Weakly supervised anatomical feature learning for cross-dataset ejection fraction estimation from echocardiography videos
TL;DR
The findings support controlled external validation under explicit image and target harmonization, while further prospective validation, calibration, and quality-control mechanisms are needed before clinical deployment.
Abstract
Ejection fraction (EF) is a central measure of cardiac function, but echocardiographic EF assessment remains reader-dependent and sensitive to acquisition quality. Deep learning can automate EF estimation, yet performance measured on a single development dataset may not transfer to data acquired under different imaging and annotation conventions. We investigated whether anatomically constrained weakly supervised learning improves cross-dataset EF estimation. We propose CAFEx, a contrastive-augmented feature extraction pipeline that combines left ventricular segmentation, echocardiography-specific augmentation, mask-derived anatomical features, and temporal EF regression. The model was trained on EchoNet-Dynamic using video-level EF labels with limited dense annotations for the anatomical component. External evaluation was performed on CAMUS after image harmonization and recalculation of an apical-four-chamber monoplane EF endpoint. Performance was compared with reproduced segmentation-based, direct video-regression, graph-based, transformer-based, and feature-extraction baselines using Dice score, mean absolute error, and coefficient of determination. In the harmonized train-on-EchoNet/test-on-CAMUS benchmark, anatomically constrained models degraded less than direct video-regression baselines. CAFEx achieved 90.73% Dice on CAMUS and improved CAMUS EF prediction over the strongest reproduced feature-extraction baseline, increasing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} from 0.38 to 0.54 and reducing mean absolute EF error from 7.89 to 6.70 percentage points. Poor-quality videos and extreme EF values remained challenging. Weakly supervised EF estimation can benefit from an anatomical bottleneck learned with limited dense annotation. The findings support controlled external validation under explicit image and target harmonization, while further prospective validation, calibration, and quality-control mechanisms are needed before clinical deployment.