Echocardiography Video Segmentation via Neighborhood Correlation Mining.

Area of Science:

• Medical imaging analysis
• Cardiovascular disease diagnostics
• Artificial intelligence in healthcare

Background:

• Accurate left ventricle segmentation in echocardiography is crucial for cardiovascular disease diagnosis and treatment.
• Current segmentation methods struggle with ultrasound imaging limitations and sparse annotations.
• Existing approaches fail to effectively address noise and boundary refinement challenges in echocardiography segmentation.

Purpose of the Study:

• To propose a novel semi-supervised segmentation framework, NCM-Net, for echocardiography.
• To enhance segmentation accuracy and temporal consistency in cardiac ultrasound images.
• To overcome limitations of sparse annotations and noise in existing segmentation methods.

Main Methods:

• Developed the Neighborhood Correlation Mining (NCM) module to mine spatiotemporal correlations and refine features, reducing noise impact.
• Introduced Unreliable-Pixels Masked Attention (UMA) to focus on refining segmentation boundaries by prioritizing unreliable pixels.
• Implemented cross-frame boundary constraints to optimize temporal consistency of segmentation predictions.

Main Results:

• NCM-Net achieved state-of-the-art performance on the CAMUS and EchoNet-Dynamic datasets.
• The proposed framework demonstrated outstanding temporal consistency in echocardiography segmentation.
• Experimental results validate the effectiveness of the NCM module and UMA in improving segmentation accuracy.

Conclusions:

• NCM-Net offers a significant advancement in semi-supervised echocardiography segmentation.
• The framework effectively addresses noise and annotation sparsity, leading to improved diagnostic accuracy.
• The proposed methods enhance both segmentation accuracy and temporal consistency for cardiovascular imaging analysis.