Predicting Ejection Fraction from Multi-View Echocardiography with Scarce Labels
ORIGINAL / Learning from Scarce Labels: Multi-View Echocardiography for Ejection Fraction Prediction
This study introduces the first publicly available dataset for predicting left ventricular ejection fraction (EF) from parasternal long-axis (PLAX) echocardiography, overcoming label scarcity via an innovative data generation strategy, and demonstrates the feasibility of PLAX-based EF estimation and the benefit of multi-view fusion.
01 ABSTRACT
The authors propose a novel data generation method that leverages temporal correlations between clinical notes and echocardiographic videos, combined with view classifier fine-tuning and proxy labeling, to create a labeled dataset of over 25,000 PLAX videos. Using this, they train the first reproducible PLAX EF model, achieving a mean absolute error (MAE) of 6.86%, comparable to the clinical standard A4C method (6%-7%). Furthermore, simple unweighted late fusion of PLAX and A4C predictions reduces MAE to 6.37%, outperforming single views. The authors conclude that EF estimation from PLAX views is feasible and clinically relevant.
02 KEY FINDINGS
- First publicly available PLAX-EF dataset with over 25,000 labeled PLAX videos.
- Data generation strategy based on temporal correlation and proxy labeling.
- PLAX EF model achieves MAE of 6.86%, comparable to A4C methods.
- Multi-view fusion (PLAX+A4C) reduces MAE to 6.37%, outperforming single views.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG