Automated echocardiographic diastolic function grading: A hybrid multi-task deep learning and machine learning approach.

Abstract

Assessing left ventricular diastolic function (LVDF) with echocardiography as per ASE guidelines is tedious and time-consuming. The study aims to develop a fully automatic approach of this procedure by a lightweight hybrid algorithm combining deep learning (DL) and machine learning (ML). The model features multi-modality input and multi-task output, measuring LV ejection fraction (LVEF), left atrial end-systolic volume (LAESV), and Doppler parameters: mitral E wave velocity (E), A wave velocity (A), mitral annulus e' velocity (e'), and tricuspid regurgitation velocity (TRmax). The algorithm was trained and tested on two internal datasets (862 and 239 echocardiograms) and validated using three external datasets, including EchoNet-Dynamic and CAMUS. The ASE diastolic function decision tree and total probability theory were used to provide diastolic grading probabilities. The algorithm, named MMnet, demonstrated high accuracy in both test and validation datasets, with Dice coefficients for segmentation between 0.922 and 0.932 and classification accuracies between 0.9977 and 1.0. The mean absolute errors (MAEs) for LVEF and LAESV were 3.7 % and 5.8 ml, respectively, and for LVEF in external validation, MAEs ranged from 4.9 % to 5.6 %. The diastolic function grading accuracy was 0.88 with hard criteria and up to 0.98 with soft criteria which account for the top two probability in total probability theory. MMnet can automatically grade ASE diastolic function with high accuracy and efficiency by annotating 2D videos and Doppler images.