Deep Learning Classification of Skin Diseases with Explainable Predictions

Abstract

Skin Diseases are common health problems worldwide. The signs of infections are invisible, which can cause physical health distress as well as mental depression. It sometimes leads to skin cancer in severe cases. Diagnosing skin diseases from the clinical images is one of the most challenging tasks in medical imaging analysis. Moreover, when performed manually by medical experts, diagnosing skin diseases is time-consuming and subjective. As a result, both patients and dermatologists require an automated skin disease prediction model to expedite treatment planning. This paper describes a deep learning system that classifies skin lesions using GRAD-CAM visualizations to make model decisions inspectable. We trained on two standard benchmarks, the ISIC 2019 challenge and HAM10000, across eight lesion types: melanoma, melanocytic nevus, basal cell carcinoma, actinic keratosis, benign keratosis, dermatofibroma, vascular lesion, and squamous cell carcinoma. A decision tree classifier handles the extracted features, with Swin attention maps providing a second interpretability layer alongside GRAD-CAM. The performance metrics are solid. That tracks for a model trained on balanced data.