Explainable Deep Learning for Skin Disorder Classification in Resource-Constrained Settings

Abstract

Skin conditions are among the most common reasons patients seek medical attention worldwide, ranging from self-limiting infections to chronic, disfiguring, or malignant disease. Although deep learning models for dermatological image classification frequently report high overall accuracy, this metric often conceals critical performance failures on clinically vulnerable minority classes such as cancer. To address this, we propose a hybrid classification framework combining DenseNet121 and EfficientNetB0 architectures with transformer-based feature fusion, trained on a DermNet-derived corpus of clinical photographs. Furthermore, we implement a multi-method post- hoc explainability protocol comparing input-gradient saliency, Gradient-weighted Class Activation Mapping (Grad-CAM), SHapley Additive exPlanations (SHAP), and Local Interpretable Model-agnostic Explanations (LIME). The framework was evaluated on a held-out test set of 897 images, measuring overall accuracy, balanced accuracy, and cancer recall alongside cross- method attribution consistency on misclassified cases. The results demonstrate that a high overall accuracy of 91.0% masked a severe cancer recall deficit of 40.0%, whereas multi-method explainability exposed shortcut learning vulnerabilities, with SHAP partially attributing decisions to image watermarks. This work contributes evidence that single-number accuracy and single-method explanations are insufficient, and it supports disaggregated performance reporting and multi-method explanation as minimum requirements for reliable clinical artificial intelligence (AI) evaluation