Deep Learning vs manual techniques for assessing left ventricular ejection fraction in 2D echocardiography: validation against CMR

Accurate and reproducible echocardiographic assessment of left ventricular (LV) volumes and ejection fraction (EF) is crucial in clinical decision-making and risk stratification (1–6). Hence LVEF thresholds are used for decision making in heart failure (7) and coronary artery (8) and valvular heart (9) diseases. Simpson’s method from two-dimensional (2D) 2- and 4-chamber views is currently the preferred approach for evaluation of LV volumes and EF by echocardiography. Yet it is time consuming, subject to wide interobserver variability (10,11), and the reproducibility is highly affected by various factors such as operator experience and image quality. Accordingly, 2D echocardiography has shown to be less reproducible than cardiac magnetic resonance (CMR)(12), which is currently considered the reference standard for evaluation of LVEF and volumes. This approach however suffers from higher costs and less frequent availability. Deep learning (DL) allows automated contour detection offering the promise of faster and potentially more accurate and reproducible evaluation of LV volumes and EF by echocardiography. In this work, we developed a new DL algorithm for manual contouring based on a U-Net convolutional network architecture, using an anonymised database of echocardiographic images. The aim of the present study was to evaluate the generalizability and accuracy of this new algorithm relative to manual contouring and against cardiac MR volumes and EF as a reference. We evaluated our algorithm using a set of multimodal data from 171 subjects from two centres. We also compared the automated contouring and resulting LV volumes to their manual counterparts, as obtained by different junior and senior observers across the two centres and used CMR as an independent 3D modality to evaluate differences in bias between manual and automated contouring. Finally, we benchmarked our DL algorithm against two other DL implementations from the literature and made the 2DE database, CMR LVEF values and the Simpson bi-plane code publicly available for reproducibility of this paper and further benchmarks.