Semi-supervised medical image classification with adaptive threshold pseudo-labeling and unreliable sample contrastive loss

Semi-supervised learning (SSL) may employ unlabeled data to improve model performance, which has great significance in medical imaging tasks. However, pseudo-labeling-based semi-supervised approaches suffer from two problems in medical image datasets: (1) the models' predictions are biased toward the majority class in imbalanced datasets, and (2) discarding unlabeled data with confidence below the thresholds results in the loss of useful information. To solve these issues, we propose a novel SSL framework, FullMatch, which improves the model's performance by utilizing all unlabeled data. Specifically, we propose adaptive threshold pseudo-labeling (ATPL), a method for generating pseudo-labels based on the model's current learning status. ATPL dynamically adjusts the thresholds for each class during the training process, which can generate more pseudo-labels for classes with learning difficulties, thus alleviating the problem of data imbalance. Unlike existing semi-supervised methods based on pseudo-labeling, we do not discard unlabeled data with confidence below the thresholds. We propose an unreliable sample contrastive loss (USCL) to leverage useful information from unlabeled data with confidence below the thresholds by learning the similarities and differences between sample features. To evaluate the performance of the proposed method, we conducted experiments on the ISIC 2018 skin lesion classification dataset and the blood cell classification dataset. The experimental results show that our method outperforms the state-of-the-art SSL methods.