Centerline-Guided Atlas Registration

Code for centerline-guided mesh registration of deformed ex vivo organ surfaces to 3D anatomical atlases while preserving local geometry.

Overview

This repository contains the centerline-based registration component used in the manuscript:

Globally Deformable, Locally Rigid Registration of Ex vivo Pancreas to a 3D Atlas Using a Centerline-Based Method

The method was developed to map a deformed ex vivo pancreas surface to a 3D atlas using a globally deformable, locally rigid transformation. In the full study workflow, a specimen surface is reconstructed from smartphone video using a Neural Radiance Field (NeRF), scaled using ex vivo CT, and then transformed to atlas space through cylindrical coordinates defined by corresponding centerlines.

This repository focuses on the registration and geometric evaluation stages of that workflow. It does not implement wet-lab cut mapping directly. Instead, it registers the ex vivo pancreas surface to atlas space and provides a common anatomical frame that can support downstream mapping of cuts, blocks, and slides.

What this repository contains

This repository is intended to host the centerline registration stage of the workflow, including:

• centerline resampling and processing,
• vertex-to-centerline correspondence,
• cylindrical-coordinate transformation,
• mesh warping from specimen space to atlas space,
• radial distortion analysis utilities,
• simple YAML-driven runner scripts for the paper pipeline.

Related repositories

The NeRF-based surface reconstruction pipeline used upstream in the study is available separately:

• https://github.com/MASILab/NerfOrganRecon

Method summary

The core idea is to:

• extract a centerline from the ex vivo mesh,
• extract a centerline from the atlas mesh,
• establish a one-to-one correspondence between the two centerlines,
• represent each mesh vertex in a local coordinate system relative to the source centerline,
• transfer that representation to the atlas centerline.

This lets the specimen bend toward the atlas trajectory while preserving local radial geometry relative to the centerline.

Workflow summary

After NeRF scaling and rough alignment, centerlines are extracted externally using VMTK within 3D Slicer. The resulting source and atlas centerlines are saved as .ply point clouds and passed back to the Python registration script for resampling and centerline-guided mesh transformation.

The paper workflow is:

1. CT preprocessing and mesh generation
2. NeRF mesh scaling and rough alignment
3. Manual step: centerline extraction in VMTK / 3D Slicer
4. Centerline-guided registration to atlas space
5. Radial distortion and smoothing sensitivity analysis

Given:

• an ex vivo specimen mesh,
• a target atlas mesh,
• a source centerline,
• a target centerline,

the method:

1. Resamples the specimen and atlas centerlines to matching point counts.
2. Assigns each specimen vertex to its nearest centerline sample.
3. Expresses each vertex in a local cylindrical frame:
   • radial distance r
   • angular position theta
   • tangent-direction offset h
4. Transfers (r, theta, h) to the corresponding atlas centerline frame.
5. Reconstructs the transformed vertex in atlas space.

This produces a deformation that is globally flexible but locally constrained by the centerline frame.

Manuscript context

In the current study, the workflow was applied to a single ex vivo human pancreas specimen. The wet-lab sampling plan consisted of 25 tissue blocks and 100 sub-blocks, and the specimen centerline was resampled to 26 points to reflect the sequential dissection plan.

The registered geometry was then used to place specimen-derived locations into atlas space, providing the geometric basis for downstream localization of tissue blocks and histology.

Geometric side effects were evaluated by measuring vertex-wise radial deviation before and after transformation. Moderate Laplacian smoothing (approximately 4–6 iterations) provided the best trade-off between surface stabilization and minimizing smoothing-induced shrinkage.

Inputs

Expected inputs include:

• specimen surface mesh,
• atlas surface mesh,
• specimen centerline,
• atlas centerline,
• optional smoothing parameters.

Outputs

Expected outputs include:

• transformed specimen mesh in atlas space,
• transformed specimen centerline,
• centerline correspondence files,
• radial distortion measurements,
• summary tables and figures for quality control.

Downstream mapping of wet-lab cuts or histology can be built on top of these outputs, but is not the main purpose of this repository.

Data acquisition used in the paper

For the feasibility study described in the manuscript, specimen video was acquired manually using the main rear camera of an iPhone 14 Pro at:

• 1920 × 1080 resolution
• 30 frames/s
• H.264 encoding
• ~180.8 s duration

The upstream NeRF mesh was scaled using ex vivo CT before centerline-based registration.

Installation

uv venv --python 3.10
source .venv/bin/activate
uv pip install -e .

How to run

python scripts/run_pipeline.py --config configs/pancreas_example.yaml --stage pre_vmtk

Extract the centerlines manually in VMTK / 3D Slicer and save them to the paths listed in the YAML config:

• paths.nerf_centerline
• paths.atlas_centerline

Then run registration and analysis:

python scripts/run_pipeline.py --config configs/pancreas_example.yaml --stage post_vmtk
python scripts/run_sensitivity_analysis.py --config configs/pancreas_example.yaml

You can also run stages separately:

python scripts/run_ct_preprocess.py --config configs/pancreas_example.yaml
python scripts/run_scale_align.py --config configs/pancreas_example.yaml
python scripts/run_registration.py --config configs/pancreas_example.yaml
python scripts/run_sensitivity_analysis.py --config configs/pancreas_example.yaml

Repository structure

centerline2atlas/
├── README.md
├── environment.yml
├── requirements.txt
├── configs/
├── scripts/
├── src/centerline2atlas/
└── figures/manuscript/

Citation

If you use this code, please cite the associated manuscript:

@article{hucke2026centerline,
  title={Globally Deformable, Locally Rigid Registration of Ex vivo Pancreas to a 3D Atlas Using a Centerline-Based Method},
  author={Hucke, Andre T.S. and Kim, Michael E. and Remedios, Lucas W. and others},
  journal={TBD},
  year={2026}
}