An extension of digital volume correlation for multimodality image registration

Erika Tudisco, Clément Jailin, Arturo Mendoza, Alessandro Tengattini, Edward Andò, Stephen A. Hall, Gioacchino Viggiani, François Hild, Stéphane RouxMeasurement 2017 · DOI

TL;DR

This paper extends Digital Volume Correlation (DVC) to register images from different modalities (like X-ray and neutron tomography), a crucial step for combining complementary information. They introduce two novel potential functions for this registration, one based on Gaussian mixtures and another on a log-likelihood approach, demonstrating their effectiveness on real 3D data.

What Problem Does This Paper Address?

Combining data from different imaging modalities (e.g., X-ray and neutron tomography) offers a richer understanding of materials by capturing complementary properties. However, these images are often acquired in different frames of reference and can have vastly different contrasts, making it challenging to align them accurately. This misalignment prevents a unified analysis and local characterization of material properties. Traditional image registration methods often struggle with these significant contrast differences. This paper aims to bridge this gap by developing a robust method to register images from different modalities, enabling their combined analysis.

Methods

The core of the proposed method is an extension of Digital Volume Correlation (DVC), a technique typically used for registering images from the same modality. The key innovation lies in defining a "potential" function, $\Phi$, that quantifies the similarity between images from different modalities. Two specific approaches for defining this potential are presented:

  • Gaussian Mixture (GM) Approach: This method models the joint distribution of gray levels between the two images as a mixture of Gaussian distributions. This approach assumes that distinct phases in the material will have characteristic gray level signatures in each modality. By fitting these Gaussian components, the method can also achieve a natural segmentation of the material into phases.
  • Log-Likelihood (LL) Approach: This method is more data-driven and does not rely on a pre-defined model for the joint distribution. Instead, it directly uses the observed joint histogram of gray levels and aims to maximize the likelihood of the observed data under the transformation. This approach is less prone to subjective choices about the number of phases but might be less robust to initial mismatches.

Both approaches are implemented within a multiscale framework (pyramidal scheme). This involves starting the registration at a coarse resolution and progressively refining it at finer resolutions. This hierarchical approach significantly improves robustness and helps avoid getting trapped in local minima, especially when the initial alignment is poor. The transformation itself is modeled as a combination of rigid body motions (translation and rotation) and isotropic dilation, allowing for up to 7 parameters in 3D.

Datasets

The authors used a 3D test case consisting of two volumes of the same Bentheim sandstone sample, acquired using X-ray tomography and neutron tomography. These datasets were chosen because they exhibit significantly different contrasts, posing a challenge for registration. The sample had specific properties: average porosity of 23%, mean grain diameter of 300 µm, and a composition of 95% quartz, 3% kaolinite, and 2% orthoclase. The voxel size for both modalities was approximately 30 µm.

Key Insights and Ideas

  • Bridging Modality Gaps: The fundamental insight is that DVC, a powerful tool for intra-modality registration, can be extended to inter-modality registration by carefully designing a potential function that accounts for differing contrast characteristics.
  • Potential Function Design is Key: The success of inter-modality registration hinges on the choice of the potential function. The paper demonstrates that both model-based (GM) and data-driven (LL) approaches can be effective, each with its own trade-offs.
  • Multiscale for Robustness: Employing a multiscale (pyramidal) approach is crucial for achieving robust registration, especially when dealing with significant initial misalignments or complex joint histograms. This hierarchical strategy allows the algorithm to converge from a wider range of initial conditions.
  • Segmentation as a Byproduct: A significant and valuable outcome of the GM approach is the inherent segmentation of the material into phases, directly derived from the registration process. This provides an additional layer of information without requiring separate segmentation steps.
  • Artifact Detection: The residual field analysis revealed an unexpected apparent strain between the X-ray and neutron images, potentially indicating a detector alignment issue. This highlights the power of such registration methods for quality control and artifact identification.

Key Results

  • Accurate Registration: Both the GM and LL approaches successfully registered the X-ray and neutron tomography volumes with sub-pixel accuracy.
  • Comparable Transformations: The estimated transformation parameters (strain, rotation, translation) from both GM and LL approaches were in good agreement, with strain components differing by less than 10⁻⁴ (except for εzz, which differed by about 1%).
  • GM Stability and Speed: The GM algorithm was found to be more stable and exhibited faster convergence within the multiscale framework.
  • LL Convergence Issues: The LL algorithm showed slower convergence, with higher iteration counts, particularly at coarser pyramid levels. Its initial registration estimates were also less reliable.
  • Segmentation: The GM approach provided a natural segmentation of the material into phases, visualized in Figure 4(b).
  • Residual Field Analysis: The residual maps (Figure 4(a) and 6(a)) indicated good convergence, with no visible bias, suggesting successful registration.

Limitations and Future Work

  • GM Simplification: The GM approach, while robust, might be an oversimplification of the true joint distribution, potentially leading to slightly larger uncertainties in the final registration compared to the LL approach.
  • LL Constraint: The LL approach, while data-driven, might not provide enough constraint, especially with limited initial data or if the joint histogram is not well-behaved.
  • Initialization Sensitivity: While the multiscale approach significantly improves robustness, the paper acknowledges that the maximum tolerable distance from the true solution for guaranteed convergence is still a limitation.
  • Future Work: The authors suggest that tailored potentials could be designed on demand for specific prior knowledge of the joint histogram. They also propose enriching the degrees of freedom for the transformation in cases where the specimen undergoes deformations (e.g., drying, swelling) between scans. The ability to enrich or correct low-resolution/noisy images from one modality with data from another is also highlighted as a promising avenue.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{tudisco2017extension,
  author = {Erika Tudisco and Clément Jailin and Arturo Mendoza and Alessandro Tengattini and Edward Andò and Stephen A. Hall and Gioacchino Viggiani and François Hild and Stéphane Roux},
  title = {{An extension of digital volume correlation for multimodality image registration}},
  booktitle = {Measurement},
  year = {2017},
  doi = {10.1088/1361-6501/aa7b48},
}
Figure