Moment evolution equations and moment matching for stochastic image EPDiff

Alexander Christgau, Alexis Arnaudon, Stefan Sommer

TL;DR

This paper introduces a novel framework for estimating noise parameters in stochastic image deformation models. By deriving first-order moment evolution equations for the stochastic EPDiff framework, the authors enable efficient parameter inference via automatic differentiation, providing a practical way to model random subject-specific variations in longitudinal medical imaging.

What Problem Does This Paper Address?

In medical imaging, researchers often use the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework to model how shapes—like a developing brain or an aging heart—change over time. While standard LDDMM is deterministic, real-world biological processes involve significant random variation.

Previous work introduced stochastic versions of these models (stochastic EPDiff) to account for this "noise," but estimating the parameters of that noise from actual image data is notoriously difficult. The challenge stems from the infinite-dimensional nature of images and the complex, non-linear way that deformations interact with image intensity. This paper seeks to bridge the gap between theoretical stochastic models and practical statistical inference.

Methods

The authors propose a "method of moments" approach to parameter estimation. Their workflow involves several key technical steps:

  1. Stochastic Formulation: They start with the SEPDA equations (Stochastic EPDiff and Advection), which describe how momentum and image intensity evolve under random perturbations.
  2. Moment Derivation: Because these equations are stochastic (specifically Stratonovich SDEs), the authors convert them to Itô form to derive equations for the expected (average) evolution of the image and momentum.
  3. First-Order Approximation: Exact moment equations are impossible to solve because they involve higher-order correlations. The authors use a first-order "cluster expansion" approximation, essentially assuming that the expectation of a product is roughly the product of the expectations.
  4. Moment Matching: They construct an estimator by matching the approximated expected image to the empirical mean of observed sample images.
  5. Implementation: The system is implemented using the mermaid library and PyTorch, leveraging automatic differentiation and adjoint methods to optimize the noise parameters efficiently without massive memory overhead.

Datasets

The study relies on synthetic datasets generated from real brain MR-slices.

  • Simulation: They generated datasets of 500 images each by numerically integrating the SEPDA equations using a Heun scheme.
  • Noise Models: They tested various parametric submodels for the noise fields, including:
  • Gaussian kernels arranged in square and hexagonal lattices.
  • Cubic B-splines (which have compact support).
  • Sinusoidal fields (sine series).

Key Insights and Ideas

  • The Power of Averages: Even with a coarse first-order approximation, matching the "average" deformed image is sufficient to recover the underlying spatial correlation of the noise.
  • Geometry Matters: The choice of lattice for noise fields significantly impacts accuracy. Hexagonal lattices outperformed square lattices for brain images because they provide a more uniform cover of the central domain and avoid "corner" inaccuracies.
  • Interference vs. Identification: Noise fields with compact support (like B-splines) are easier to estimate than those with global support (like Gaussians or Sines). This is because compact fields interfere with each other less, making their individual contributions to the total deformation easier for the optimizer to "untangle."
  • Invariance: The use of Normalized Cross-Correlation (NCC) as a similarity measure proved more effective than standard squared distances (SSD) because it is invariant to global changes in pixel intensity.

Key Results

  • Identifiability: In Experiment A, the model successfully recovered ground-truth parameters. The B-spline hexagonal lattice achieved the lowest relative error (0.0262), while the sinusoidal model struggled significantly (error of 0.320).
  • Robustness: In Experiment B (model misspecification), the framework proved capable of approximating arbitrary, non-lattice noise fields, though the accuracy was naturally limited by the resolution of the chosen grid.
  • Efficiency: The optimization procedures generally converged to accurate estimates within approximately 80 iterations.

Limitations and Future Work

  • Higher-Order Moments: The current work uses only first-order approximations. The authors suggest that including second-order moments could improve the separation of interfering noise fields, though the math would be considerably more technical.
  • Grid Resolution: The accuracy of noise estimation is currently tied to the density of the predefined grid. Future work could involve optimizing the positions and widths of the noise fields themselves, rather than just their amplitudes.
  • Dimensionality: The simulations were performed in 2D. While the mermaid implementation supports 3D, the computational cost of 3D stochastic image registration remains a hurdle for future research.
  • Real-World Application: The next logical step is applying this to real longitudinal datasets (e.g., tracking disease progression in patients) where the "ground truth" noise is unknown.

Summary updated

Abstract

Models of stochastic image deformation allow study of time-continuous stochastic effects transforming images by deforming the image domain. Applications include longitudinal medical image analysis with both population trends and random subject specific variation. Focusing on a stochastic extension of the LDDMM models with evolutions governed by a stochastic EPDiff equation, we use moment approximations of the corresponding Itô diffusion to construct estimators for statistical inference in the full stochastic model. We show that this approach, when efficiently implemented with automatic differentiation tools, can successfully estimate parameters encoding the spatial correlation of the noise fields on the image.
BibTeX
@article{christgau2021moment,
  author = {Alexander Christgau and Alexis Arnaudon and Stefan Sommer},
  title = {{Moment evolution equations and moment matching for stochastic image EPDiff}},
  journal = {J. Math. Imaging Vis.},
  year = {2021},
  url = {https://openreview.net/forum?id=femNRAle_0},
}
Figure