DM-GAN: A Data Augmentation-Based Approach for Imbalanced Medical Image Classification

In medical data processing, the issue of data imbalance often leads to biased learning outcomes in machine learning methods. Generative Adversarial Networks (GAN) can generate realistic and diverse augmented samples to balance sample distribution. However, the problems of intra-class imbalance and sparse isolated samples often affect the performance of GAN models. Based on this, this paper proposes a multi-generator GAN model architecture, named Diversity Multi-Generator GAN (DM-GAN). This model combines self-attention mechanisms and diversity loss functions to improve the quality and diversity of generated samples. Specifically, we designed multiple independently trained generators to capture more sample patterns and introduced self-attention modules in both the generator and discriminator to enhance the model’s ability to capture image details. Additionally, we proposed an improved generator loss function that combines mode-seeking loss and mutual exclusion loss. By encouraging the generation of different samples and reducing sample overlap, this approach enhances the diversity and coverage of generated samples. Experimental results on two real-world medical image datasets demonstrate that DM-GAN shows significant advantages in handling intra-class imbalance and isolated sample issues. Compared to existing methods, our approach not only achieves better results in terms of the quality and diversity of generated samples but also effectively improves the performance of downstream classification tasks.