The accurate segmentation of myocardial structure is crucial for heart disease diagnosis. With the rapid advancements in computer-aided diagnosis, deep learning has surfaced as a promising instrument for segmenting cardiac ultrasound images. Nevertheless, the limited imaging quality and structural complexity of these images often hinder achieving optimal segmentation accuracy, particularly in edge regions. To address this issue, this paper innovatively proposes a deep learning segmentation method based on attention mechanism. The method adopts a cascade network structure, first using a deep convolutional neural network to obtain feature maps and preliminary prediction results. Subsequently, by introducing a difference analysis module, it deeply analyzes the difference information between the preliminary prediction results and the rough labels generated by ordinary network. Then, using attention mechanism, it enhances the feature vectors generated in the preliminary prediction stage, and generates more accurate prediction labels on this basis. To verily the effectiveness of this method, experiments were conducted on a private myocardial dataset and achieved remarkable results. It is anticipated that this approach will offer robust assistance in achieving precise segmentation of cardiac ultrasound images, subsequently furnishing crucial technical backing for the auxiliary diagnosis of cardiac diseases.