RSSL: A Unified Reliability-Aware Semi-Supervised Learning Framework for Medical Image Classification

Semi-supervised learning (SSL) is emerging as the dominant paradigm in medical image analysis, which can leverage abundant unlabeled data under the assumption of the same distribution with labeled data. However, out-of-distribution (OOD) unlabeled data are ubiquitous in clinical practice, often originating from diverse populations, devices, and settings. Most of current SSL methods leverage unlabeled data in a uniform way, which deteriorates the model performance caused by: 1) the unreliable knowledge derived from OOD unlabeled data, and 2) the confirmation bias induced by unreliable pseudo labels. In this work, we propose a unified reliability-aware SSL framework (RSSL), which integrates automatic identification of reliable unlabeled data with pseudo-label self-refinement schemes to tackle the above challenges. Specifically, we introduce a weight network to automatically assess the reliability of unlabeled data. This allows us to capture the reliable knowledge derived from the unlabeled data and safely blend it into the downstream tasks. Additionally, we introduce a pseudo-label self-refinement scheme to alleviate the negative effect of confirmation bias on model training. The synergistic learning between reliable unlabeled data identification and pseudo-label self-refinement during the training process enhances the robustness and generalizability of the model. Extensive experiments conducted on three public medical image datasets demonstrate the superior performance of the proposed method.