-class Imbalance in GAN-based Medical Image Augmentation for Improved Imbalanced Data Classification

Currently, medical image classification is challenged by performance degradation due to imbalanced data. Balancing the data through sample augmentation proves to be an effective solution. However, traditional data augmentation methods and simple linear interpolation fall short in generating more diverse new samples, thereby limiting the enhancement of results with augmented data. Although Generative Adversarial Networks (GAN) models have the potential to generate more diverse samples, current GAN models struggle to effectively address the issue of intra-class mode collapse. In this research, we propose a GAN model structure named LEGAN, based on Local Outlier Factor (LOF) and information entropy, to address this problem. The LEGAN model focuses on resolving mode collapse caused by intra-class imbalances. Firstly, LOF is used to detect sparse and dense sample points in intra-class imbalance, and affine transformations are performed on sparse sample points to enhance the diversity of sample data and features. Then, we train LEGAN jointly using the augmented sparse samples and dense samples to effectively learn the sample distribution in sparse regions, thereby generating more diverse sparse samples. Secondly, we propose a decentralization constraint based on information entropy. This method measures the diversity of generated samples using information entropy during the training process and provides feedback to the generator, encouraging it to optimize towards better diversity. We conducted extensive experiments on three medical datasets, namely BloodMNIST, OrgancMNIST, and PathMNIST, demonstrating that LEGAN can achieve more diverse intra-class sample generation. The quality of the generated images and the classification performance are both significantly improved.