Classification of 24 skin conditions using swin transformer: leveraging DermNet & healthy skin dataset

作者：Jameel Ahmad, Muhammad Umar Farooq, Fouqia Zafeer, Usama, Nauman Shahid · 发表于：IET conference proceedings. · 年份：2025 · DOI：10.1049/icp.2025.1160 · 被引用次数：8 · 研究领域：Cutaneous Melanoma Detection and Management

Skin disease classification is a pivotal task in dermatological diagnostics, often hindered by the visual similarities betwee n conditions and diverse disease presentations. Traditional convolutional neural network (CNN) models face limitations in capturing global dependencies and multi-scale features, leading to challenges in achieving high classification accuracy. To address these issues, this study leverages the Swin Transformer, a hierarchical vision Transformer architecture, to classify 24 skin conditions, including 23 diseased skin classes from the DermNet dataset and an additional healthy skin class assembled from multiple datasets. By incorporating a healthy skin class, the model addresses a critical real-world scenario, ensuring the accurate differentiation of healthy skin from diseased conditions. Unlike prior studies that merge datasets or classify subse ts of DermNet classes, our approach focuses on comprehensive coverage of all DermNet super-classes. It adds a healthy skin category, enhancing practicality and diagnostic reliability. The Swin Transformer overcomes CNN limitations by utilizing localized self-attention mechanisms and multi-scale feature extraction, achieving an accuracy, precision, recall, and F1 score of 97%. These results establish the effectiveness of Swin Transformers for real-world dermatological applications and highlight their potential as a robust framework for medical image analysis.