Fully Automatic Left Atrial Strain Quantification via Multi-task Learning on Cardiac Cine MRI

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Biomedical Engineering

Background:

• Cardiac magnetic resonance feature tracking (CMR-FT) for left atrial (LA) strain is limited by contouring errors, motion inconsistencies, and temporal drift.
• Manual and landmark-based methods for LA strain lack reproducibility and scalability.

Purpose of the Study:

• To develop and validate a fully automatic, multi-task learning framework for accurate and efficient LA strain quantification using CMR cine images.
• To benchmark the proposed method against existing feature tracking algorithms.

Main Methods:

• A multi-task learning model was developed using a multi-center, two-vendor cine MRI dataset, coupling groupwise registration and segmentation networks via spatiotemporal cross-attention.
• Performance was evaluated against optical flow, pairwise registration, and VoxelMorph using metrics like mean-squared error, contour distance, and drift error.
• Diagnostic performance was assessed using ROC analysis to differentiate healthy from diseased subjects.

Main Results:

• The proposed method demonstrated superior tracking accuracy and significantly reduced drift compared to other algorithms, achieving inference in 0.5 seconds.
• Automatic strain measurements showed high agreement with manual segmentation (reservoir r=0.95, conduit r=0.96, booster r=0.92).
• In an external dataset, diseased subjects exhibited significantly lower LA strain components (p<0.001), with reservoir strain achieving high discriminative power (AUC: 0.81-0.97).

Conclusions:

• A fully automatic, multi-task learning framework for LA strain quantification has been developed and validated on multi-center, two-vendor data.
• This framework enhances tracking accuracy and speed, enabling rapid and scalable assessment of atrial function in routine clinical practice.
• The source code is publicly available, promoting further research and application.