Development of Automated Neural Network Prediction for Echocardiographic Left ventricular Ejection Fraction

TL;DR

The automated neural network-based calculation of LVEF is comparable to expert clinicians performing time-consuming, frame-by-frame manual evaluations of cardiac systolic function.

Abstract

The echocardiographic measurement of left ventricular ejection fraction (LVEF) is fundamental to the diagnosis and classification of patients with heart failure (HF). In order to quantify LVEF automatically and accurately, this paper proposes a new pipeline method based on deep neural networks and ensemble learning. Within the pipeline, an Atrous Convolutional Neural Network (ACNN) was first trained to segment the left ventricle (LV), before employing the area-length formulation based on the ellipsoid single-plane model to calculate LVEF values. This formulation required inputs of LV area, derived from segmentation using an improved Jeffrey's method, as well as LV length, derived from a novel ensemble learning model. To further improve the pipeline's accuracy, an automated peak detection algorithm was used to identify end-diastolic and end-systolic frames, avoiding issues with human error. Subsequently, single-beat LVEF values were averaged across all cardiac cycles to obtain the final LVEF. This method was developed and internally validated in an open-source dataset containing 10,030 echocardiograms. The Pearson's correlation coefficient was 0.83 for LVEF prediction compared to expert human analysis (p<0.001), with a subsequent area under the receiver operator curve (AUROC) of 0.98 (95% confidence interval 0.97 to 0.99) for categorisation of HF with reduced ejection (HFrEF; LVEF<40%). In an external dataset with 200 echocardiograms, this method achieved an AUC of 0.90 (95% confidence interval 0.88 to 0.91) for HFrEF assessment. This study demonstrates that an automated neural network-based calculation of LVEF is comparable to expert clinicians performing time-consuming, frame-by-frame manual evaluation of cardiac systolic function.

Figures (8)

• Figure 1: (a), (b) and (c) were from the Stanford dataset; (d) was from the CAMUS dataset. (a) presented human labelled coordinate points in one frame. A Euclidean distance between two pink points was the LV length; (b) shown the mask generated from these coordinate points, which was used for training our segmentation network; (c) illustrated the LV area, LV widths, LV heights and LV length; (d) presented annotations information including the left ventricle endocardium, the left ventricle myocardium and the left atrium.
• Figure 2: (a) Flow chart of the pipeline. There were three main steps, including LV segmentation, LVEF calculation and HFrEF assessment. The area information from segmentation could also be used for ED and ES identification, beat-to-beat analysis of the heart, as well as visualising changes in volume (for example due to an arrhythmia such as atrial fibrillation). ED $=$ end diastole; ES $=$ end systole; HFrEF$=$ Heart failure with reduced LVEF; LV $=$ left ventricle; LVEF$=$ left ventricular ejection fraction. (b) was the input of the pipeline. (c) was the proposed AI system. (d) was outputs information, including the segmentation result and the beat-to-beat visualizer. The calculated LVEF values was presented in this visualizer along with the result of the HF phenotype classification. (e) was the outcome.
• Figure 3: Overall segmentation architecture. The segmentation network combined ResNet-50 (a), atrous convolutions, and atrous spatial pyramid pooling (ASPP) (b) to resample features at different scales and to capture multi-scale information. As an example, p0, r2, and s1 in the figure denotes padding = 0, atrous convolution with rate = 2, and stride = 1, respectively.
• Figure 4: Ensemble learning model: including Extra Tree (ET), AdaBoosting (AD), Lasso and a stacking algorithm combining Ridge, K-Nearest Neighbors (KNNs) and Gradient Boosting Decision Tree (GBDT). The predicted LV lengths from these regressors were finally ensembled by a voting mechanism.
• Figure 5: (a) showed three scenarios used for selecting true peaks, which are identified as ED and ES phases. (b) was the improved Jeffrey’s method used to fine tune LV areas computed from segmentation. Here, three parts were averaged to compute final LV areas at ED or ES.