Enhancing the impact of model performance gains for semi-supervised medical image segmentation

Semi-supervised methods aim to alleviate the high cost of annotating medical images by incorporating unlabeled data into the training set. Recently, various consistency regularization methods based on the mean-teacher model have emerged. However, their performance is limited by the small number and poor quality of confident pixels in the pseudo-labels. Based on experimental observations, we propose a new argument: the performance gains of the model do not proportionally translate into improvements in pseudo-label quality, mainly due to constraints in pixel diversity representation and model expressiveness. Therefore, we propose a novel semi-supervised framework, DOC-MLE, which consists of two key components: a dynamic orthogonal constraint (DyOrCon) method and one multi-level election (MLElect) strategy. Specifically, DyOrCon imposes orthogonal constraints on multiple intermediate projection heads to enhance pixel diversity and fully exploit the model’s potential representation capacity. MLElect is designed considering both unsupervised pixel-level and supervised feature-level strategies, to generate reliable pseudo-labels. Moreover, to generate more robust prototype representations, this paper proposes new threshold filtering, edge erosion, and dynamic convolution strategies to address errors associated with low-confidence, high-confidence, and local morphological constraints. Extensive experiments on coronary angiography, polyp dataset, and retinal fundus images have proven the effectiveness of the proposed method.