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

Objective and Impact Statement . There is a need to develop high-performance low-cost data augmentation strategies for intelligent skin cancer screening devices that can be deployed in rural or underdeveloped communities. The proposed strategy not only improve the classification performance of lesions but also highlight potential regions interest clinicians’ attention. This implemented broad range clinical disciplines early automatic diagnosis many other diseases low resource settings. Methods We propose search space 10 1 , which combined with any model through plug-and-play mode best argumentation method medical database cost. Results With EfficientNets as baseline, BACC HAM10000 0.853, outperforming published models “single-model no-external-database” ISIC 2018 Lesion Diagnosis Challenge (Task 3). average AUC on 2017 achieves 0.909 (±0.015), exceeding most ensembling those using external datasets. Performance Derm7pt archives 0.735 (±0.018) ahead all related studies. Moreover, model-based heatmaps generated by Grad-CAM++ verify accurate selection lesion features judgment, further proving scientific rationality diagnosis. Conclusion greatly reduces computational cost clinically lesions. It may facilitate research low-cost, portable, AI-based mobile therapeutic guidance.