Coercive Region-level Registration for Multi-modal Images

Yu-Hui Chen, Dennis L. Wei, Gregory E. Newstadt, Jeffrey P. Simmons, Alfred O. Hero Iii

Abstract

This paper introduces a "coercive" framework that simultaneously segments and registers multi-modal microscopy images at the region level. By forcing segmentations from different sensors (such as BSE and EBSD) to be topologically similar while respecting their unique statistical properties, the method avoids the artifacts of pixel-level interpolation and significantly improves boundary detection accuracy. Evaluation on simulated and real datasets demonstrates the framework's ability to recover missing boundaries and correct local misalignments.

Introduction

In materials science, Scanning Electron Microscopy (SEM) is frequently used to characterize samples through various modalities. Two primary techniques are Backscattered Electrons (BSE), which provide scalar intensity values related to surface texture and composition, and Electron Backscatter Diffraction (EBSD), which maps crystal orientation data. Combining these images is difficult due to differing data types, local distortions from beam geometry, and interpolation bias. Standard registration methods often blur sharp grain boundaries by interpolating values between pixels.

Methodology

The proposed approach employs a "coercive" objective function that balances intra-modal data fit with inter-modal topological consistency.

Statistical Modeling

BSE intensities are modeled using a Gaussian distribution, while EBSD orientation data (represented as quaternions) are modeled via a von Mises-Fisher (vMF) mixture model to account for crystal symmetry.

Coercive Framework

The framework minimizes an energy function consisting of:

  • Intra-modal Energy: Measures how well the segmentation fits the raw data of each modality.
  • Inter-modal Energy: A penalty term that "coerces" the two segmentations to share the same topology (the same number and arrangement of grains).
Visual challenges in multi-modal registration: (a) misalignment and (b) missing edges.

To handle specific artifacts, the authors developed a Generalized Likelihood Ratio Test (GLRT) to identify missing edges and an adaptive misalignment test based on the Rayleigh distribution. These tests allow the algorithm to distinguish between genuine structural differences and simple alignment errors.

Results and Discussion

The framework was evaluated on simulated data from the "Dream3D" toolbox and real IN100 Nickel-base alloy microscopy data.

Quantitative Performance

In simulations, the coercive method achieved a significantly higher overlapping rate—a measure of boundary precision—compared to independent segmentation or standard B-spline registration. The results indicate that the coercive approach maintains superior accuracy across varying boundary widths.

Overlapping rate comparison: The coercive method significantly outperforms independent segmentation and B-spline registration.

Qualitative Boundary Recovery

On real IN100 data, the algorithm successfully identified "missing" boundaries in EBSD data by leveraging the BSE signal. By treating the image as a collection of regions rather than a grid of pixels, the method avoids ad-hoc smoothing and exploits the synergy between modalities.

Conclusion

The coercive region-level registration framework provides a robust solution for aligning multi-modal microscopy data. By integrating segmentation and registration into a single statistical framework, it ensures physical consistency and improves boundary recovery. Future work will address initialization dependencies and refine geometric assumptions for complex grain shapes.

Summary updated

BibTeX
@inproceedings{chen2015coercive,
  author = {Yu-Hui Chen and Dennis L. Wei and Gregory E. Newstadt and Jeffrey P. Simmons and Alfred O. Hero III},
  title = {{Coercive Region-level Registration for Multi-modal Images}},
  booktitle = {ICIP},
  year = {2015},
  url = {https://openreview.net/forum?id=fK5vfTR9gUTL},
}
Figure