Segment-Aware Contrastive Representation Learning With Vision Transformers: TransCon-Skin

Abstract

Skin cancer is one of the most common malignant diseases worldwide, and its early and accurate diagnosis is critical for the prognosis and treatment success. In this study, we propose TransCon-Skin, a new deep learning model based on segmentation and contrastive learning for high-accuracy classification of dermoscopic images. The model provides an effective learning structure that strengthens class discrimination by optimizing the representations extracted with the Vision Transformer (ViT) architecture with the MoCo framework. In the experiments, TransCon-Skin demonstrated outstanding classification performance, achieving 99.79% accuracy, 99.89% F1-score, and 100% recall in all ViT configurations. Furthermore, classification times of 1.5 to 4.9 milliseconds demonstrate that the model is not only highly accurate but also fast and efficient, making it suitable for integration into real-time systems. These results demonstrate that the TransCon-Skin model offers a reliable, scalable, and clinically applicable approach to skin cancer diagnosis.