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

CASIMIR (Child Anatomy Segmentation In MRI with Increased Robustness) is a 2D nnU-Net v2 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. The model and its evaluation are described in a manuscript that is not yet published, see Citation.

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

Inference

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:

@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).

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.

Downloads last month: -; Downloads are not tracked for this model. How to track