Hybrid Deep Learning Framework for Skin Disease Classification Using an Enhanced Swin Transformer Network

Automated skin disease classification from dermoscopic and clinical images is challenging because lesions can exhibit substantial variation in color, texture, shape, illumination, scale, and background artifacts. This paper proposes a hybrid image-analysis framework centered on an Enhanced Swin Transformer Network (E-SwinTransNet) for multi-class skin disease classification The framework obtains labeled images from a benchmark dataset, applies median filtering for noise suppression and Contrast Limited Adaptive Histogram Equalization for local contrast enhancement, and then uses rotations, flips and controlled brightness variations to diversify the training set and reduce over fitting. The augmented images are transformed into visual tokens and processed by a hierarchical Swin Transformer with attention refinement and multi-scale feature fusion. A leakage-aware training protocol and comprehensive evaluation strategy are also specified to reach the desired output.