Research On Cardiac MRI Segmentation Algorithm Based On Convolution And Self-attention

Improving the accuracy of early diagnosis of the heart can greatly reduce the mortality of patients with cardiovascular diseases.Meanwhile,because manual segmentation requires a lot of manpower,it is particularly important to study accurate and fast automatic heart segmentation algorithms.In this paper,a method based on deep convolution and self-attention mechanism is used to realize automatic cardiac MRI image segmentation.The main research content of this paper is as follows:(1)Aiming at the problem of poor edge segmentation of the heart,a multi-scale cross self-attention transformer network(CST-Unet)is proposed based on the U-shaped medical image segmentation framework.The CST-Unet model effectively combines the convolutional window sliding and the long-distance dependence of self-attention;and proposes a cross attention skip module(AS)in the skip connection to bridge the information gap between the encoder and decoder.The network can effectively enhance the global extraction ability of the model,and the proposed module can increase the attention to the required target features,and suppress the influence of background and other irrelevant factors on the cardiac image segmentation algorithm.Finally,the model achieves 90.44% on the average DSC metric,and 1.50 mm and 2.86 mm on the edge segmentation metrics ASSD and HD,respectively.(2)For the problem of low segmentation accuracy of the myocardium and the large consumption of computing resources of the model,this paper combines the structure of CST-Unet to propose a depthwise separable convolutional network(DS-UNe Xt).In the DS-UNe Xt model,a parallel depthwise separable spatial pooling module and a lightweight decoder are proposed;and depthwise separable convolution is deeply integrated in the overall structure.The model deeply combines the inherent locality of convolution and the excellent global information learning ability of the self-attention mechanism.The proposed parallel depthwise separable spatial pooling module enhances advanced semantic features.The DS-UNe Xt model is lightweight and has less parameters and computational complexity.Finally,the proposed model achieved an average DSC index of 90.89%,an edge segmentation index of ASSD and HD of 1.21 mm and 2.61 mm,respectively,and the model parameters and computational complexity were reduced to10.9M and 2.6G,respectively.(3)The two proposed models were analyzed experimentally on the ACDC dataset.The experimental results show that the proposed algorithm can effectively alleviate the problem of poor segmentation of the cardiac edge,and greatly improve the difficulty of segmentation in cardiac MRI.The segmentation effect of the top slice can more accurately segment the left ventricle,right ventricle,and myocardial structure of the heart,and can also alleviate the problems caused by the consumption of computing resources.