Learning Geometric Information Propagation for Semi-supervised 3D Medical Image Segmentation

Semi-supervised learning enhances medical image segmentation by leveraging unlabeled data, reducing reliance on extensive labeled datasets. We observe that medical images usually exhibit consistent anatomical structures defined by geometric features, while most existing methods fail to fully utilize global geometric priors. Meanwhile, the distribution discrepancy between limited labeled data and abundant unlabeled data hinders model generalization, and we find that transferring information about regions where the semantics of labeled and unlabeled data are similar can alleviate it. Inspired by the above two observations, we propose a novel semi-supervised medical image segmentation method incorporating global geometry attention and information propagation (GGIP). Firstly, we design a geometric moment attention mechanism to extract richer global geometric features. Secondly, we introduce a global geometric perturbation consistency strategy to allow the model to perceive the geometric deformation caused by individual differences in different medical data. Thirdly, based on Mamba which is a selective linear state-space propagation model for efficient sequence modeling, we design a global information propagation Mamba module to reduce the distribution discrepancy between labeled and unlabeled data. Our method achieves the state-of-the-art (SOTA) performance on a series of datasets, including the NIH pancreas, left atrium, and brain tumor datasets. Code is available at https://github.com/Lianyuan-Yu/GGIP .