Anatomy-aware, label-informed approach improves image registration for challenging datasets

Rachel A. Roston, Nicholas J. Tustison, A. Murat MagabioRxiv 2025 · DOI

TL;DR

This paper introduces a novel "label-informed" image registration method that uses anatomical segmentations (labels) to guide the alignment of medical images. This approach significantly improves registration accuracy for datasets with substantial morphological variation, such as developmental abnormalities, which are challenging for traditional intensity-only methods. The improved registration unlocks new possibilities for quantitative morphological analysis in diverse biological studies.

What Problem Does This Paper Address?

Quantitative morphometrics, which uses image registration to measure shape differences, is a powerful tool for understanding biological variation. However, a major hurdle arises when dealing with datasets exhibiting significant morphological differences, often seen in developmental studies or comparisons across species. Traditional image registration methods, which rely solely on pixel intensity values to align images, struggle with these "overt phenotypes" – large variations in size, shape, orientation, or even missing structures. This failure in registration renders subsequent quantitative analyses unreliable or impossible, limiting the scope of morphometric studies. The authors aim to overcome this by developing a more robust registration method that can handle such challenging datasets.

Methods

The core innovation is the label-informed image registration function within the ANTsX ecosystem. Unlike traditional methods that solely use intensity information, this new approach incorporates anatomical labels (segmentations of organs or regions) to guide the registration process. These labels, which can be generated manually, semi-automatically, or via deep learning, provide a priori correspondences. The registration process is multi-stage:

  1. Initialization: Uses the centers of mass of corresponding label maps for an initial alignment.
  2. Deformable Registration: Employs intensity-based similarity metrics, but crucially, the labels act as "biologically meaningful anchors" to guide the deformation.
  3. Optional Refinement: Can include label-based similarity metrics.

This contrasts with intensity-only registration, which relies on initial linear transformations (rigid, affine) based on intensity data to achieve a coarse alignment before deformable registration.

Datasets

The study utilizes micro-CT scans of E15.5 mouse embryos. Specifically, they compare wildtype embryos (Gli2+/+) with knockout embryos (Gli2-/-) exhibiting severe scoliosis and internal organ rearrangements. The template image used for registration was generated by averaging thirty-five wildtype micro-CT volumes from various inbred and outbred mouse lines.

Key Insights and Ideas

  • Labels as Anatomical Anchors: The fundamental idea is that anatomical labels provide crucial structural information that intensity alone cannot capture, especially when shapes deviate significantly. These labels act as reliable landmarks to guide the registration.
  • Bridging Segmentation and Morphometrics: This method effectively bridges the gap between segmentation workflows and voxel-based morphometrics, allowing segmentations to actively inform, rather than just be a downstream output of, the registration process.
  • Generalizability: The label-informed approach is presented as a general solution applicable to a wide range of morphological variations, not just specific to the tested phenotype.
  • Improved Statistical Power: By achieving better alignment, the method leads to more sensitive and powerful downstream statistical analyses, revealing differences that were previously obscured.

Key Results

  • Qualitative Improvement: Label-informed registration produced significantly better visual alignment of knockout embryos to the wildtype template compared to intensity-only registration.
  • Tensor-Based Morphometry (TBM) Enhancement: Label-informed TBM identified 50% more statistically significant voxels with volume differences (FDR < 0.05) compared to intensity-only TBM. It also increased the range of Jacobian magnitudes by 45%, indicating greater sensitivity to local volume changes.
  • Reflected Anatomy: TBM results from label-informed registration more accurately reflected observable anatomical differences, including subtle shape variations not detected by intensity-only methods. For instance, it captured localized volume reductions in the liver and spinal cord that corresponded to observable morphological changes.
  • Principal Component Analysis (PCA) Separation: PCA on intensity-only registration failed to separate wildtype and knockout phenotypes. In contrast, PCA on label-informed registration transforms clearly separated the genotypes, with PC2 specifically capturing the scoliosis phenotype.

Limitations and Future Work

  • Label Dependency: The quality and selection of labels are critical. Poorly chosen or distributed labels can skew registration. Corresponding structures must be present in all images being registered, limiting applications where structures develop at different stages.
  • Scaling Consistency: The study notes that scaling from label-informed registration might differ from intensity-only methods and between different datasets, suggesting that scaling interpretation needs careful consideration.
  • Future Work: The authors suggest that label-guided atlas-based segmentation could be improved by using label-informed registration to transfer labels from an atlas. They also highlight the potential for this method in developmental and comparative morphometrics, enabling studies across species and developmental stages where homologous regions can be identified.

Summary updated

Abstract

Image registration-based volumetric morphometrics have emerged as a valuable method for identifying subtle morphological differences in neuroimaging and other biomedical images. However, accurate registration out-of-the-box remains challenging when overt morphological phenotypes-such as those observed in developmental and comparative studies-are present in a dataset. A new label-informed image registration function developed in the ANTsX ecosystem provides an easy to use, generalizable solution for anatomy-aware registration of a wide diversity of morphological variation. In this approach, segmentations (<i>i.e.</i>, labels) provide <i>a priori</i> regional correspondences that guide the registration. These labels can be generated by any method-manually, using semi-automated tools, or through deep learning-based approaches-and allow morphological experts to define regions of correspondence based on biological concepts of homology (<i>e.g.</i>, tissue origin, gene expression patterns). Here we demonstrate the utility of this label-informed image registration approach for improving the registration knockout mouse embryos which fail to register to a wildtype (normative) template image by traditional registration methods. E15.5 <i>Gli2</i> <sup><i>-/-</i></sup> mouse embryos show a severe scoliosis and radical topological rearrangement of the internal organs. Compared to traditional, intensity-only registration, the new label-informed image registration improved the correspondence of knockout subjects to the canonical template image, which resulted in increased power and sensitivity of downstream statistical analyses. All in all, label-informed image registration provides a flexible and customizable method to allow image registration in datasets for which registration-based morphometrics were previously unfeasible, unlocking new potential applications of registration-based morphometrics in developmental, comparative, and evolutionary studies.
BibTeX
@misc{roston2025anatomy,
  author = {Rachel A. Roston and Nicholas J. Tustison and A. Murat Maga},
  title = {{Anatomy-aware, label-informed approach improves image registration for challenging datasets}},
  year = {2025},
  doi = {10.1101/2025.08.11.669599},
  howpublished = {bioRxiv},
}

Loading related content...

Figure