A Vision Transformer-Based Model for Skin Disease Classification
Document Type
Conference Proceeding
Publication Date
2026
Abstract
A timely and accurate diagnosis of dermatological conditions is critical for improving treatment outcomes and reducing the burden on healthcare systems. This paper proposes a vision transformer-based model for skin disease classification using the Data-efficient Image Transformer (DeiT-Tiny). The proposed model is applied on a publicly available dataset called subirbiswas19 comprising eight skin disease categories. Two configurations are evaluated: a baseline with minimal preprocessing, and an enhanced pipeline incorporating Gaussian blur, Contrast-Limited Adaptive Histogram Equalization (CLAHE), data augmentation, and controlled minority oversampling applied exclusively to the training set. Gradient-weighted Class Activation Maps (Grad-CAM) are used to provide visual explanations of model decisions, supporting clinical interpretability. The suggested model has a test accuracy of 98.30% and an F1-score of 98.17%, exceeding competitor CNN-based techniques.
Recommended Citation
Abdalla, M. A., & Sayed, G. I. (2026, July). A Vision Transformer-Based Model for Skin Disease Classification. In 2026 Intelligent Methods, Systems, and Applications (IMSA) (pp. 318-323). IEEE. https://doi.org/10.1109/IMSA70415.2026.11700304