Frontiers in medicine · 2026
Adaptive class rebalancing and spatial-aware metadata fusion for long-tailed skin lesion classification.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Ma Tianming, Han Xianwei, Liu Guijun
Аннотация
INTRODUCTION: Dermoscopic skin lesion classification is challenged by class imbalance and the underutilization of clinical metadata, limiting the diagnostic reliability of existing deep learning systems.
METHODS: Building upon a Vision Transformer baseline that fuses dermoscopic images with patient metadata via cross-modal attention, we introduce three complementary improvements: class-balanced focal loss with square-root effective-number sampling, a patch-level cross-attention metadata-guided attention module, and supervised contrastive regularization. The extended model is evaluated on the ISIC 2019 and BCN 20000 datasets.
RESULTS: Ablation studies demonstrate consistent improvements. On the ISIC 2019 validation set, the full model achieves a Macro AUC of 0.9613 and a Macro balanced accuracy of 0.8426, with strong performance on rare categories such as dermatofibroma (AUC = 0.991) and vascular lesions (AUC = 0.999).
DISCUSSION: Principled loss design, spatially aware fusion, and representation-level regularization jointly improve the robustness of skin lesion classification under extreme class imbalance.
METHODS: Building upon a Vision Transformer baseline that fuses dermoscopic images with patient metadata via cross-modal attention, we introduce three complementary improvements: class-balanced focal loss with square-root effective-number sampling, a patch-level cross-attention metadata-guided attention module, and supervised contrastive regularization. The extended model is evaluated on the ISIC 2019 and BCN 20000 datasets.
RESULTS: Ablation studies demonstrate consistent improvements. On the ISIC 2019 validation set, the full model achieves a Macro AUC of 0.9613 and a Macro balanced accuracy of 0.8426, with strong performance on rare categories such as dermatofibroma (AUC = 0.991) and vascular lesions (AUC = 0.999).
DISCUSSION: Principled loss design, spatially aware fusion, and representation-level regularization jointly improve the robustness of skin lesion classification under extreme class imbalance.