Spatial temporal graph convolution network for the analysis of regional wall motion in left ventricular opacification echocardiography

The analysis of regional wall motion (RWM) in echocardiography is critical for the diagnosis of coronary heart disease (CHD). Traditional detections of regional wall motion abnormalities (RWMAs) mainly rely on manual observation and experience. Despite the rapid development of deep learning analysis of medical data, current automated analyses of myocardial function are mainly quantitative measurements, limited by the accuracy of left ventricular (LV) segmentation results. We expect to develop a qualitative deep learning method, not relying excessively on quantitative measurements, to detect RWMAs and determine the infarct coronary artery, in left ventricular opacification (LVO). The architecture we proposed constructs the LV wall segments into graphs, using spatial temporal graph convolution network (ST-GCN) and Multi-head Self-attention (MSA) to extract lesion features. We adopted modified dynamic and channel-non-shared topological strategy, with specific structural topology for RWM. The three ST-GCN classifiers are used to detect RWMAs related to the three coronary artery branches respectively, and summarize results. In our experiments, using the data we collected in the clinic, we trained and evaluated three ST-GCNs with different data sets, achieving the best performance compared with several state-of-the-art models. The comprehensive accuracy, sensitivity and specificity are 81.83%, 80% and 83.67% respectively. The experimental results demonstrate the feasibility of GCN analysis for the RWM in LVO.