Explainable Multimodal Skin Lesion Classification via Supervised Contrastive Learning and Attention-Based Fusion

Early skin-lesion diagnosis is challenging due to high visual similarity between lesion categories and severe class imbalance in clinical datasets. We propose FusionClassifier, a multimodal framework that combines a Swin-Small Transformer with supervised contrastive learning and attention-based feature fusion. A two-phase training strategy first learns discriminative image representations and then integrates them with patient metadata, including age, sex, and anatomical site. Grad-CAM and Grad-CAM++ are further employed to provide visual explanations of model predictions. Experiments on the HAM10000 dataset using five-fold cross-validation achieve 98.17% accuracy, 97.28% F1-score, and 99.80% AUC-ROC. Ablation results show that supervised contrastive learning substantially improves the Swin-Small baseline from 73.58% to 98.21% accuracy. The results demonstrate the effectiveness of combining discriminative representation learning, multimodal fusion, and explainability for skin-lesion classification.