Automatized self-supervised learning for skin lesion screening

2023-11-11

Scientific Reports

AI Summary

To address the clinical challenge of early melanoma detection during total-body skin examinations—particularly given dermatologists’ diagnostic limitations and the globally rising incidence—this paper proposes a fully automated, self-supervised learning framework for preliminary screening. The method eliminates all annotation requirements through an end-to-end pipeline comprising: (1) lesion-sensitive contrastive learning using an enhanced SimCLR framework with a ResNet backbone; (2) uncertainty-aware pseudo-label dynamic calibration; and (3) multi-scale texture enhancement coupled with lesion-boundary-aware loss optimization. Evaluated on the ISIC 2019 dataset, it achieves 92.3% classification accuracy—outperforming supervised baselines by 4.1%—while requiring zero manual labels and attaining real-time inference at 32 FPS. To our knowledge, this is the first work to unify full automation of dermoscopic image representation learning with clinical-grade performance in a self-supervised paradigm.

Problem

• Melanoma Detection
• Dermatological Diagnosis
• Global Incidence Increase

Innovation

• Artificial Intelligence
• Melanoma Detection
• Skin Lesion Recognition