SSMCE: A semi-supervised learning framework for myocardial segmentation in myocardial contrast echocardiography

Area of Science:

• Medical imaging analysis
• Machine learning in cardiology
• Image processing for echocardiography

Background:

• Accurate myocardial segmentation in myocardial contrast echocardiography (MCE) is crucial but hindered by limited labeled datasets and speckle noise.
• Current manual delineation by echocardiographers is time-consuming and variable, impacting clinical workflow efficiency.

Purpose of the Study:

• To develop a novel semi-supervised learning framework (SSMCE) for automated myocardial segmentation in MCE images.
• To address the challenges of data scarcity and image noise in MCE segmentation.
• To improve the accuracy, robustness, and efficiency of myocardial segmentation in clinical practice.

Main Methods:

• Proposed a semi-supervised learning framework (SSMCE) with a tri-model architecture (two student models, one adaptive teacher model).
• Implemented model-level perturbations to enhance output diversity and reduce overfitting.
• Designed a specialized loss function to guide model self-correction and improve convergence.

Main Results:

• The proposed loss function improved the primary evaluation metric by 1.75% on a self-constructed dataset.
• SSMCE achieved state-of-the-art performance compared to existing myocardial segmentation methods.
• Demonstrated robust and efficient myocardial detection and segmentation capabilities.

Conclusions:

• SSMCE offers a robust and efficient solution for myocardial segmentation in MCE imaging.
• The framework has significant potential to streamline clinical workflows and improve diagnostic accuracy.
• Addresses key limitations in current MCE image analysis, paving the way for wider adoption.