Adaptive class rebalancing and spatial-aware metadata fusion for long-tailed skin lesion classification

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning models for dermoscopic skin lesion classification face challenges due to imbalanced datasets and limited use of clinical metadata.
• This imbalance and underutilization reduce the diagnostic reliability of current systems.

Purpose of the Study:

• To improve the diagnostic reliability of deep learning systems for skin lesion classification.
• To address class imbalance and effectively integrate patient metadata into classification models.

Main Methods:

• A Vision Transformer baseline was enhanced with three key improvements: class-balanced focal loss, a patch-level cross-attention metadata-guided attention module, and supervised contrastive regularization.
• The model fuses dermoscopic images with patient metadata using cross-modal attention.
• Evaluations were conducted on the ISIC 2019 and BCN 20000 datasets.

Main Results:

• Ablation studies confirmed consistent performance improvements with the proposed enhancements.
• The full model achieved a Macro AUC of 0.9613 and a Macro balanced accuracy of 0.8426 on the ISIC 2019 validation set.
• The model demonstrated strong performance on rare categories, including dermatofibroma (AUC = 0.991) and vascular lesions (AUC = 0.999).

Conclusions:

• The combination of principled loss design, spatially aware fusion, and representation-level regularization significantly enhances the robustness of skin lesion classification.
• These methods effectively address the challenge of extreme class imbalance in dermoscopic datasets.
• The findings contribute to more reliable AI-driven diagnostic tools in dermatology.