Weakly supervised anatomical feature learning for cross-dataset ejection fraction estimation from echocardiography videos

Figures

• 

  Fig. 1. Distribution of EF values across datasets and EF calculation variants. EchoNet-Dynamic and CAMUS differ not only in image format and annotation type, but also in their EF distributions and released target definitions

• 

  Fig. 2. Target harmonization for CAMUS. Panel (A) shows the distribution of EF deviations between the recalculated apical-four-chamber monoplane endpoint and the original CAMUS biplane EF, with the zero-deviation reference and mean bias indicated. Panel (B) compares original biplane EF with recalculated monoplane EF and reports the correlation and mean absolute error between the two endpoints. Panel (C) shows the bland–Altman agreement plot, with mean bias and 95% limits of agreement. The recalculated monoplane endpoint is used as the explicit single-plane target for the external evaluation, but the observed bias and limits of agreement indicate that it should not be treated as identical to the original CAMUS biplane reference

• 

  Fig. 3. Geometric feature extraction from a predicted left ventricular segmentation. The line-length feature family approximates the clinical Simpson workflow by estimating the long axis and measuring short-axis diameters along the ventricle

• 

  Fig. 4. CAFEx prediction pipeline. The input video is sampled and padded, a contrastive-augmented segmentation model predicts left ventricular masks, anatomical features \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_n$$\end{document} are extracted from each mask, and a temporal regressor predicts EF

• 

  Fig. 5. Contrastive pretraining used for the segmentation backbone. Two augmented views of the same echocardiographic frame are encouraged to have nearby representations, whereas different frames are pushed apart. This stage uses image structure rather than manual EF or mask labels and is therefore compatible with annotation-efficient training

• 

  Fig. 6. CAMUS EF predictions from representative baselines. In each panel, samples are ordered by ground-truth EF; the blue curve represents sorted ground-truth EF values, and orange markers represent the corresponding predictions at the same sorted indices. These plots are intended as qualitative error-structure visualizations rather than substitutes for agreement plots. Direct video-regression approaches show substantial error structure in the evaluated dataset shift, whereas feature-based prediction is less degraded in this benchmark

• 

  Fig. 7. Absolute EF prediction errors of representative baselines stratified by dataset and CAMUS video quality. The x-axis categories are the EchoNet-Dynamic test set and the CAMUS good-, medium-, and poor-quality subsets. The quality stratification reveals that poor-quality videos are challenging in this benchmark, but model architecture is associated with the magnitude of external-benchmark degradation

• 

  Fig. 8. Ground-truth EF values and corresponding predictions on CAMUS, sorted by ground-truth EF. In each panel, the blue curve represents sorted ground-truth EF values, and orange markers represent the corresponding predictions. CAFEx reduces the variance of the prediction errors compared with the feature-extraction baseline, although high EF values remain underestimated

• 

  Fig. 9. CAFEx EF prediction behavior. Panel (a) shows EchoNet-Dynamic test-set predictions sorted by ground-truth EF. Panel (b) shows absolute EF errors for the EchoNet-Dynamic test set and CAMUS good-, medium-, and poor-quality subsets. The model performs well for many mid-range EF values, but the quality-stratified CAMUS results show that poor-quality videos remain a limiting factor

• 

  Fig. 10. CAFEx segmentation Dice scores split by dataset, CAMUS video quality, and ED/ES frame type. ED frames are consistently more difficult than ES frames, and poor-quality CAMUS videos show the largest variance