Myocardial Segmentation of 2D Echocardiography Based on the Improved Self-Attention Mechanism

The measurement of different clinical indicators, such as the myocardial wall's thickness and the ventricle's volume, depends on accurately segmenting the left ventricle in 2D echocardiography. This segmentation is essential for making various clinical decisions, such as whether surgery is necessary. However, segmenting the myocardium requires clinical expertise and makes accurate segmenting time-consuming due to the poor quality of echocardiograms. It is essential to develop an automatic myocardial segmentation method. Some deep learning methods, especially the U-net structure, were proposed to support the automatic myocardium segmentation in echocardiography. The model must simultaneously capture the local and global dependency relationships in the ultrasound images for myocardial segmentation in echocardiography. Given the inherent attribute of local operation of convolution operations, this poses significant limitations for U-Net based on the traditional CNN as the basic architecture. For example, when dealing with complex myocardial textures and structures, we need to accurately identify the myocardium's global morphological features and overall structural patterns. Moreover, this locality may lead to deviations and errors when segmenting the boundary regions or the myocardial parts closely related to the surrounding tissues. Our primary focus will be redefining the skip connection using the attention mechanism. More integrated attention mechanisms can further enhance the segmentation effect of myocardial in echocardiography. Therefore, we integrated the cross-attention mechanism when concatenating features in the skip connection. Through conducting rigorous ablation studies on multiple echocardiography segmentation datasets, the performance of the proposed mechanism was proven.