Improved Strain Computation for Transesophageal Echocardiography Acquisitions

Cardiac surgeries are major interventions prone to serious complications that can occur during and after the operation. Monitoring the heart allows to anticipate such complications and has thus become a standard practice to perform through the perioperative period. Echocardiographic assessment of the myocardial contractility is a common monitoring procedure that is generally performed by visual inspection. The qualitative nature of this technique makes it highly vulnerable to the operator's subjectivity and thus drove cardiologists to develop standardized quantitative measures of cardiac function, such as strain. Strain estimation still requires manual annotation of images and still suffers from inter- and intra-observer variability.
This thesis is presented as a contribution towards the complete automatization of the strain estimation task in transesophageal echocardiographic (TEE) images. A novel strain estimation pipeline is proposed. It focuses on the estimation of the longitudinal strain in the basal segments of the 4-chamber, 2-chamber and apical long-axis views of the heart. This pipeline uses the segmentation model U-Net and a custom thinning algorithm to automatically extract myocardial points from the first frame of a B-mode sequence and estimates their motion with optical flow methods. Strain is then computed based on the estimated displacement of these points through cardiac cycles. Four optical flow models are experimented, among which two have a convolutional neural network-based architecture. Integration of tissue velocity imaging (TVI) data and a novel tracking method based on Kalman filtering are developed in order to improve the motion estimation and tracking processes.
U-Net and the two CNN-based models are trained on B-mode recordings from 70 patients. A test set of 18 patients is used to evaluate the tracking and strain estimation performances of the different models. The myocardial point extraction algorithm gives usable results in 50% and 57% of the cases when applied to high and low frame rate B-mode sequences respectively. Three optical flow algorithms present outstanding tracking performances in five of the six basal segments. It is shown that exploiting TVI data improves tracking performances. The same observation is made when the Kalman filtering-based tracking method is applied to high frame rate sequences. The proposed techniques achieve state-of-the-art strain estimation performances. A mean absolute error of (2.74 + 2.38)% is achieved in the inferoseptal segment. The inferior and anterior segments are the segments in which the correlation between strain estimates and ground truth values is the highest: the Pearson correlation coefficient reaches the value 0.77 in the inferior segment and 0.79 in the anterior segment.