Contour-constrained branch U-Net for accurate left ventricular segmentation in echocardiography

Area of Science:

• Medical Imaging
• Cardiology
• Artificial Intelligence

Background:

• Echocardiography is vital for assessing left ventricular (LV) function, but image quality issues like noise and artifacts hinder accurate LV segmentation.
• Accurate LV segmentation is crucial for deriving key clinical parameters, including ejection fraction, yet challenging due to boundary delineation difficulties.

Purpose of the Study:

• To develop a simplified yet effective method for precise LV segmentation in echocardiographic images.
• To enhance the accuracy of LV boundary segmentation, thereby improving the reliability of derived cardiac diagnostic parameters.

Main Methods:

• Introduced a novel Branch U-Net (BU-Net) by integrating a branch sub-network into the decoder of a standard U-Net architecture.
• Utilized LV contour features for supervising the branch decoding process and employed a cross-attention module to enhance feature interaction.
• Tested the model on the CAMUS and EchoNet-dynamic public echocardiography segmentation datasets.

Main Results:

• The proposed BU-Net achieved superior segmentation performance compared to existing state-of-the-art models on public datasets.
• Demonstrated improved accuracy specifically in segmenting challenging LV boundary regions.
• Achieved high performance without relying on complex architectures like residual connections or transformers.

Conclusions:

• The BU-Net offers a simplified and effective approach for accurate left ventricular segmentation in echocardiography.
• This method holds potential for improving the clinical diagnosis of cardiac function by providing more reliable segmentation results.
• The publicly available code facilitates further research and application in cardiac image analysis.