X-GAN: Explainable Generative Adversarial Networks for Rare Disease Data Augmentation and Clinical Insights

Rare disease imaging faces dual challenges: limited annotated data and the lack of interpretability in AI-based diagnostic models. To address these issues, we propose X-GAN, a novel framework that integrates StyleGAN2 for synthetic data generation with advanced explainable AI (XAI) techniques, including Grad-CAM++ and SHAP. The model enhances data diversity while providing transparency into both generative and diagnostic processes. Synthetic images achieved high fidelity with a Frechet Inception Distance (FID) of 18.4 and Structural Similarity Index (SSIM) of 0.91. Diagnostic models trained on real and synthetic data achieved a classification accuracy of 92.6% and a Dice Coefficient of 0.89, outperforming state-of-the-art methods. Clinical experts rated the outputs highly for diagnostic relevance and trustworthiness. This integration of GAN and XAI techniques demonstrates a significant advancement in addressing data scarcity, improving model interpretability, and fostering clinical adoption of AI in rare disease workflows.