---
license: apache-2.0
tags:
  - medical-imaging
  - ct
  - cbct
  - synthetic-ct
  - image-synthesis
  - nnunet
  - regression
  - torchscript
language: []
---

# SimCBCT Generator — Synthetic CT, Pelvis

A 3D regression model that generates synthetic CT (sCT) images from Cone-Beam CT (CBCT) scans of the PELVIS region. Part of the [SimCBCTGenerator](https://github.com/openvoxelmed/simcbctgenerator) framework, which was used to generate the training data for this model (aligned input/output pairs by simulating CBCT data).

The model is trained with an nnUNet regression pipeline (`nnUNetTrainerRegression_mae_deep`, 3D full-resolution).

## Intended Use

Convert acquired CBCT images to synthetic CT for pelvis anatomy. Typical applications include adaptive radiotherapy, where sCT is needed for dose recalculation or replanning directly from CBCT without a new planning CT acquisition. It was extensively tested for Elekta machines but initial inspection also results in good performance for Varian datasets!

## Model Files

| File | Description |
|---|---|
| `checkpoints/model.pt` | TorchScript compiled model (~390 MB) |
| `checkpoints/metadata.json` | Patch size, normalization stats, inference config |
| `sct_generator.py` | Self-contained inference class — no nnUNet install required |

The `model.pt` is a **TorchScript** module exported with `torch.jit.save()` and must be loaded with `torch.jit.load()`. 

## Requirements

```
torch
scipy
numpy
```


## Quick Start

```python
import numpy as np
from sct_generator import StandaloneRegressionInference

# Load model (pass the directory containing model.pt and metadata.json)
model = StandaloneRegressionInference(
    model_path="checkpoints/",
    device="cuda"   # or "cpu"
)

# cbct_volume: 3D numpy array of HU values, shape (D, H, W)
cbct_volume = np.load("your_cbct_volume.npy")

# Run inference — returns sCT volume in HU, same spatial shape as input
sct_output = model.predict(cbct_volume)
```

`predict()` handles internally:
- Z-score normalization of the CBCT input
- Sliding-window tiled inference with Gaussian patch blending
- Denormalization of the sCT output back to HU values with the precomputed statistics


## Training Details

| Field | Value |
|---|---|
| Trainer | `nnUNetTrainerRegression_mae_deep` |
| Configuration | `3d_fullres` |
| Fold | all |
| Loss | MAE |
| Input | Physics-based simulated CBCT |
| Output | CT |
| Anatomy | Pelvis |

## Citation

If you use this model or the SimCBCTGenerator framework, please cite:

```bibtex
@article{zimmermann2026simcbct,
  title   = {Eliminating Registration Bias in Synthetic CT Generation:
             A Physics-Based Simulation Framework},
  author  = {Zimmermann, Lukas and Rauter, Michael and Schmid, Maximilian
             and Georg, Dietmar and Kn\"{a}usl, Barbara},
  journal = {arXiv preprint arXiv:2602.02130},
  year    = {2026}
}
```

**Framework repository:** [github.com/openvoxelmed/simcbctgenerator](https://github.com/openvoxelmed/simcbctgenerator)