FixMatch-LS: Semi-supervised skin lesion classification with label smoothing

Recent research on semi-supervised learning (SSL) is mainly based on the method of consistency regularization, which relies on data augmentation in the specific domain. Pseudolabelling is a more general method that has no such restrictions but is limited by noisy training. Medical datasets are a particular domain that exhibit a long-tail distribution. Combining these two limitations, we focus on the widespread use of weak augmentation to generate pseudolabels. We propose FixMatch-LS and a variant FixMatch-LS-v2 for medical image classification. First, we introduce label smoothing to change the pseudolabel threshold, which reduces the influence of noisy pseudolabels. In addition, pseudolabelling should be matched with consistency. A suitable consistency can constrain pseudolabelling to improve the quality of the pseudolabels. We validate our framework on skin lesion diagnoses from the ISIC 2018 and ISIC 2019 challenges, obtaining AUCs of 91.63%, 93.70%, 94.46%, and 95.44% on the four proportions of labelled data from ISIC 2018.

Automated summary

Recent research in semi-supervised learning focuses on consistency regularization using data augmentation, while the more general method of pseudolabelling is limited by noisy training. Medical datasets have a long-tail distribution, and combining these limitations, we propose FixMatch-LS and its variant FixMatch-LS-v2 for medical image classification. We introduce label smoothing to adjust the pseudolabel threshold and reduce the influence of noisy pseudolabels, and emphasize the importance of consistency for pseudolabelling to improve pseudolabel quality. The framework is validated on skin lesion diagnoses from the ISIC 2018 and ISIC 2019 challenges, achieving AUCs of 91.63%, 93.70%, 94.46%, and 95.44% on different proportions of labelled data from ISIC 2018.