Federated ViT: A Distributed Deep Learning Framework for Skin Cancer Classification

Skin cancer remains one of the most prevalent and life-threatening diseases globally, where early and accurate diagnosis is crucial for effective treatment and improved patient outcomes. Although deep learning and digital pathology have advanced lesion classification performance, significant challenges remain due to complex histopathological patterns, privacy concerns over centralized data storage, and inconsistencies across datasets in terms of annotation, quality, and structure. These disparities hinder model generalization and complicate performance benchmarking. To overcome these issues, this study proposes a novel skin lesion classification framework that integrates Federated Learning (FL) with Vision Transformers (ViTs). FL enables decentralized model training across multiple clinical sites without sharing sensitive data, inherently accommodating data heterogeneity while preserving privacy. ViTs, with their self-attention mechanisms, effectively capture global contextual features, making them robust to structural and visual variations in non-standardized datasets. The framework incorporates key components such as client-side preprocessing, data augmentation, federated transfer learning, model aggregation, and feature optimization. Comprehensive experiments were conducted on two publicly available skin lesion datasets—HAM10000 and ISIC2019. The proposed model achieved a global test accuracy of 90%, with a sensitivity of 88.2% and specificity of 91.4% on the HAM10000 dataset. On the ISIC2019 dataset, the model attained a test accuracy of 87.6%, with Area Under the Curve (AUC) values reaching 0.96 for the best-performing class, demonstrating strong discriminative ability. These results confirm the framework’s robustness in real-world scenarios with non indentical data and its potential to support privacy-preserving, scalable clinical applications in dermatology. This work contributes to advancing FL-based computer vision solutions for skin cancer detection and lays the groundwork for future exploration of ViTs in decentralized healthcare environments.