Dual-View Joint Estimation of Left Ventricular Ejection Fraction with Uncertainty Modelling in Echocardiograms

Echocardiography (echo) is a standard-of-care imaging technique for characterizing heart function and structure. Left ventricular ejection fraction (EF) is the single most commonly measured cardiac metric and a powerful prognostic indicator of cardiac events. In two-dimensional transthoracic echo, EF is measured via (1) segmentation of left ventricle on multiple cross-sectional 2D views; and/or (2) visual assessment of echo cines. However, due to high inter- and intra-observer in both approaches, robust EF estimation has proven challenging. In this paper, we propose a dual-stream multi-tasking network for segmentation-free joint estimation of both segmentation- and visual assessment-based EF, across two echo views. To account for variability in EF labels, we introduce an uncertainty modelling layer, which enables the network to inherently capture the variability in expert-annotated clinical labels, of both regression and classification types. We trained a model on 1,751 apical two- and four-chamber pairs of echo cine loops and their corresponding EF labels, and achieved an R-2 of 0.90, mean absolute error of 4.5%, and classification accuracy of 91% on a test set of 430 patients. Our proposed framework (1) requires no segmentation; (2) provides estimates for four clinical EF measurements derived from the two views; (3) recognizes the inherent uncertainties in echo measurements and encodes it; (4) provides measurements with corresponding uncertainties, which may help increase the interpretability and adoption of computer-generated clinical measurements. The proposed framework can be used as a generic approach for deriving other cardiac function parameters from echo. [GRAPHICS]