Generative Myocardial Motion Tracking via Latent Space Exploration with Biomechanics-informed Prior

Chen Qin, Shuo Wang, Chen Chen, Wenjia Bai, Daniel Rueckert

Abstract

Myocardial motion and deformation are rich descriptors that characterize cardiac function. Image registration, as the most commonly used technique for myocardial motion tracking, is an ill-posed inverse problem which often requires prior assumptions on the solution space. In contrast to most existing approaches which impose explicit generic regularization such as smoothness, in this work we propose a novel method that can implicitly learn an application-specific biomechanics-informed prior and embed it into a neural network-parameterized transformation model. Particularly, the proposed method leverages a variational autoencoder-based generative model to learn a manifold for biomechanically plausible deformations. The motion tracking then can be performed via traversing the learnt manifold to search for the optimal transformations while considering the sequence information. The proposed method is validated on three public cardiac cine MRI datasets with comprehensive evaluations. The results demonstrate that the proposed method can outperform other approaches, yielding higher motion tracking accuracy with reasonable volume preservation and better generalizability to varying data distributions. It also enables better estimates of myocardial strains, which indicates the potential of the method in characterizing spatiotemporal signatures for understanding cardiovascular diseases.

Introduction

Tracking the motion of the heart muscle (the myocardium) from MRI scans is essential for diagnosing cardiovascular diseases. However, it is a classic "ill-posed" inverse problem: because the heart is a complex, deforming object and MRI data can be noisy or low-resolution, there are infinitely many mathematical ways to "warp" one image frame to the next.

Traditional methods try to solve this by adding generic constraints, like "make the motion smooth." While helpful, these constraints don't capture the actual physics of how heart tissue behaves. On the other hand, modern deep learning (DL) methods are fast but often "black boxes" that require thousands of perfectly labeled images to train. These DL models frequently fail when moved from the hospital where they were trained to a new site with different scanners or protocols (a problem known as "domain shift").

Methodology

The authors propose a hybrid approach that merges biomechanical priors with latent space exploration. The workflow consists of two main stages:

  1. Learning the Prior (The Generative Model): Instead of training on real MRIs, the researchers used Finite Element Methods (FEM) to simulate thousands of physically realistic heart deformations based on the laws of solid mechanics. They then trained a Temporal Variational Autoencoder (VAE)—incorporating Recurrent Neural Networks (RNNs)—to learn a low-dimensional "latent space" (a manifold) that represents all these plausible motions.
  2. Tracking via Exploration (Inference): When presented with a new patient's MRI, the model doesn't just "guess" the motion in one shot. Instead, it performs an optimization: it searches (traverses) the pre-learned latent space to find the specific deformation that best aligns the MRI frames.
Overview of the proposed framework including the biomechanical prior learning, VAE architecture, and latent space exploration.

By forcing the solution to live within the VAE’s latent space, the motion is guaranteed to be biomechanically plausible. Because the model is "searching" at test time rather than just regressing, it can adapt to new data distributions much more effectively.

Experimental Setup

The model was evaluated on three major public cardiac cine MRI datasets:

  • UK Biobank (UKBB): 300 sequences (used for primary testing and as the source for simulation geometries).
  • ACDC Dataset: 100 sequences featuring various cardiac pathologies (used to test generalization).
  • M&Ms (Multi-Centre, Multi-Vendor, Multi-Disease) Dataset: 150 sequences from different MRI vendors (used to test robustness to domain shift).

Results and Discussion

Key Insights

  • Physics as a Substitute for Big Data: By using biomechanical simulations to train the VAE, the authors created a "data-efficient" model. It learns what a heart can do from physics, meaning it doesn't need thousands of manually annotated human scans to understand motion.
  • Optimization vs. Regression: Most DL trackers are "feed-forward," meaning they give one answer and stop. This paper uses the DL model as a parameterized space and then uses classical optimization to find the best fit. This "best of both worlds" approach is why the model generalizes so well to different scanners.
  • Temporal Coherence: By using a temporal VAE with RNN layers, the model ensures that the heart's motion is consistent over the entire cardiac cycle, preventing the "jitter" often seen in frame-by-frame tracking methods.

Performance Metrics

  • Superior Accuracy: The proposed method outperformed state-of-the-art DL baselines (like Motion-Net and BINN) and classical methods (like FFD). It achieved higher Dice coefficients (region overlap) and lower Mean Contour Distance (MCD) across all datasets.
  • Robust Generalization: On the ACDC and M&Ms datasets—which the model never saw during training—it showed an average improvement of 2.4% in Dice score and 0.28 mm in MCD compared to other DL methods.
  • Clinical Utility: The method provided better estimates of myocardial strain (radial and circumferential), which are critical biomarkers for identifying diseases like hypertrophic cardiomyopathy or heart failure.

Conclusion

While the current study focuses on 2D short-axis slices, the results are promising for clinical practice. Future work will address the extension to 3D simulations and tracking to account for through-plane motion. Although the optimization step at test time (1–3 seconds per sequence) is slower than one-shot DL methods, it remains significantly faster than traditional iterative methods while offering superior robustness and physical plausibility.

Summary updated

Abstract

Myocardial motion and deformation are rich descriptors that characterize cardiac function. Image registration, as the most commonly used technique for myocardial motion tracking, is an ill-posed inverse problem which often requires prior assumptions on the solution space. In contrast to most existing approaches which impose explicit generic regularization such as smoothness, in this work we propose a novel method that can implicitly learn an application-specific biomechanics-informed prior and embed it into a neural network-parameterized transformation model. Particularly, the proposed method leverages a variational autoencoder-based generative model to learn a manifold for biomechanically plausible deformations. The motion tracking then can be performed via traversing the learnt manifold to search for the optimal transformations while considering the sequence information. The proposed method is validated on three public cardiac cine MRI datasets with comprehensive evaluations. The results demonstrate that the proposed method can outperform other approaches, yielding higher motion tracking accuracy with reasonable volume preservation and better generalizability to varying data distributions. It also enables better estimates of myocardial strains, which indicates the potential of the method in characterizing spatiotemporal signatures for understanding cardiovascular diseases.
BibTeX
@article{qin2023generative,
  author = {Chen Qin and Shuo Wang and Chen Chen and Wenjia Bai and Daniel Rueckert},
  title = {{Generative Myocardial Motion Tracking via Latent Space Exploration with Biomechanics-informed Prior}},
  journal = {Medical Image Analysis},
  year = {2023},
  doi = {10.1016/j.media.2022.102682},
}
Figure