An efficient, scalable, and adaptable plug-and-play temporal attention module for motion-guided cardiac segmentation...

Area of Science:

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

Background:

• Cardiac anatomy segmentation is crucial for assessing cardiac function and diagnosing diseases.
• Deep learning (DL) models have improved segmentation accuracy, especially when incorporating cardiac motion dynamics.
• Existing methods for integrating motion information are computationally expensive or prone to registration errors.

Purpose of the Study:

• To develop a computationally efficient and robust Temporal Attention Module (TAM) for enhancing deep learning-based cardiac anatomy segmentation.
• To create a plug-and-play module adaptable to various segmentation network architectures (CNN, transformer, hybrid).
• To improve the accuracy and efficiency of cardiac motion-aware segmentation.

Main Methods:

• Proposed a novel Temporal Attention Module (TAM) using a multi-headed, cross-temporal attention mechanism.
• Integrated TAM into diverse segmentation networks including UNet, FCN8s, UNetR, SwinUNetR, I²UNet, and DT-VNet.
• Validated TAM on multiple cardiac imaging datasets: 2D echocardiography (CAMUS, EchoNet-Dynamic), 3D echocardiography (MITEA), and 3D cardiac MRI (ACDC).

Main Results:

• TAM consistently improved segmentation performance across all tested datasets and networks.
• Integration of TAM into SAM and MedSAM reduced Hausdorff distance (HD) significantly.
• TAM-UNet and TAM-DT-VNet demonstrated substantial HD reductions on the ACDC 3D dataset.
• TAM training requires only sparse temporal annotation, not full ground truth segmentation for all frames.

Conclusions:

• The Temporal Attention Module (TAM) offers a robust, generalizable, and adaptable solution for motion-aware cardiac image segmentation.
• TAM enhances segmentation accuracy and efficiency without drastic architectural modifications, scaling from 2D to 3D.
• The proposed method reduces computational costs and reliance on complex registration techniques, making it a practical advancement for clinical applications.