Cardiac motion tracking with bidirectional latent neural ODEs from cine cardiac MRI

Area of Science:

• Medical Imaging
• Computational Biology
• Machine Learning

Background:

• Accurate cardiac motion tracking from cine MRI is crucial for assessing heart function and diagnosing diseases.
• Current learning-based methods often fail to capture long-term temporal dependencies, leading to errors and physically inconsistent motion.
• Existing approaches typically operate in discrete time, limiting their ability to model continuous physiological processes.

Purpose of the Study:

• To develop a unified, unsupervised framework for accurate and temporally consistent cardiac motion tracking.
• To address the limitations of discrete-time models in capturing complex cardiac dynamics.
• To improve the physiological plausibility and long-term consistency of motion tracking in cardiac MRI.

Main Methods:

• A novel framework utilizing latent neural ordinary differential equations (ODEs) for continuous-time cardiac motion modeling.
• A frame-aware encoder with temporal embeddings to extract motion-sensitive features.
• Bidirectional forward-backward evolution in a latent space with a conditional MLP-based dynamics module.
• Bidirectional Lagrangian regularization to ensure temporal consistency and motion reversibility.

Main Results:

• The proposed method achieves state-of-the-art performance on the ACDC and M&Ms datasets.
• Demonstrates highly temporally consistent and physiologically plausible cardiac motion fields.
• Achieves these results with a lightweight and computationally efficient architecture.

Conclusions:

• The latent neural ODE framework offers a significant advancement in unsupervised cardiac motion tracking.
• The method effectively models continuous-time deformation dynamics, overcoming limitations of discrete approaches.
• This approach provides a robust and efficient tool for cardiac functional assessment and disease analysis using MRI.