Multi-class classification of autoimmune skin disease by efficient localization of overlapped lesion boundaries using SAM

BACKGROUND: Detection of Autoimmune skin disease is found challenging due to overlapping features and irregular skin lesion boundaries. Although numerous deep learning models have been proposed for skin disease classification, most are primarily designed for general skin lesions and do not capture the complex and irregular visual characteristics specific to autoimmune skin conditions. RESEARCH DESIGN AND METHODS: The proposed study develops a hybrid deep learning model, AutoImmune-HybridNet, that uses the Segment Anything Model for intelligent isolation of lesion regions, and ResNet50 with a deformable attention transformer for adaptively modeling the irregular shapes and spatial patterns of autoimmune skin lesions. A deformable attention transformer is improved with contrastive learning and the focal loss function to enhance feature discrimination among visually similar conditions. Explainability is incorporated with uncertainty estimation to ensure reliable predictions. RESULTS: The AutoImmune-HybridNet was evaluated on a Human Skin Diseases dataset, and the model achieved an accuracy of 98%, which is better than currently available approaches. CONCLUSIONS: The proposed explainable deep learning framework effectively classifies multiple autoimmune skin diseases while providing an interpretable visualization. Limitations include dataset size and reliance on image-only data will be addressed in future work using multi-modal clinical information.