Velocity-based cardiac contractility personalization from images using derivative-free optimization

Ken C. L. Wong, Maxime Sermesant, Kawal Rhode, Matthew Ginks, Christopher A. Rinaldi, Reza Razavi, Hervé Delingette, Nicholas AyacheJournal of the Mechanical Behavior of Biomedical materials/Journal of Mechanical Behavior of Biomedical Materials 2014 · DOI

TL;DR

This paper introduces a novel method for personalizing cardiac contractility models using velocity data from medical images and derivative-free optimization. This approach allows for simultaneous estimation of contraction and relaxation rates, overcoming limitations of previous methods that focused only on peak contraction, and demonstrating improved accuracy and robustness.

What Problem Does This Paper Address?

Accurately modeling the mechanical behavior of the heart is crucial for understanding and treating cardiovascular diseases. Personalizing these models to individual patients using medical imaging data is a key step towards clinical application. However, existing methods often struggle to identify not just the maximum force a heart muscle can generate (contractility), but also how quickly it contracts and relaxes. This is partly because estimating these rates requires complex calculations that are difficult to perform with traditional gradient-based optimization techniques, especially when dealing with intricate cardiac models. The authors aim to address this by developing a method that can simultaneously estimate these key contractility parameters, including the rates, using a more robust optimization approach.

Methods

The core of this work lies in two main innovations:

  • Velocity-Based Objective Function: Instead of using the position of heart structures to measure how well a model matches real data, the authors propose using the velocity of these structures. They argue that velocity data better captures the temporal dynamics of contraction and relaxation, which are crucial for estimating the rates. This is achieved by defining an objective function that minimizes the difference between simulated and measured velocities.
  • Derivative-Free Optimization (SUBPLEX): Traditional optimization methods rely on calculating the gradient (slope) of the objective function. For complex cardiac models, this is computationally expensive and prone to errors. The authors adopt a derivative-free optimization algorithm called SUBPLEX. This method makes fewer assumptions about the objective function's properties, making it more suitable for the complex, potentially non-smooth landscapes encountered when estimating multiple contractility parameters. SUBPLEX is a generalization of the Nelder-Mead simplex method, designed to handle higher-dimensional problems more efficiently by searching in subspaces.

Datasets

The authors evaluated their framework using:

  • Synthetic Data: Generated from a patient-specific cardiac model with known ground truth parameters. This allowed for quantitative comparison of their method against others and assessment of its robustness to noise and initial parameter choices.
  • Clinical Data:
    • Patient Data: From three patients with heart conditions (myocardial infarction, dilated cardiomyopathy, left bundle branch block). This data included cine MRI for geometry and motion, and electrophysiology data for activation times.
  • Healthy Volunteer Data: From three volunteers with cine MRI and heart period measurements.

Key Insights and Ideas

  • Velocity is Key for Rates: Using velocity differences, rather than position differences, in the objective function significantly improves the ability to identify both contraction and relaxation rates, not just the peak contraction stress. This is because velocity directly reflects the dynamic changes over time.
  • Derivative-Free Optimization is Essential: The complex, multi-parameter nature of cardiac contractility estimation leads to objective function landscapes that are not always well-behaved (e.g., flat valleys). Derivative-free methods like SUBPLEX are more robust to these characteristics than gradient-based methods, especially when estimating multiple parameters simultaneously.
  • Robustness to Initial Conditions: The proposed velocity-based framework with SUBPLEX optimization demonstrates good robustness to variations in initial parameter guesses, which is a significant advantage for clinical applications where precise initializations might be difficult.
  • Clinical Relevance: The personalized models generated using this framework show improved consistency with personalized kinematics and, in patient cases, reflect known pathologies like reduced contractility in infarcted regions.

Key Results

  • Synthetic Data Performance: The velocity-based optimization consistently outperformed position-based optimization, especially when estimating contraction and relaxation rates. It converged faster and achieved better accuracy, particularly in higher-dimensional scenarios (e.g., AHA-zone representation).
  • Robustness: The framework showed good robustness to measurement noise (down to SNR=1) and variations in initial parameters.
  • Clinical Data Consistency: Personalized simulations using the framework showed improved agreement with personalized kinematics. For patients, the estimated contractility patterns reflected known pathologies (e.g., reduced contractility in scar tissue). For healthy volunteers, the results were consistent with normal cardiac function.
  • Comparison to Other Optimizers: When compared to L-BFGS-B and BOBYQA on synthetic data, SUBPLEX demonstrated superior performance in estimating both peak contraction stress and the rates, with L-BFGS-B showing significant deviations.

Limitations and Future Work

  • Passive Properties: The current framework assumes passive mechanical properties (stiffness) are known and homogeneous. Future work could incorporate personalization of these properties, which are known to vary, especially in diseased hearts.
  • Model Complexity: The model has limitations in accurately simulating valve dynamics and precise pressure curves, which affects the perfect matching of simulated and measured pressures.
  • Manual Adjustments: Some calibrations, like time delays between electrical and mechanical events, were manually adjusted. Future work could automate these.
  • Image Modalities: The study relied on cine MRI for velocity information. The availability of phase-contrast MRI could provide more direct velocity measurements, potentially improving patient-specific results.
  • Computational Cost: While derivative-free optimization helps, estimating parameters for many zones can still be computationally intensive. Exploring hierarchical approaches or parameter decomposition could reduce this burden.
  • Invasive Measurements: Incorporating invasive measurements like blood pressure directly into the objective function is a challenging but promising area for future research.

Summary updated

BibTeX
@article{wong2014velocity,
  author = {Ken C. L. Wong and Maxime Sermesant and Kawal Rhode and Matthew Ginks and Christopher A. Rinaldi and Reza Razavi and Hervé Delingette and Nicholas Ayache},
  title = {{Velocity-based cardiac contractility personalization from images using derivative-free optimization}},
  journal = {Journal of the Mechanical Behavior of Biomedical materials/Journal of Mechanical Behavior of Biomedical Materials},
  year = {2014},
  doi = {10.1016/j.jmbbm.2014.12.002},
}

Loading related content...

Figure