A deep learning framework for explainable skin lesion analysis

Skin diseases, particularly viral infections such as monkeypox, have emerged as significant global health concerns due to their rapid transmission, visual similarity to other dermal conditions, and the need for timely and accurate diagnosis. In this study, a hybrid deep learning approach combining U-Net++, FastViT, and the artificial bee colony (ABC) optimization algorithm is proposed for the automatic classification of skin lesions. In the initial phase, lesion regions were segmented using the U-Net++ architecture with pseudo-masks generated through Otsu thresholding. The lesion-focused images obtained from segmentation were subsequently classified using the FastViT-T8 model. Deep features extracted from the FastViT model were optimized via the ABC algorithm to identify the most discriminative feature subset. The selected features were classified using the majority-voting method to enhance decision stability. To ensure model explainability, Gradient-weighted Class Activation Mapping was applied to visualize lesion regions influencing classification decisions. The proposed system was evaluated on the PoxData dataset, comprising monkeypox, chickenpox, cowpox, and healthy classes. Experimental results demonstrated that the proposed approach achieved a classification accuracy of 99.38%, outperforming the convolutional neural network- and vision transformer-based models considered in the study.