Not for clinical or commercial use. Research use only.
This repository includes the inference code and models used in our paper on modality transfer from CT to MRI for pediatric image segmentation. If you use it, please cite our corresponding paper:
@article{eicke2026paediatric,
title={Paediatric thoracal and abdominal organ segmentation in CT and MRI using global intensity non-linear augmentation},
author={Marcel Eicke and Kai Geissler and Jasmin Heepe and Markus Benedikt Krueger and Hinrich B. Winther and Ann-Katrin Heymann and Fabian Knörr and Andreas M. Bucher and Jochen Herrmann and Clemens Benoit and Hans-Joachim Mentzel and Andrea Schenk and Wilhelm Wößmann and Diane M. Renz and Bianca Lassen-Schmidt},
journal={European Journal of Radiology Artificial Intelligence},
pages={100096},
year={2026},
publisher={Elsevier},
doi = {https://doi.org/10.1016/j.ejrai.2026.100096}
}Install uv first if it is not already available.
uv python install 3.13
uv sync
This creates a local virtual environment in .venv and installs the project in editable mode with its dependencies.
You need the deep learning model in ONNX format and the input image in NIfTI format. As we use the MONAI LoadImage class for image loading, other formats should also be supported but were not tested. The image has to be resampled to a voxel size of 1.5 x 1.5 x 1.5 mm^3 and needs to be in LPS (left-posterior-superior) orientation. The script will take care of intensity normalization. The output is an integer label encoded mask in NIfTI format.
uv run python -m pedseggin.apply_model --image_path <path_to_image> --model_path <path_to_model> --output_path <output_file_name>
After installation, you can also use the console command:
uv run pedseggin --image_path <path_to_image> --model_path <path_to_model> --output_path <output_file_name>
Trained models to be used with this script can be found in the folder models. They are tracked using git lfs and can be added as normal files. .gitattributes is configured to automatically track all .onnx models with git lfs.
We offer the following models:
- pediatric_ct_1p5mm_96x96x96.onnx Trained on Pediatric-CT-SEG (TCIA dataset). Images were resampled to a voxel size of 1.5^3 mm^3 with a patch size of 96^3 during training.
- pediatric_ct_1p5mm_96x96x96_gin.onnx Trained on Pediatric-CT-SEG (TCIA dataset) using GIN augmentation. Images were resampled to a voxel size of 1.5^3 mm^3 with a patch size of 96^3 during training.
- totalsegmentator_v2_1p5mm_96x96x96_gin.onnx Trained on TotalSegmentatorV2 CT data using GIN augmentation. Images were resampled to a voxel size of 1.5^3 mm^3 with a patch size of 96^3 during training.
Install the development dependencies and run the end-to-end test suite with coverage:
uv sync --group dev
uv run pytest
The pytest configuration reports terminal coverage for pedseggin and also writes coverage.xml for CI integration.
This project is under a non-commercial license, for details refer to the license file.
The software is not qualified for use as a medical product or as part thereof. It is provided 'as is' without specific verification or validation and without guarantees to implement requested features.
Kai Geissler, Fraunhofer Institute for Digital Medicine MEVIS, Bremen, Germany.