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

Shaofeng Zhou, Shenwei Tian, Long Yu, Weidong Wu, Dezhi Zhang, Zhen Peng, Zhicheng Zhou, Junwen Wang

Abstract: Highlights•We propose a novel semisupervised medical image classification framework that integrates four loss functions, including fully-supervised classification loss, consistency regularization, the sample relation consistency paradigm (SRC), and pseudolabelled loss. The framework makes better use of unlabelled image features to improve classification performance.•Combining label smoothing with pseudolabelling, we alleviate the effect of noisy artificial labels with consistency regularization and relieve the imbalance of various categories of data by generating pseudolabels.•We observe the following:• The model performs better using KL loss than MSE loss when trained with fewer labelled samples.• The model performs better using KL loss than MSE loss when trained with more pseudolabells (and similarly more noisy labels), which is consistent with the results of [1].• Pseudolabeling needs to be matched with consistency. Inappropriate consistency will cause the model to underutilize the unlabelled data.