[Paper Review] Coercive Region-level Registration for Multi-modal Images

Yu‐Hui Chen, Dennis Wei+3 more|arXiv (Cornell University)|Feb 26, 2015

Medical Image Segmentation Techniques16 references3 citations

TL;DR

This paper proposes a coercive region-level registration and segmentation framework for multi-modal microscopy images, jointly optimizing alignment and segmentation using statistical models and hypothesis testing to detect missing boundaries. It outperforms state-of-the-art methods by eliminating interpolation, preserving region correspondences, and achieving superior boundary detection accuracy on synthetic and real SEM data.

ABSTRACT

We propose a coercive approach to simultaneously register and segment multi-modal images which share similar spatial structure. Registration is done at the region level to facilitate data fusion while avoiding the need for interpolation. The algorithm performs alternating minimization of an objective function informed by statistical models for pixel values in different modalities. Hypothesis tests are developed to determine whether to refine segmentations by splitting regions. We demonstrate that our approach has significantly better performance than the state-of-the-art registration and segmentation methods on microscopy images.

Motivation & Objective

• To address the challenge of registering and segmenting multi-modal microscopy images—such as BSE and EBSD—where pixel-level methods suffer from interpolation bias and segmentation fails due to low contrast.
• To develop a joint registration-segmentation framework that preserves topological correspondences between regions across modalities for effective data fusion.
• To detect and correct missing grain boundaries in low-contrast modalities using statistical hypothesis tests grounded in pixel value distributions.
• To eliminate reliance on interpolation by performing registration at the region level, improving robustness to resolution and intensity differences between modalities.

Proposed method

• The method minimizes a composite objective function combining intra-modal fitting (J(S_i, I_i)) and inter-modal topological similarity (D(S_1, S_2)), with a regularization parameter λ.
• Statistical models are used for pixel values: Gaussian for BSE and von Mises-Fisher for EBSD, enabling probabilistic segmentation and region-level alignment.
• Alternating minimization is performed between segmentation and registration, refining segmentations iteratively using region-level correspondences.
• A hypothesis test based on Rayleigh-distributed displacement d is used to detect missing boundaries; the test threshold is adaptively set using a user-defined false positive rate α.
• The inter-modal energy D(S_1, S_2) is defined as the number of segment boundaries present in one modality but not the other, promoting topological consistency.
• Boundary detection is performed by testing whether the observed displacement between corresponding region edges exceeds a statistical threshold derived from the Rayleigh distribution.

Experimental results

Research questions

• RQ1 Can joint region-level registration and segmentation outperform independent segmentation and pixel-level registration in multi-modal microscopy images?
• RQ2 How can missing grain boundaries in low-contrast modalities be reliably detected using statistical models of pixel intensities?
• RQ3 To what extent does region-level registration without interpolation improve alignment accuracy compared to interpolation-based pixel-level methods?
• RQ4 Can hypothesis testing based on statistical distributions of inter-region displacements effectively identify and correct misaligned or missing boundaries?
• RQ5 How well does the method preserve topological correspondences between regions across modalities for downstream data fusion?

Key findings

• The proposed method achieved significantly higher boundary overlapping rates than independent segmentation and B-spline registration, especially for narrow boundaries (w = 1–3 pixels), demonstrating superior detection of misaligned and missing edges.
• On simulated data, the coercive approach achieved an overlapping rate of over 90% for boundary width w = 3, outperforming the next-best method by more than 20 percentage points.
• In real IN100 data, the method successfully realigned EBSD boundaries (blue lines) to better match BSE image intensities, while detecting previously missing boundaries (green lines) within grain regions.
• The hypothesis test with α = 0.05 and σ_d = 3 pixels correctly identified boundary misalignments and omissions, with thresholds adaptively set based on equivalent radius r and false positive rate.
• The method preserved all region correspondences across modalities, enabling robust data fusion and eliminating interpolation artifacts common in pixel-level methods.
• The algorithm demonstrated robust performance on real materials data, accurately registering grain boundaries and improving segmentation quality in both BSE and EBSD modalities.