Synthetic Data Augmentation for Multi-Class Skin Lesion Classification: A Hybrid EfficientNet-Vision Transformer Framework with Explainable AI

Abstract

The research explores the utilization of Stable Diffusion models for synthetic image augmentation aimed at generating an improved dataset to enhance the performance of deep learning models in diagnosing skin diseases. The study utilizes the HAM10000 dataset, which consists of seven skin lesion categories exhibiting severe class imbalance, to fine-tune a class-conditional Stable Diffusion model for minority-class image synthesis. The resultant synthetic images, evaluated at a Fréchet Inception Distance (FID) of 98.0, are combined with real training images to train a dual-branch hybrid architecture,SkinHybrid,fusing EfficientNet and Vision Transformer (ViT) representations. An ablation study demonstrates that incorporating Stable Diffusion-generated samples improves the weighted F1-score from 84.1% to 97.0%. Post-hoc explainability via Grad-CAM confirms that the model attends to clinically relevant dermoscopic morphology