Boundary attention with multi-task consistency constraints for semi-supervised 2D echocardiography segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Accurate segmentation of 2D echocardiograms is crucial for assessing cardiac function and diagnosing heart diseases.
• Current automatic segmentation methods struggle with loss of boundary and localization information, and data limitations.

Purpose of the Study:

• To propose a novel semi-supervised echocardiography segmentation method to overcome existing limitations.
• To improve the accuracy and efficiency of cardiac structure segmentation in 2D echocardiograms.

Main Methods:

• Developed a Boundary Attention Transformer Net (BATNet) to capture spatial and location information using self-attention.
• Introduced a multi-task level semi-supervised model (semi-BATNet) with boundary feature consistency constraints.
• Utilized consistency loss across different scales for unlabeled data between student and teacher networks.

Main Results:

• On the CAMUS dataset, semi-BATNet significantly improved segmentation with only 25% labeled images, outperforming five state-of-the-art methods.
• Achieved Dice coefficient of 0.936 and Jaccard similarity of 0.881 on a self-collected dataset with 50% labeled images.
• Demonstrated superior performance in segmenting cardiac structures compared to existing semi-supervised approaches.

Conclusions:

• The proposed semi-BATNet effectively enhances segmentation accuracy by leveraging unlabeled echocardiogram data.
• This method shows significant potential for accurately and efficiently assisting cardiologists in clinical practice.
• Addresses key challenges in 2D echocardiogram segmentation, including boundary and localization information loss.