Hybrid Vision Transformers and Federated Learning for Precision and Privacy in Skin Lesion Detection

Skin cancer is a serious health concern globally, and melanoma is the cause of most deaths as it is usually diagnosed too late or misdiagnosed. Early diagnosis of skin lesion identifications may help to reduce the mortality rate. To address this issue, this study proposes a new algorithm called Hybrid Explainable Vision-Aided Deep Learning (HEVADL). The proposed algorithm consists of four core steps: Image Acquisition and Preprocessing, Image Segmentation, Feature Extraction and Classification. This algorithm combines Vision Transformers (ViT) and Deep Convolutional Neural Networks (DCNN) enhanced by a Feature Weight Optimization Module (FWOM) that utilizes Ant Colony Optimization (ACO). The novelty of this framework would bring together the segmentation accuracy with Possibilistic Fuzzy C-Means (PFCM) clustering. The HEVADL model was tested on the HAM10000, ISIC 2020, and DERMOFIT datasets. This model outperformed other methods, including Xception, ResNet50, and DenseNet201, Inception-v3 through adaptive feature rebalancing, multi-modal attention mechanisms, and privacy-protecting Federated Learning. DERMOFIT is the best with an accuracy of 98.70%, followed by ISIC 2020 with 98.46% and HAM10000 with 98.22%. The accuracy of the test results shows 12% increase in comparison with old deep learning models. False negatives decreased by 15%. The research has defined a new benchmark for such easy-to-understand, effective, and safe skin cancer detection and classification systems. HEVADL is the best deep learning model for skin cancer detection, and DERMOFIT is the best dataset.