A Low-Cost High-Performance Data Augmentation for Deep Learning-Based Skin Lesion Classification.

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Computational Dermatology

Background:

• Intelligent skin cancer screening requires high-performance, low-cost data augmentation, especially for rural deployment.
• Current methods may not be efficient or accessible in resource-limited settings.

Purpose of the Study:

• To develop a high-performance, low-cost data augmentation strategy for AI-based skin cancer screening.
• To improve classification performance and highlight regions of interest for clinicians.
• To enable early screening and diagnosis for various diseases in low-resource settings.

Main Methods:

• A plug-and-play data augmentation strategy with a search space of 10^1 was proposed.
• The strategy was tested with EfficientNets as a baseline on medical databases.
• Grad-CAM++ was used to generate heatmaps for model interpretability.

Main Results:

• Achieved best BACC of 0.853 on HAM10000, outperforming existing single-model approaches.
• Obtained average AUC of 0.909 on ISIC 2017, surpassing ensembling and external database models.
• Reached best BACC of 0.735 on Derm7pt, exceeding all related studies. Grad-CAM++ heatmaps confirmed accurate feature selection.

Conclusions:

• The proposed data augmentation strategy significantly reduces computational costs for intelligent skin lesion diagnosis.
• This approach supports the development of cost-effective, portable AI devices for skin cancer screening and treatment guidance.
• The strategy has broad applicability for early disease screening and diagnosis in diverse clinical settings.