Bias correction medical image segmentation method based on multiple disturbance

Abstract Semi-supervised Learning (SSL) has shown significant advantages in using large-scale unlabeled data, particularly in medical image segmentation tasks. However, model perturbation may occur during the training process, which affects the segmentation outcome of the model, especially for edge regions. In order to solve this issue, we construct a multi-perturbed two-student correction network (MPDCN), which is based on a mean-teacher network and designed with a two-student network structure. To enhance the training data, multiple data perturbations, including feature-level perturbations and image-level perturbations, are used to obtain more diverse output features. Meanwhile, the bias between different output predictions to correct the training of the model by learning potential prediction bias from prediction difference features between outputs. The experimental outcomes show that MPDCN achieves prominent performance improvement over the baseline model and outperforms SS-Net, MC-Net, and MCF on multiple datasets, proving its effectiveness and wide applicability. MPDCN provides a new perspective for the application of SSL in medical image segmentation.