DermaLite - Multi Modal Evidential Deep Learning for Skin Lesion Classification

Abstract

Automated skin lesion classification from dermoscopic images is challenging due to class imbalance and the need for reliable confidence estimates in clinical settings. This paper proposes DermaLite, a multi-modal evidential deep learning framework for skin lesion classification. We use StyleGAN2-ADA to generate synthetic dermoscopic images that balance the HAM10000 dataset, which originally has severe imbalance (melanocytic nevi comprise 67% of samples while vascular lesions are under 2%). The model combines EfficientNet-B3 visual features with clinical metadata such as patient age and lesion location through a cross-attention fusion layer. An evidential classification head outputs Dirichlet-distributed probability, providing explicit uncertainty scores alongside predictions. Grad-CAM is used for visual explanation. On the HAM10000 test set, DermaLite achieves 96.17% accuracy. The evidential head gives well-calibrated uncertainty (ECE = 0.023), and ablation experiments confirm that each component contributes meaningfully to performance.