ScholarWorks@동국대학교

초록

Accurate classification of skin lesions is essential for early diagnosis and treatment planning. However, severe class imbalance in dermatological datasets hinders the effective training of multi-class classification models. To address this challenge, we propose an end-to-end framework combining dual contrastive learning with segmentation-guided attention. Our model uses a ResNet18-based U-Net encoder, pretrained with Self-Supervised Contrastive Learning (SSCL) and Supervised Contrastive Learning (SCL). The U-Net decoder generates a spatial attention map that leverages segmentation information to identify lesion boundaries. This segmentation-guided attention map is element-wise multiplied with the original image to create lesion-focused input for classification. This enhanced input is then processed by a classification head for final diagnosis. Evaluated on SLICE-3D and HAM10000 datasets, the proposed method achieved 72.19% accuracy, 72.96% weighted F1-score, and 88.39% macro AUC. Ablation studies confirm the effectiveness of both segmentation and attention, as well as the synergy of the dual contrastive strategy. The framework demonstrates robust and balanced performance, making it clinically applicable for the skin lesion classification. © 2025 IEEE.

키워드