Abstract: Heart disease is one of the main causes of death worldwide,which seriously threatens human life and health.As a non-invasive medical imaging technology,cardiac magnetic resonance imaging(MRI) is widely used in clinical diagnosis,helping doctors accurately and efficiently diagnose and treat heart diseases.However,cardiac MRI segmentation faces great challenges in practical applications,manual segmentation methods are time-consuming and subjective,while existing fully supervised and semi-supervised segmentation methods are not effective in dealing with complex cardiac structures and pathological changes,limited by the scarcity of data sets.This study aims to solve the challenges of cardiac MRI segmentation by proposing a 3D left atrial semi-supervised segmentation framework based on global-local information fusion to address the time-consuming and subjective pro-blems of manual segmentation.Although the current fully supervised heart segmentation methods are effective,they are limited by the scarcity of data sets.The semi-supervised methods come into being,but they are still limited by the small amount of data,especially when dealing with complex cardiac structure and pathological changes.To solve this problem,this study propose a new cardiac MRI segmentation method that combines Linformer and Performer merge V-Net(LPV-Net) and 3D Enhanced Discriminator with Attention(3D-EDA) technologies to achieve an effective fusion of global-local information.The LPV-Net module,created by LinPerBlock and improved V-Net,aims to standardize the training process of the model and achieve effective fusion of global and local information,thus improving the accuracy and robustness of segmentation.In addition,we also introduced a new discriminator 3D-EDA for the specification of unlabeled data.The most critical module in the model is CARELayer,which integrates a custom attention module to enhance the ability of the model to capture important information in the feature,and the auxiliary segmentation network improves the segmentation performance.By conducting a comprehensive experiment on the left atrial dataset,comparing the proposed method with several advanced semi-supervised methods.The experimental results show that the proposed method performs well on the baseline dataset,especially when training with limited label data.For example,when training with only 10% and 20% labeled data,the Dice coefficients of 88.50％ and 90.39％ were obtained.

Key words: Semi-supervised segmentation, Global and local information, Fuse, LPV-Net, 3D-EDA, Left atrium, CARELayer