Confidence-Guided Online Knowledge Distillation for Semi-supervised Medical Image Classification

In medical image analysis, semi-supervised learning (SSL) classification algorithms are crucial due to the time-consuming and expensive nature of acquiring annotated medical images compared to readily available unlabeled images. In this paper, we propose confidence-guided online knowledge distillation for SSL to enhance model performance through a single-stage reliable mutual learning process. Specifically, we set two SSL models to focus on pseudo-labeling and consistency learning, respectively. Leveraging online knowledge distillation, we orchestrate collaborative learning between two models, facilitating mutual enhancement. In addition, we introduce dynamic model confidence to guide the independent and mutual learning of these models, encouraging the models to learn more reliable information from unlabeled data. Extensive experiments on publicly available datasets for blood cell and skin disease classification show significant performance improvements with our proposed framework.