Dynamic-Guided Spatiotemporal Attention for Echocardiography Video Segmentation.

Area of Science:

• Medical Imaging
• Cardiovascular Imaging
• Artificial Intelligence in Medicine

Background:

• Left ventricle (LV) endocardium segmentation is crucial for quantifying LV ejection fraction from echocardiography videos.
• Existing methods often rely on 2D convolutional networks and struggle with temporal consistency, sometimes using noisy optical flow estimation (OFE).

Purpose of the Study:

• To develop a novel semi-supervised method for echocardiography video segmentation that overcomes limitations of existing approaches.
• To improve temporal consistency and accuracy in LV endocardium segmentation by effectively integrating dynamic and appearance information.

Main Methods:

• Fine-tuning the RAFT network for OFE on echocardiography data to obtain reliable dynamic information.
• Utilizing a dual-encoder structure to separately extract motion and appearance features, with inter-frame flows as input.
• Proposing a dynamic-guided spatiotemporal attention (DSA) mechanism with deformable attention for enhanced temporal modeling and feature calibration.

Main Results:

• The proposed DSA method achieved state-of-the-art performance on the CAMUS and EchoNet-Dynamic echocardiography datasets.
• The approach effectively integrates dynamic continuity with semantic consistency, enhancing segmentation accuracy.
• The bilateral feature calibration and offset estimation modules contribute to robust feature enhancement.

Conclusions:

• The dynamic-guided spatiotemporal attention (DSA) offers a significant advancement in semi-supervised echocardiography video segmentation.
• This method provides a more robust and accurate approach for quantifying LV ejection fraction.
• The findings highlight the potential of integrating refined dynamic information for improved medical image analysis.