Study on Echocardiographic Image Segmentation Based on Attention U-Net

To interpret cardiac function through the use of echocardiography requires considerable expertise and years of diagnostic experience. To construct the support system for the evaluation of cardiac function from echocardiographic images, in this paper, we consider an automatic segmentation in a two-chamber view of echocardiographic images based on Attention U-Net. To improve accuracy, we made two ingenuity. 1) In the dataset, we merge the left ventricle as a medial constraint to its 6 parts of the left ventricular wall. 2) the weight of the corresponding loss function of each class is then set according to the area ratio of each class of echocardiography. Training and testing were performed using annotated data produced under the guidance of an echocardiographic expert.