Cyclical Self-Supervision for Semi-Supervised Ejection Fraction Prediction From Echocardiogram Videos

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Left-ventricular ejection fraction (LVEF) is a critical metric for diagnosing heart failure.
• Current LVEF estimation methods from echocardiogram videos demand extensive labeled data, hindering broader application.
• Manual video annotation is labor-intensive and time-consuming.

Purpose of the Study:

• To develop the first semi-supervised approach for accurate LVEF prediction from echocardiogram videos.
• To reduce the reliance on large, fully annotated datasets for LVEF estimation.
• To improve the efficiency and applicability of LVEF analysis in clinical settings.

Main Methods:

• Proposed a novel Cyclical Self-Supervision (CSS) method for unsupervised learning of left ventricular (LV) segmentation from video.
• Integrated LV segmentation mask prediction with LVEF regression to provide spatial context.
• Employed teacher-student distillation to transfer segmentation knowledge into an end-to-end LVEF regression model.

Main Results:

• Achieved a Mean Absolute Error (MAE) of 4.17 for LVEF prediction, competitive with supervised methods.
• Utilized half the number of labeled videos compared to state-of-the-art supervised approaches.
• Demonstrated improved generalization capabilities on an external dataset.

Conclusions:

• The proposed semi-supervised method effectively estimates LVEF with reduced data requirements.
• Incorporating LV segmentation knowledge enhances LVEF prediction accuracy and model robustness.
• This approach offers a promising direction for efficient and scalable heart failure assessment using echocardiography.