RegistrationBias-sCT: checkpoints, registrations and simulated CBCT

Model weights, registration files and simulated CBCT of “When Misalignment Becomes Supervision: Structured Label Noise in Supervised Synthetic CT Generation” (arXiv:2609.29387).

The code, the configurations, the patient-level results and the step-by-step reproduction guide are on GitHub: vboussot/RegistrationBias-sCT.

Content

checkpoints/Task_{1,2}/{ELX,IMPACT}/{MAE,SAM,VGG}/CV_{0..4}.pt      50 checkpoints, 15.7 GB
checkpoints/manifest.csv                                            size and SHA-256 of each checkpoint
transforms/synthrad2025-elx-registration/Task_{1,2}/{AB,HN,TH}/     169 Elastix transforms, 0.2 GB
sim-cbct/{AB,HN,TH}/{case}.mha                                      103 simulated CBCT, 0.5 GB

Checkpoints

Synthetic CT generators: a 2.5D U-Net++ with a ResNet-34 encoder (five adjacent axial slices, 26 M parameters), trained with KonfAI on the SynthRAD2025 abdomen, head-and-neck and thorax cases.

Path element	Values	Meaning
Task_1, Task_2		MR-to-CT, CBCT-to-CT
ELX, IMPACT		registration used to build the training pairs: organizers' Elastix, or IMPACT-Reg
MAE, SAM, VGG		training loss: MAE alone, MAE + SAM 2.1 features, MAE + VGG (IMPACT pairs only)
CV_0 … CV_4		cross-validation fold; each file is the checkpoint with the lowest validation MAE of its fold

That is 2 tasks × (ELX: MAE, SAM; IMPACT: MAE, SAM, VGG) × 5 folds = 50 files, about 313 MB each. A prediction of the paper averages the five folds of a model, each applied to the image and two flipped copies.

ELX transforms

Elastix B-spline transform parameter files of the 169 held-out SynthRAD2025 cases (Task 1: 66, Task 2: 103), one {case}.txt per case. They register the planning CT to the MR or CBCT and are computed with the organizers' parameter files (SynthRAD2025/preprocessing, commit 8a5b125).

The IMPACT-Reg transforms are in VBoussot/synthrad2023-impact-registration and VBoussot/synthrad2025-impact-registration.

Simulated CBCT

The Sim-CBCT test set of the paper: one CBCT simulated from the planning CT of each of the 103 held-out SynthRAD2025 Task 2 cases (abdomen 32, head and neck 37, thorax 34) by RTK forward projection, noise, scatter and FDK reconstruction (scripts/preprocessing/cbct_synthesis.py in the GitHub repository). Each volume is on the grid of its CT, in HU, and set to -1024 outside the patient mask. The CT and the masks are in the SynthRAD2025 training set.

License

The checkpoints and the ELX transforms are released under Apache-2.0. The simulated CBCT derive from SynthRAD2025 images and are released under their license, CC BY-NC 4.0.

Use

The GitHub repository downloads these files to the right place and checks the SHA-256 of each checkpoint:

git clone https://github.com/vboussot/RegistrationBias-sCT && cd RegistrationBias-sCT
python scripts/download.py --only checkpoints elx sim-cbct
python scripts/predict.py --gpu 0                      # see docs/REPRODUCE.md

One model alone:

hf download VBoussot/RegistrationBias-sCT --include "checkpoints/Task_1/IMPACT/MAE/*" --local-dir .

Citation

@article{boussot2026misalignment,
  title   = {When Misalignment Becomes Supervision: Structured Label Noise in Supervised Synthetic CT Generation},
  author  = {Boussot, Valentin and H{\'e}mon, C{\'e}dric and Lafond, Caroline and Nunes, Jean-Claude and Dillenseger, Jean-Louis},
  journal = {arXiv preprint arXiv:2609.29387},
  year    = {2026}
}