A CMR Short-Axis Images Segmentation Method based on Multi-Attention Mechanism and Boundary Distance Map

Heart disease is a common illness, and currently, the most commonly used technique for diagnosing it is cardiac magnetic resonance (CMR) imaging. CMR semantic segmentation has problems such as poor segmentation performance and blurred edges. To address the problem of poor semantic segmentation of CMR short-axis images, a heart structure segmentation and post-processing method based on multiple attention mechanisms and boundary distance map is proposed. By using an image segmentation network based on multiple attention mechanisms, the pixel-level classification of different cardiac structures was basically achieved. Meanwhile, to address the problem of poor prediction results for the heart base, regression prediction was performed on the boundary distance maps of the left ventricle, left ventricular myocardium, and right ventricle to complete the post-processing task of cardiac segmentation images and further improve the accuracy of CMR short-axis image semantic segmentation. Experimental results show that the proposed method performs well in comparison with similar methods in Dice and HD metrics on both the ACDC public dataset and our private dataset; The proposed post-processing method achieves good results in optimizing the edges of cardiac segmentation images.