A Self-Supervised Vision Transformer Approach for Dermatological Image Analysis

Main Article Content

Fan Guo
Xuan Wu
Lu Zhang
Hao Liu
Anda Kai

Abstract

This paper proposes a novel skin disease classification method, which combines visual transformers (ViT) with self-supervised learning and is validated on the ISIC 2018 dataset. Experimental results show that the method can effectively extract feature representations through feature pre-training through self-supervised contrastive learning, and shows significant advantages over traditional convolutional neural network (CNN) models in multiple indicators such as AUC, precision, recall and F1 score. T-SNE visualization further reveals the obvious clustering characteristics in the feature space, confirming the superiority of the method. Compared with existing mainstream technologies, this method shows higher discrimination ability in handling complex skin lesion classification tasks, while reducing the dependence on large-scale labeled data and enhancing the practicality of the model. Analysis of the loss function curve shows that the method achieves fast and stable convergence during training, highlighting its efficiency and stability. This study verifies the potential of ViT in medical image analysis and provides an efficient solution for the automatic classification of skin diseases.

Article Details

How to Cite
Guo, F., Wu, X., Zhang, L., Liu, H., & Kai, A. (2025). A Self-Supervised Vision Transformer Approach for Dermatological Image Analysis. Journal of Computer Science and Software Applications, 5(4). https://doi.org/10.5281/zenodo.15165057
Section
Articles