Research On Semi-supervised Cardiac Medical Image Segmentation Methods

Cardiac image segmentation is an important task in the field of computer vision and has significant implications for the diagnosis and treatment of heart disease.However,annotating medical images requires not only the involvement of experts but is also time-consuming and laborintensive.Deep learning-based cardiac image segmentation methods can effectively reduce the need for expert involvement but require large datasets for model training.Meanwhile,cardiac images have issues such as unclear boundaries and lower resolution due to their unique characteristics,which can affect the accuracy of cardiac image segmentation.To improve the automation and segmentation accuracy of cardiac images,this paper studies a semi-supervised method enhanced by an external attention mechanism,with the main work including:(1)By combining the external attention mechanism with the Swin Transformer model,an improved EA-Swin UNet(External Attention augmented Swin UNet)segmentation method is proposed.This method introduces the external attention mechanism into the skip connections of Swin UNet,enabling the utilization of global semantic information of cardiac images to enhance the feature correlation between global samples.It effectively solves the accuracy reduction problem caused by small sample dataset training in existing Transformer-based cardiac segmentation methods,improving the accuracy and generalization ability of cardiac segmentation.(2)To address the limitation of the teacher model in the Mean Teacher semi-supervised cardiac segmentation framework,which only provides single supervision for the student model,this paper proposes an improved cross-pseudo-supervised cardiac image segmentation method.Based on EA-Swin UNet,two networks with the same structure but different initial weights are designed,with the output results mutually supervising and cross-promoting the training of the other network.This method further improves the automation and efficiency of cardiac semi-supervised segmentation by integrating consistency learning and self-training,two semi-supervised learning strategies.Experiments show that the two semi-supervised cardiac image segmentation methods proposed in this paper—EA-Swin UNet enhanced by an external attention mechanism and the improved cross pseudo supervision segmentation network—achieve significant improvements on the MICCAI2017 ACDC dataset.Both region segmentation accuracy and boundary clarity are significantly improved.Compared with the existing baseline,the Dice metric increases by nearly 2%,while the 95% HD metric reaches 3.52.