Ventricle Segmentation Model

Automatic segmentation of the ventricular system and brain stem in brain MRI (T1 and T2), based on nnU-Net v2. Accepts a DICOM series or NIfTI volumes and returns either a DICOM RTSTRUCT or a NIfTI label map.

Research use only. Not a medical device and not intended for clinical decision making.

---

Citation

If you use this code or the trained models, please cite:

tbd

---

License

• Code: Apache License 2.0, modified by the Commons Clause and attribution terms in the License file (see LICENSE). \
• Model weights: Creative Commons Attribution Non Commercial Share Alike 4.0 (CC BY-NC-SA 4.0) (see LICENSE).

---

Requirements

• Python 3.10+
• NVIDIA GPU with CUDA (the script asserts torch.cuda.is_available(); CPU is not supported)
• Trained model weights (see below)

Model weights

Weights are not included in this repository. Download them from <WEIGHTS URL / Zenodo DOI> and unpack so the tree looks like this:

nnUNet_results/
├── Dataset501_VenSeg3DT1/
│   └── nnUNetTrainer__nnUNetPlans__2d/
│       ├── fold_0/checkpoint_best.pth
│       ├── ...
│       ├── dataset.json
│       └── plans.json
└── Dataset502_VenSeg3DT2/
    └── nnUNetTrainer__nnUNetPlans__2d/
        └── ...

nnUNet_results must sit next to run_inference.py; the script sets the environment variable itself.

---

Quick start

python run_inference.py \
  --input_folder  /path/to/patient_001 \
  --output_folder /path/to/output \
  --mr_sequence   T1

Argument	Description
--input_folder	Folder with either the DICOM slices of one series, or NIfTI volumes
--output_folder	Folder where predictions are written (must exist)
--mr_sequence	T1 (Dataset501) or T2 (Dataset502)

Input and output behaviour

The format is decided from the first file in the input folder, so do not mix DICOM and NIfTI in one folder.

• DICOM in (*.dcm): the series is converted to NIfTI in a temporary folder, segmented, and written back as an RTSTRUCT at output_folder/prediction_<folder_name>.dcm with the ROIs Ventricle and Brain stem.
• NIfTI in (*.nii.gz): all volumes in the folder are segmented and label maps are written to output_folder. Files must follow the nnU-Net naming convention with a modality suffix, i.e. case_0000.nii.gz.

Label values in the NIfTI output: 1 = ventricle, 2 = brain stem, 0 = background.

---

Repository layout

.
├── run_inference.py     # CLI entry point: preprocessing, inference, RTSTRUCT export
├── requirements.txt
├── nnUNet_results/      # trained models (downloaded separately, git-ignored)
├── LICENSE
└── README.md

Training

Models were trained with standard nnU-Net v2 (nnU-Net) in the 2d configuration (nnUNetTrainer__nnUNetPlans__2d), using checkpoint_best.pth for inference.

Known limitations

• GPU is mandatory; there is no CPU fallback.
• The DICOM branch handles one series per input folder.
• Check orientation; verify overlays in your viewer before downstream use.

Contact

tbd

Acknowledgements

Built on nnU-Net and rt-utils.