Symmetric Algorithmic Components for Shape Analysis with Diffeomorphisms

Nicolas Guigui, Shuman Jia, Maxime Sermesant, Xavier Pennec

Abstract

In computational anatomy, diffeomorphic registration is essential for mapping anatomical variations. However, standard Large Deformation Diffeomorphic Metric Mapping (LDDMM) often suffers from numerical drift because registration residuals—the unavoidable matching errors—are typically ignored in subsequent geometric operations. This paper introduces a symmetric framework that incorporates these residuals directly into fundamental geometric operators. By redefining midpoints and symmetries to account for registration error, the authors achieve higher numerical consistency in parallel transport and statistical shape analysis. Evaluation on cardiac meshes demonstrates significantly improved consistency and more reliable statistical mapping of anatomical features.

Introduction

The study of anatomical shape variation relies on the ability to "warp" one shape into another smoothly and reversibly. While LDDMM provides a robust mathematical framework for these deformations, practical implementations are limited by regularization. This results in a "matching error" or residual between the warped template and the target. When these residuals are ignored during geometric steps like parallel transport (moving a deformation from one subject to a common atlas), numerical errors accumulate, compromising the reliability of clinical statistical comparisons.

Methodology: A Symmetric Approach to Residuals

The authors propose treating the space of deformations as a symmetric space. To handle numerical imperfections, they decompose a target shape $S$ into a deformation part $\phi(T)$ and a residual displacement field $\delta$.

Modified Geometric Operators

The core innovation lies in redefining the building blocks of shape geometry:

  • Midpoints: Instead of a purely geodesic midpoint, the new operator calculates the halfway point of the geodesic plus half of the residual displacement.
  • Involution (Symmetry): The symmetry operator is adjusted to include the residual, ensuring that applying the operation twice more accurately returns to the starting point.

Consistency Metrics

To evaluate the framework, the authors define three key geometric properties:

  1. Centrality: Measures how accurately a midpoint sits between two shapes.
  2. Involutivity: Measures the error when a symmetry operation is applied twice.
  3. Transvectivity: The composition of two symmetries, serving as a proxy for the gradient of the curvature of the deformation space.
Geometric definitions of Centrality (a), Involutivity (b), and Transvectivity (c) used to evaluate algorithmic consistency.

Results and Discussion

The method was evaluated using a dataset of 138 cardiac meshes, focusing on the deformation between end-diastole and end-systole.

Numerical Consistency

The inclusion of residuals significantly improved numerical stability. For a standard regularization parameter ($\alpha^2 = 1$), the centrality error was reduced from 0.50mm to 0.10mm. As shown in the boxplots below, the "withResidual" method consistently outperforms the standard "noResidual" approach across various regularization scales ($\alpha^2$).

Quantitative comparison of Involutivity and Transvectivity errors. The proposed 'withResidual' method (blue) shows significantly lower error and higher stability across different regularization parameters compared to the standard 'noResidual' method (orange).

Parallel Transport of Cardiac Strain

By using a "Symmetric Pole Ladder" informed by these modified operators, the researchers transported "area strain" measurements from individual subjects to a population atlas. The residual-aware approach produced smoother maps with fewer numerical artifacts, particularly in complex regions like the heart valves. The results also highlight an "optimal" regularization range where the space behaves most symmetrically, balancing warp smoothness with residual magnitude.

Conclusion

This work demonstrates that registration residuals should be treated as informative features rather than mere noise. By integrating these errors into a symmetric geometric framework, researchers can achieve more consistent and reliable shape analysis. Future work may extend this to 3D medical images, where residuals could account for intensity biases in addition to geometric misalignments.

Summary updated

BibTeX
@inproceedings{guigui2019symmetric,
  author = {Nicolas Guigui and Shuman Jia and Maxime Sermesant and Xavier Pennec},
  title = {{Symmetric Algorithmic Components for Shape Analysis with Diffeomorphisms}},
  booktitle = {GSI},
  year = {2019},
  doi = {10.1007/978-3-030-26980-7\_79},
}
Figure