# TotalSynth: whole-body synthetic CT from MRI and CBCT in 3D Slicer (IMPACT-Synth and KonfAI updated)

**URL:** https://discourse.slicer.org/t/totalsynth-whole-body-synthetic-ct-from-mri-and-cbct-in-3d-slicer-impact-synth-and-konfai-updated/48242
**Category:** Development
**Created:** [September 23, 2026, 2:23pm UTC](https://discourse.slicer.org/t/totalsynth-whole-body-synthetic-ct-from-mri-and-cbct-in-3d-slicer-impact-synth-and-konfai-updated/48242 "2026-09-23T14:23:25Z")
**Posts on this page:** 1
**Page:** 1

<div class="post-metadata">

### Author: ![vboussot](https://avatars.discourse-cdn.com/v4/letter/v/839c29/32.png) [@vboussot](https://discourse.slicer.org/u/vboussot)
#### Post date: [September 23, 2026, 2:23pm UTC](https://discourse.slicer.org/t/totalsynth-whole-body-synthetic-ct-from-mri-and-cbct-in-3d-slicer-impact-synth-and-konfai-updated/48242/1 "2026-09-23T14:23:25Z")

</div>

Dear all,

In February we announced the [KonfAI](https://github.com/vboussot/SlicerKonfAI) and [IMPACT-Synth](https://github.com/vboussot/SlicerImpactSynth) extensions. Here is what is new.

**TotalSynth models**

We released **TotalSynth** , pretrained models for **synthetic CT from MRI and from CBCT** , from the head to the pelvis ([arXiv:2609.13838](https://arxiv.org/abs/2609.13838)). They were trained on 1450 quality-controlled pairs of the SynthRAD2023 and SynthRAD2025 challenges plus four prostate cohorts, with the planning CT registered to each input by IMPACT-Reg. Three 5-fold families are available on [Hugging Face](https://huggingface.co/VBoussot/ImpactSynth):

| Model | Input | MAE (HU) | SSIM | PSNR (dB) |
| --- | --- | --- | --- | --- |
| MRI-to-CT | MRI | 67.5 | 0.920 | 29.3 |
| CBCT-to-CT | CBCT | 53.6 | 0.939 | 32.1 |
| Unified | MRI or CBCT | 67.7 / 54.2 | 0.920 / 0.938 | 29.2 / 31.9 |

On an external whole-body Dixon MRI dataset (BIC-MAC), fine-tuning on 45 local cases brought the MAE from 100.9 to 62.2 HU. Each app ships its training configuration, so `konfai-apps fine-tune` starts from the released weights on your own pairs.

**IMPACT-Synth**

The extension runs these models on a loaded MRI or CBCT and checks the result:

- **Synthesis** : pick the model, the checkpoints to ensemble and the test-time augmentations, click Run. The sCT is loaded over the input, in HU. About 15 s per checkpoint on a 24 GB GPU.
- **Evaluation with a reference CT** : MAE map, PSNR, SSIM, and the Dice of TotalSegmentator structures on the sCT and on the CT.
- **Evaluation without reference** : uncertainty map from the ensemble and TTA spread, conformity map from the disagreement of the segmentations.
- **Segmentation tab** : TotalSegmentator, MRSegmentator and IMPACT-Seg on the input or on the sCT.

[Video (2 min)](https://github.com/vboussot/SlicerImpactSynth/blob/main/Screenshots/SlicerImpactSynth-tutorial.mp4) and [tutorial](https://github.com/vboussot/SlicerImpactSynth/blob/main/TUTORIAL.md).

**KonfAI**

The generic extension runs any KonfAI App on the volumes of the scene. Since February: apps added from any Hugging Face repository or local folder, with per-checkpoint download; an **Advanced** dialog for patch size, batch size and the parameters an app exposes, saved as a local app if you want; **remote servers** (`konfai-apps-server` on a GPU machine, its GPUs and memory shown in the panel); live RAM and VRAM gauges; fine-tuning setup; a launcher for KonfAI Studio. ImpactSynth and ImpactReg are built on the same library and get all of it.

Three videos of about a minute: [run a published app](https://github.com/vboussot/SlicerKonfAI/blob/main/Screenshots/SlicerKonfAI-inference.mp4), [quality assurance](https://github.com/vboussot/SlicerKonfAI/blob/main/Screenshots/SlicerKonfAI-qa.mp4), [apps and remote servers](https://github.com/vboussot/SlicerKonfAI/blob/main/Screenshots/SlicerKonfAI-apps.mp4); [tutorial](https://github.com/vboussot/SlicerKonfAI/blob/main/TUTORIAL.md).

**Install**

View → Extensions Manager → **ImpactSynth** or **KonfAI** (Slicer 5.10 or newer). PyTorch comes from the SlicerPyTorch extension; a GPU with 8 GB of VRAM is enough.

Feedback and requests for other models are welcome.

Valentin Boussot, Cédric Hémon

Boussot V. et al., _TotalSynth: Robust Whole-Body Synthetic CT from MRI and CBCT_, arXiv:2609.13838, 2026.  
Boussot V., Dillenseger J.-L., _KonfAI: A Modular and Fully Configurable Framework for Deep Learning in Medical Imaging_, arXiv:2508.09823, 2025.