Segmentation Guided Explainable Deep Learning for Multi-Class Skin Disease Detection

Abstract

This research centers on the creation of a "Segmentation-Guided" Explainable Deep Learning system, built specifically to sharpen the accuracy of skin disease detection and fix the "trust gap" currently plaguing digital dermatology. To move past the limitations of typical "black-box" models—where the inner logic is hidden—this study pushes for a smarter approach: the system first maps out the exact physical boundaries of a lesion using an Attention-U-Net module before it even attempts a diagnosis. We put three cutting-edge hybrid models to the test using the HAM10000 dataset: an efficiency-focused EfficientNetV2 and MobileNetV3 pairing, a structural Swin Transformer mix, and an InceptionV4 plus Vision Transformer (ViT) combo for deep textural analysis. To make sure the data was clean enough for medical use, we applied specialized preprocessing, like the DullRazor method to strip away hair and Contrast Stretching to clarify blurry images. We also tackled the massive problem of data imbalance—where rare, deadly cancers like Melanoma often get drowned out—by using Weighted Frequency Sampling and Focal Loss to ensure these critical cases aren't missed. The real heart of this project is the inclusion of Explainable AI (XAI) via Grad-CAM++, which generates visual heatmaps that show doctors exactly why the AI made its choice. With a high validation accuracy of 90.99%, our results show that this framework successfully ignores distracting artifacts like skin hair or poor lighting to focus purely on the pathology. Ultimately, this work turns AI from a mysterious, opaque algorithm into a transparent assistant, helping bridge the gap between human medical expertise and automated screening.