Co-learning of appearance and shape for precise ejection fraction estimation from echocardiographic sequences

Area of Science:

• Cardiovascular Imaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Accurate ejection fraction (EF) estimation from echocardiography is crucial for cardiac function assessment.
• Current methods face challenges including image noise, myocardial motion, sparse annotations, and quality degradation.
• Traditional Simpson's bi-plane method relies on left ventricle (LV) segmentation in keyframes.

Purpose of the Study:

• To propose a multi-task semi-supervised framework (MCLAS) for precise EF estimation from echocardiographic sequences.
• To automate the entire EF estimation pipeline, from view identification to EF calculation.
• To improve the accuracy and temporal consistency of cardiac segmentation and tracking.

Main Methods:

• Developed a co-learning mechanism for iterative segmentation and myocardium tracking, enhancing appearance and shape.
• Implemented auxiliary tasks: view classification for feature extraction and EF regression for spatiotemporal regularization.
• Utilized a semi-supervised approach to leverage available annotations effectively.

Main Results:

• Achieved superior performance in LV volumes (ED/ES phases) and EF estimation compared to existing methods.
• Demonstrated high Pearson correlations: 0.975 for LV volumes (ED), 0.983 for LV volumes (ES), and 0.946 for EF.
• Generated accurate and temporally consistent segmentation for intermediate frames, enabling dynamic function evaluation.

Conclusions:

• The MCLAS framework significantly improves automated EF estimation from echocardiography.
• The proposed co-learning and auxiliary tasks enhance segmentation accuracy and temporal consistency.
• This method holds substantial potential for clinical application and automated cardiac function assessment.