Skin cancer is a major global healthcare burden in which delayed diagnosis substantially reduces treatment effectiveness. Existing AI-based dermatological systems typically rely on centralized, image-only training, limiting privacy, contextual understanding, and interpretability. This work proposes a privacy-preserving multi-agent multimodal framework for multiclass skin cancer classification using dermoscopic images and patient metadata from the HAM10000 dataset. The framework integrates preprocessing, synthetic balancing, augmentation, Optuna-based hyperparameter optimization, transfer learning, multimodal fusion, adaptive ensemble learning, explainability, and federated optimization. Ten candidate transfer-learning architectures are benchmarked, and the top five—EfficientNet-B4, DenseNet201, SE-ResNeXt50, ConvNeXt Base, and Swin Transformer—are combined via a five-agent intelligence layer performing model assessment, ensemble optimization, feature-aware routing, clinical reasoning, and federated aggregation. Federated Averaging keeps raw patient data localized across distributed clients, while Grad-CAM and attention-based explainability improve clinician trust. Evaluated under centralized, equal-client federated, and heterogeneous federated settings, the framework achieves a best classification accuracy of 91.12% with robust multiclass discrimination across all deployment environments.

Skin Cancer Classification,Multimodal Learning,Federated Learning,Privacy-Preserving AI,Adaptive Ensemble Learning,Explainable AI,Clinical Decision Support