UPMatch: Enhancing Semi-Supervised Medical Image Classification through Contrastive Learning with Unreliable Pseudo Labels

Pseudo-labeling based semi-supervised learning (SSL) framework has proven highly successful in medical image analysis (MIA) by addressing the problem of a shortage of labeled samples. However, the existing SSL methods use a fixed or flexible confidence threshold to filter reliable samples, leaving large number of unlabeled samples unused. This is a more serious issue in MIA because of the low inter-class distance and imbalanced categories. We argue that effectively mining useful information hidden in ambiguous unlabeled sample is the key to improve model performance, so we propose UPmatch, a new pseudo labeling-based SSL framework. Our framework introduces a contrastive unreliable pseudo label learning module (CUPM) that incorporates unreliable pseudo label samples into the training process. Additionally, we propose an informative sample selection strategy (ISSS) that selects samples used in contrastive learning iteratively in each mini-batch. Our experiments on TissueMNIST and ISIC2019 dataset with various training settings demonstrate the effectiveness of our proposed strategy.