4942Automated recognition of regional wall motion abnormalities by deep neural network interpretation of echocardiography

Background The deep neural network assisted in automated echocardiography interpretation joint to cardiologist final confirmation has now been gradually emerging. There were applications applied in echocardiography views classification, chamber size and myocardium mass evaluation, and certain disease detections already published. Our aim, instead of frame-by-frame “image-level” interpretation in previous studies, is to apply deep neural network in echocardiography temporal relationship analysis – “video-level” – and applied in automated left ventricle myocardium regional wall motion abnormalities recognition. Methods We collected all echocardiography performed in 2017, and preprocessed them into numeric arrays for matrix computations. Regional wall motion abnormalities were approved by authorized cardiologists, and processed into labels whether regional wall motion abnormalities presented in anterior, inferior, septal, or lateral walls of the left ventricle, as the ground truth. We then first developed a convolutional neural network (CNN) model to do view selection, and gathered parasternal long/short views, and apical four/two chamber views from each exam, as well as developing view prediction confidence for strict image quality control. Within these images, we annotated part of images to develop the second CNN model, known as U-net, for image segmentation and mark each regional wall. Finally, we developed the major three-dimensional CNN model with the inputs composed of four views of echocardiography videos and then output the final label for motion abnormalities in each wall. Results In total we collected 13,984 series of echocardiography, and gathered four main views with quality confidence level above 90%, which resulted in 9,323 series for training. Within these images, we annotated 2,736 frames for U-net model and resulted in dice score of segmentation 73%. With the join of segmentation model, the final three-dimensional CNN model predict regional wall motion with accuracy of 83%. Conclusions Deep neural network application in regional wall motion recognition is feasible and should mandate further investigation for promoting performance. Acknowledgement/Funding None