VBoussot/synthrad2025-impact-registration

🧭 SynthRAD2025 IMPACT Registrations (BSpline Transforms)

This repository provides Elastix B-spline transformation parameter files generated using the **IMPACT** method on the **SynthRAD2025** dataset.

Each file corresponds to a non-rigid registration between a reference CT and another modality (MRI or CBCT), aligned into CT space using Elastix with the IMPACT similarity metric.

• —Task 1: 411 transforms (102 excluded cases)
• —Task 2: 638 transforms (136 excluded cases)
• —Excluded: All cases from Center D (restricted to challenge use only)

🚀 Overview

High-quality multimodal registration is **essential for supervised sCT generation**. Inaccurate alignment between MRI/CBCT and CT images can lead to blurred, anatomically inconsistent, or artifact-prone synthetic CTs.

By leveraging features from pretrained segmentation models, IMPACT improves the anatomical consistency of cross-modality alignments, ensuring that each voxel correspondence reflects a true anatomical match.

The B-spline transforms provided here can be directly applied to warp MRI or CBCT images into CT space for training or evaluation of sCT generation models.

🔎 B-spline Transform Details

All registrations were performed using a 3rd-order B-spline transform with a final grid spacing of 10 mm across 4 resolution levels. The IMPACT loss was configured as a multi-metric combination of MIND and M730 features extracted from the final network layers.

🔧 Usage

To apply a transformation, use Transformix (from Elastix):

transformix -in Task1/HN/1HNA013/mr.mha -tp Task_1/HN/1HNA013.txt -out output/

Where:

• —mr.mha → Input image (MRI or CBCT) from the SynthRAD2025 dataset (not included here)
• —Task_1/HN/1HNA013.txt → B-spline transformation file from this repository
• —output/ → Directory where the warped image will be saved

📂 Repository Structure

SynthRAD2025_IMPACT_Registrations/
├── Task_1/
│   ├── HN/
│   │   ├── 1HNA013.txt
│   │   ├── 2HNA015.txt
│   │   └── ...
│   └── AB/
│   │   ├── 1ABA011.txt
│   │   └── ...
│   └── TH/
│   │   ├── 1THA011.txt
│   │   └── ...
│   │── Exclude.txt
└── Task_2/
    ├── HN/
    ├── AB/
    └── TH/
    │── Exclude.txt

• —Task 1: MRI → CT registrations
• —Task 2: CBCT → CT registrations
• —All transforms are in standard Elastix parameter file format (.txt)

⚠️ Restrictions

🏥 Center D

The SynthRAD2025 dataset includes multiple centers. Due to dataset licensing restrictions, Center D data are limited to challenge use only and cannot be redistributed. Corresponding B-spline transforms are therefore excluded from this release.

🗑️ Excluded Cases

102 cases were excluded from Task 1 due to poor image quality. The list of excluded cases is provided in Task_1/Exclude.txt.

📚 References

If you use these transformations, please cite the following works:

1. IMPACT Method Boussot V., Hémon C., Nunes J.-C., Dowling J., Rouzé S., Lafond C., Barateau A., Dillenseger J.-L.

IMPACT: A Generic Semantic Loss for Multimodal Medical Image Registration. arXiv:2503.24121, 2025. https://arxiv.org/abs/2503.24121

2. SynthRAD2025 Dataset Thummerer A., van der Bijl E., Galapon A. Jr., Kamp F., Savenije M., Muijs C., Aluwini S., Steenbakkers R.J.H.M., Beuel S., Intven M.P.W., Langendijk J.A., Both S., Corradini S., Rogowski V., Terpstra M., Wahl N., Kurz C., Landry G., Maspero M.

SynthRAD2025 Grand Challenge Dataset: Generating Synthetic CTs for Radiotherapy from Head to Abdomen. arXiv:2502.17609, 2025. https://arxiv.org/abs/2502.17609

3. Registration for sCT Synthesis Boussot V., Hémon C., Nunes J.-C., Dillenseger J.-L.

Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration. arXiv:2510.21358, 2025. https://arxiv.org/abs/2510.21358

🧠 License

All transformation files are released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). You may reuse, modify, and redistribute them for non-commercial research purposes only, with appropriate attribution.

https://creativecommons.org/licenses/by-nc/4.0/