Sammendrag
Objective: Segmental longitudinal strain (SLS) of the left ventricle is an important prognostic indicator in critically ill patients undergoing cardiac surgery and intensive care for the detection of myocardial infarctions. Unfortunately, manual strain measurements are too resource-intensive for the monitoring of critically ill patients. In the present study, we proposed a method for automatic estimation of regional LV function by predicting segmental SLS in critically ill patients by transesophageal echocardiography (TEE).
Methods: We proposed a RAFT-based neural network for optical flow estimation named EchoFlow, to estimate myocardial motion in TEE. We adopted the SIMUS simulation pipeline to generate synthetic TEE data with ground-truth references of the myocardium motion through the cardiac cycle to train EchoFlow. We trained our EchoFlow network on a subset of the synthetic data and tested EchoFlow on the remaining synthetic data. We evaluated the EchoFlow performance by providing a myocardial mesh for the first frame, warping the mesh based on the estimated dense displacement field and comparing the mesh position with ground truth references for all frames of the recording. We integrated the network into a pipeline for SLS computation.
Results: We generated realistic synthetic TEE B-mode ultrasound images of 80 patients. EchoFlow achieved an end-point-error between the estimated and reference flow field of 0.03 mm, and a mean distance of 2.3 mm between the estimated and reference myocardial meshes. Our autoStrain pipeline automatically estimated SLS on the test data. Evaluated against ground truth measures, we achieved a mean
distance (95% standard deviation) of -2.2 (-11.0 to 6.5) %.
Conclusion: We successfully created a synthetic TEE data set. Our EchoFlow network showed promising results for precise motion estimation in TEE. Our autoStrain pipeline demonstrated EchoFlow’s ability to estimate the SLS of the left ventricle in TEE.
Vis fullstendig beskrivelse
