Data Augmentation Improves OOD Generalization in Dermoscopic Skin Cancer Classification

A new research study on arXiv has highlighted the advantages of data augmentation in improving out-of-domain (OOD) generalization in skin cancer classification using dermoscopy techniques. The team utilized a ConvNeXt-Large architecture and evaluated various augmentation methods on the diverse ISIC Archive dataset, which included Derm7pt. For out-of-domain testing, they incorporated the HAM10000 and ISIC 2019-2020 datasets. Notably, the composite augmentation method outperformed others, particularly with photometric adjustments. The study revealed a significant enhancement in OOD performance, with an average gain of 0.053, suggesting the potential for better classification accuracy of skin cancer types.