An Explainable Uncertainty-Aware Deep Learning Framework for Robust Skin Disease Diagnosis

Abstract

Automated skin-disease diagnosis from clinical and dermoscopic images remains challenging because of illumination variability, acquisition artifacts, irregular lesion boundaries, limited feature diversity, and the low interpretability of conventional deep learning models. This study proposes EDLF-Skin, an integrated framework that combines Physio-Color Consistency preprocessing, Attentive Multi-scale U-Trans segmentation, Multi-Modal Feature Fusion, a Hybrid Ensemble with Uncertainty, and a Multi-level Explainability Interface. The preprocessing stage normalizes illumination and removes artifacts while preserving diagnostically relevant color and texture information. The segmentation module integrates convolutional feature extraction with transformer-based attention to improve lesion-boundary delineation. Spatial, frequency, and texture representations are subsequently fused through cross-attention and processed by complementary convolutional and transformer classifiers. Predictive uncertainty is estimated to identify low-confidence cases requiring expert review, while Grad-CAM, SHAP attribution, and textual rationales provide multilevel explanations. Experimental results show that the proposed segmentation model achieved a Dice coefficient of 0.956 and an intersection over union of 0.935. The complete classification framework obtained 96.87% accuracy, an F1-score of 0.953, an area under the receiver operating characteristic curve of 0.981, and an expected calibration error of 0.021. These findings indicate that EDLF-Skin improves segmentation accuracy, classification performance, probability calibration, and decision transparency within a unified diagnostic workflow.