README.md

---
license: apache-2.0
library_name: casimir
pipeline_tag: image-segmentation
tags:
- medical-imaging
- mri
- whole-body
- segmentation
- nnunet
- pediatrics
---

# CASIMIR: An Open Pediatric Model for Whole-Body T1-Weighted MRI Anatomy Segmentation

[![Code on GitHub](https://img.shields.io/badge/Code-GitHub-blue?style=for-the-badge)](https://github.com/fjhorvath/casimir)    [![License](https://img.shields.io/badge/License-Apache--2.0-blue?style=for-the-badge)](https://huggingface.co/fjhorvath/casimir/blob/main/LICENSE)

CASIMIR (Child Anatomy Segmentation In MRI with Increased Robustness) is a 2D
[nnU-Net v2](https://github.com/MIC-DKFZ/nnUNet) model that segments six visceral organs and
seven skeletal structures in pediatric and young-adult whole-body T1-weighted MRI, paired
structures labeled by side. This repository holds the weights. The inference code is on
[GitHub](https://github.com/fjhorvath/casimir). The model and its evaluation are described
in a manuscript that is not yet published, see Citation.

<p align="center">
  <img class="dark:hidden" src="assets/casimir-overview.png" width="560" alt="Four axial T1 slices with the CASIMIR overlay, mid thorax to lower leg, beside a whole-body rendering of CASIMIR output with each cutting plane marked, the lower legs from a second examination">
  <img class="hidden dark:block mx-auto" src="assets/casimir-overview-dark.png" width="560" alt="Four axial T1 slices with the CASIMIR overlay, mid thorax to lower leg, beside a whole-body rendering of CASIMIR output with each cutting plane marked, the lower legs from a second examination">
</p>

## Inference

```bash
pip install git+https://github.com/fjhorvath/casimir.git@v1.0.0
casimir -i <nifti file or directory> -o <output directory>
```

Add `--device cpu` if there is no NVIDIA GPU, or `--device mps` on Apple silicon.
`casimir --help` lists all flags. The weights are downloaded from here on the first run.
For offline use, set `CASIMIR_WEIGHTS_PATH` to a local copy.

The input is NIfTI, `.nii` or `.nii.gz`. The tested volumes were converted from DICOM with
dicom2nifti and reorientation disabled (`reorient_nifti=False`), which keeps the scanner's
axis order. nnU-Net reads the voxel array as stored, so a volume saved in another axis
order is untested.

The output is one label map per input, on the input grid and under the input filename,
with the voxel values listed below. One whole-body examination takes about 1.5 minutes on
an RTX 4090 or on an Apple M2 Pro with `--device mps`, and about 4 minutes on the M2 Pro
with `--device cpu`.

## Files

`dataset.json` and `plans.json` are the nnU-Net v2.5 dataset descriptor and plans.
`fold_all/weights.pth` is the checkpoint, a single fold trained on all training cases, with
no ensemble. `LICENSE` is the Apache License 2.0.

## Structures

The 13 structures give 20 labels, one per side for the paired ones:

| | | | |
|---|---|---|---|
| 1 fibula left | 2 fibula right | 3 femur left | 4 femur right |
| 5 tibia left | 6 tibia right | 7 spine | 8 liver |
| 9 lung left | 10 lung right | 11 sacrum | 12 spleen |
| 13 humerus left | 14 humerus right | 15 heart | 16 urinary bladder |
| 17 pelvis left | 18 pelvis right | 19 kidney left | 20 kidney right |

Background is `0`. The output is the raw argmax with no post-processing.

## Intended use

CASIMIR is a research tool, not a medical device. Do not use its output for clinical
decisions.

It was trained and evaluated on contrast-enhanced axial fat-suppressed T1-weighted
gradient-echo whole-body MRI from a single 3-T PET/MR system, in patients aged 1 to 21
years with lymphoma or post-transplant lymphoproliferative disorder. External validation is
required to establish generalizability.

## License

These weights and the inference code are released under the Apache License 2.0, see
`LICENSE`.

## Citation

Until the manuscript is published, cite the software:

```bibtex
@software{horvath2026casimir,
  author  = {Horvath, Felix J. and Barrow, Michael J. and Singh, Shashi B. and Lokesha, Yashas U. and Okpokpo, Darlene and Sivanandam, Sharanya and Gatidis, Sergios and Daldrup-Link, Heike E.},
  title   = {{CASIMIR}: An Open Pediatric Model for Whole-Body T1-Weighted {MRI} Anatomy Segmentation},
  version = {1.0.0},
  year    = {2026},
  url     = {https://github.com/fjhorvath/casimir}
}
```

CASIMIR is built on nnU-Net, please also cite
[Isensee et al. (2021)](https://doi.org/10.1038/s41592-020-01008-z).

## Funding

This work was supported by a grant from the National Cancer Institute, grant number
R01 CA269231. Felix J. Horvath was supported by a travel grant from the Rolf W. Günther
Foundation for Radiological Sciences.