Evaluation of iLogDemons Algorithm for Cardiac Motion Tracking in Synthetic Ultrasound Sequence

Adityo Prakosa, Kristin McLeod, Maxime Sermesant, Xavier PennecSTACOM 2012 · DOI

TL;DR

This paper evaluates the iLogDemons algorithm for tracking cardiac motion in synthetic ultrasound sequences. By using synthetic data with known ground truth, the authors objectively assess the algorithm's performance and identify optimal parameters, finding that while the algorithm is generally accurate, the incompressibility constraint doesn't improve results due to the non-incompressible nature of the underlying biomechanical model.

What Problem Does This Paper Address?

Accurately tracking cardiac motion is crucial for understanding heart function and diagnosing diseases. Traditional methods often rely on medical images, but objectively evaluating their performance is challenging due to the lack of precise ground truth. The STACOM (Statistical Atlases and Computational Models of the Heart) challenge aims to address this by providing standardized datasets. This paper specifically tackles the problem of quantitatively evaluating a sophisticated cardiac motion tracking algorithm, iLogDemons, by leveraging a synthetic ultrasound dataset that provides the ground truth deformation. This allows for a more rigorous assessment of the algorithm's accuracy and the impact of its various parameters.

Methods

The core of this work is the evaluation of the iLogDemons algorithm. This algorithm builds upon the LogDemons framework, which is a non-linear registration method designed for tracking tissue motion. iLogDemons enhances LogDemons by incorporating physiological constraints: elasticity and incompressibility.

  • LogDemons: This method aligns a template image to a reference image by estimating a dense, non-linear transformation. It uses an energy minimization approach that balances image similarity with a regularization term to ensure smooth and invertible deformations. The transformation is parameterized by a stationary velocity vector field, ensuring invertibility.
  • iLogDemons: This extension adds two key features:
  • Elastic Regularizer: This acts as a vector filter, smoothing the velocity field using an elastic-like kernel.
  • Incompressibility Constraint: This forces the velocity field to be divergence-free, mimicking the property of incompressible materials. This is achieved by solving a Poisson equation.

The paper also describes a Cardiac Motion Tracking Strategy where the registration of sequential frames is initialized by concatenating previous frame-to-reference and current-to-previous frame velocity fields, using the Baker-Campbell-Hausdorff (BCH) formula. This strategy aims to improve tracking accuracy over time.

Datasets

The primary dataset used for evaluation is the STACOM 2012 synthetic ultrasound sequence dataset. This dataset consists of:

  • 10 synthetic ultrasound sequences, each with 23 frames.
  • Image spatial resolution of 267x355x355 with an isotropic voxel size of 0.33 mm.
  • Simulated motion and deformation patterns including normal, LBBB, RBBB, and pacing.
  • Crucially, ground truth deformation is provided as volumetric meshes, allowing for direct comparison with estimated motion.

The authors also mention using a down-sampled version of this dataset to assess the trade-off between computational time and accuracy.

Key Insights and Ideas

  • Objective Evaluation is Key: The use of synthetic data with ground truth is invaluable for objectively evaluating motion tracking algorithms, moving beyond qualitative assessments or comparisons to less precise reference data.
  • Incompressibility's Limited Impact: The paper's most significant finding is that the incompressibility constraint in iLogDemons does not improve performance on this specific synthetic dataset. This is attributed to the fact that the underlying electromechanical model used to generate the synthetic data is not incompressible, meaning enforcing this constraint doesn't align with the true physics of the simulated motion.
  • Parameter Sensitivity: The choice of the regularization parameter $\sigma$ (specifically $\sigma_{stationary velocity field}$) is shown to be important. While $\sigma = 1.5$ yielded the best results in their experiments, values of $\sigma = 1$ and $\sigma = 2$ did not offer significant improvements, suggesting a specific optimal range.
  • Down-sampling Trade-off: Down-sampling the data significantly reduces computation time (from days to hours) with only minor differences in registration error, making it a viable strategy for faster processing when absolute precision is not paramount.

Key Results

  • Accuracy: The iLogDemons algorithm, both with and without the incompressibility constraint, was able to reasonably estimate the ground truth deformation.
  • Parameter Optimization: For the first case, $\sigma = 1.5$ (without the incompressibility constraint) resulted in the lowest maximum displacement error.
  • Computational Efficiency: Down-sampling the data to a resolution of 88x117x117 with an isotropic voxel size of 1.02 mm reduced processing time from days to hours with minimal loss in accuracy.
  • Strain Estimation: The incompressibility constraint led to more realistic strain values (ranging from -15% to 25%) compared to the non-incompressible case (150% to 300%). This suggests that while it doesn't improve displacement accuracy, it can lead to more physiologically plausible strain estimates.
  • Qualitative Agreement: Visual comparisons (Fig. 8) showed good agreement between the algorithm's tracked contours and the ground truth contours at maximum contraction, indicating the algorithm captures realistic deformations.

Limitations and Future Work

  • Model Incompressibility: The primary limitation identified is that the incompressibility constraint does not improve results because the synthetic data's underlying model is not incompressible. This suggests that the benefit of this constraint might be limited to datasets generated by truly incompressible models.
  • Error Accumulation: The paper notes that error accumulation can occur during maximum contraction, particularly in frame-to-reference tracking. Future work should focus on addressing this issue.
  • Parameter Tuning: While $\sigma = 1.5$ was found to be optimal for the first case, further investigation into parameter tuning across a wider range of datasets and configurations would be beneficial.
  • Real-world Data: The evaluation was performed on synthetic data. Future work should extend this evaluation to real ultrasound sequences to assess performance in more challenging and less controlled environments.

Summary updated

BibTeX
@inproceedings{prakosa2012evaluation,
  author = {Adityo Prakosa and Kristin McLeod and Maxime Sermesant and Xavier Pennec},
  title = {{Evaluation of iLogDemons Algorithm for Cardiac Motion Tracking in Synthetic Ultrasound Sequence}},
  booktitle = {STACOM},
  year = {2012},
  doi = {10.1007/978-3-642-36961-2\_21},
}
Figure