A Deep Bayesian Video Analysis Framework: Towards a More Robust Estimation of Ejection Fraction

Ejection Fraction (EF) is a widely-used and critical index of cardiac health. EF measures the efficacy of the cyclic contraction of the ventricles and the outward pumpage of blood through the arteries. Timely and robust evaluation of EF is essential, as reduced EF indicates dysfunction in blood delivery during the ventricular systole, and is associated with a number of cardiac and non-cardiac risk factors and mortality-related outcomes. Automated reliable EF estimation in echocardiography (echo) has proven challenging due to low and variable image quality, and limited amounts of data for training data-driven algorithms which delays the integration of the technologies in the clinical workflow. In this paper, we introduce a Bayesian learning framework for automated EF assessment in echo videos. Our key contribution is to automatically estimate the epistemic uncertainty, i.e. the model uncertainty, in EF estimation. We anticipate that such information about uncertainty can be incorporated in clinical decision making. We use a ResNet18-based (2 + 1)D as the baseline architecture for video analysis and provide its side-by-side comparison of our probabilistic approach using public data from 10,031 echo exams. Our results clearly indicate the superior performance of the Bayesian model in the clinically critical lower EF population.