Fully automatic estimation of global left ventricular systolic function using deep learning in transoesophageal echocardiography.

Abstract

To improve monitoring of cardiac function during major surgery and intensive care, we have developed a method for fully automatic estimation of mitral annular plane systolic excursion (auto-MAPSE) using deep learning in transoesophageal echocardiography (TOE). The aim of this study was a clinical validation of auto-MAPSE in patients with heart disease. TOE recordings were collected from 185 consecutive patients without selection on image quality. Deep-learning-based auto-MAPSE was trained and optimized from 105 patient recordings. We assessed auto-MAPSE feasibility, and agreement and inter-rater reliability with manual reference in 80 patients with and without electrocardiogram (ECG) tracings. Mean processing time for auto-MAPSE was 0.3 s per cardiac cycle/view. Overall feasibility was >90% for manual MAPSE and ECG-enabled auto-MAPSE and 82% for ECG-disabled auto-MAPSE. Feasibility in at least two walls was ≥95% for all methods. Compared with manual reference, bias [95% limits of agreement (LoA)] was -0.5 [-4.0, 3.1] mm for ECG-enabled auto-MAPSE and -0.2 [-4.2, 3.6] mm for ECG-disabled auto-MAPSE. Intra-class correlation coefficient (ICC) for consistency was 0.90 and 0.88, respectively. Manual inter-observer bias [95% LoA] was -0.9 [-4.7, 3.0] mm, and ICC was 0.86. Auto-MAPSE was fast and highly feasible. Inter-rater reliability between auto-MAPSE and manual reference was good. Agreement between auto-MAPSE and manual reference did not differ from manual inter-observer agreement. As the principal advantages of deep-learning-based assessment are speed and reproducibility, auto-MAPSE has the potential to improve real-time monitoring of left ventricular function. This should be investigated in relevant clinical settings.