An explainable hybrid Generative Adversarial Network framework for data augmentation in imbalanced medical image classification

Class imbalance remains a major challenge in medical image analysis, often leading to biased model learning and reduced diagnostic performance, particularly in chest X-ray and histopathological image classification. To address this challenge, this study presents CiGeN 2.0, a practical hybrid GAN-based data augmentation pipeline that combines the complementary strengths of the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and the Conditional Generative Adversarial Network (cGAN) to generate diverse, class-specific synthetic medical images. WGAN-GP promotes stable training and improves image diversity, whereas cGAN enables label-conditioned image synthesis for targeted augmentation of minority-class samples. The augmented training datasets are subsequently used to train multiple deep learning classifiers, including VGG16, ResNet50, DenseNet121, LSTM, EfficientNet-B0, and Vision Transformer (ViT-B/16). Synthetic image quality is evaluated using Fréchet Inception Distance (FID), Structural Similarity Index Measure (SSIM), and Peak Signal-to-Noise Ratio (PSNR), while classification performance is assessed using accuracy, sensitivity, specificity, F1-score, and area under the receiver operating characteristic curve. Qualitative Grad-CAM visualizations are included to provide supportive insights into model attention patterns. Experimental results demonstrate consistent improvements in classification performance across the evaluated datasets under controlled image-level experimental conditions. The proposed hybrid augmentation pipeline offers a practical strategy for improving learning from imbalanced medical image data. However, the findings are limited to image-level evaluation, and additional patient-wise partitioning, external validation, and expert assessment of synthetic image realism are required to establish broader clinical applicability.