Semi-supervised segmentation method based on the contrastive learning for medical image with noisy labels

With the continuous advancement of automated analysis technologies in medical imaging, semi-supervised learning methods have gained widespread attention due to their effectiveness in label-scarce environments. However, manually annotated medical image labels often contain noise, which significantly impairs the performance of semi-supervised models. To address this issue, this paper introduces a novel semi-supervised medical image segmentation model specifically designed to tackle the problem of noisy labels in medical image segmentation tasks. The main contributions of this paper lie in the design of two core modules: the Label Filtering Module with Gaussian Mixture Model (LFG) and the Semi-supervised Learning Module for Noisy Samples (SLNS). The LFG module collaborates through two Mean-Teacher networks with different parameters, utilizing two 2D-GMMs to fit the Dice similarity loss between network predictions and original labels, effectively distinguishing between clean and noisy labels. The SLNS module enhances feature representation learning by introducing contrastive loss and optimizes the training process by calculating correction and consistency losses, thereby improving the model's stability and accuracy. Experimental results demonstrate that when noise rate is 0.4 and labeled data percentage is 20%, the model achieves an average Dice loss of 81.24% on the ACDC cardiac image dataset, while the average noise rate on the JSRT lung image dataset reaches 78.39%, significantly outperforming other State-Of-The-Art (SOTA) methods. This validates the model's effectiveness in enhancing segmentation performance in noisy environments.