nnU-Net Model for Bilateral Breast Segmentation on Chest CT

This repository provides documentation and inference instructions for a trained nnU-Net v2 model for bilateral breast segmentation from chest CT scans.

The pretrained model file is:

nnunet_CT_breastseg_best.zip

Model Overview

This model was trained using the nnU-Net framework to automatically segment bilateral breast regions on chest CT images.

Item	Description
Model name	`nnunet_CT_breastseg_best.zip`
Framework	nnU-Net v2
Task	Bilateral breast segmentation
Input modality	Chest CT
Input format	NIfTI (`.nii.gz`)
Output	Breast segmentation mask
Intended use	Research use only

Intended Use

This model is intended for automated segmentation of bilateral breast tissue on chest CT scans. The output masks may be used for downstream research tasks such as image preprocessing, radiomics analysis, breast tissue quantification, or deep learning-based risk assessment.

This model is intended for research use only and has not been validated for clinical decision-making.

Installation

Create a Python environment and install nnU-Net v2:

conda create -n nnunet_breastseg python=3.10 -y
conda activate nnunet_breastseg

pip install torch torchvision torchaudio
pip install nnunetv2

Please install the PyTorch version that is compatible with your CUDA version.

Install the Pretrained Model

After downloading nnunet_CT_breastseg_best.zip, install the model with:

nnUNetv2_install_pretrained_model_from_zip nnunet_CT_breastseg_best.zip

After installation, the model will be available in the nnU-Net results directory.

Input Data Preparation

Input CT images should be converted to NIfTI format and placed in an imagesTs folder.

For single-modality CT input, nnU-Net requires the following naming format:

imagesTs/
├── case001_0000.nii.gz
├── case002_0000.nii.gz
├── case003_0000.nii.gz
└── ...

Each case should have one CT volume, with the channel suffix _0000.

Example:

case001_0000.nii.gz

Recommended CT Preprocessing

The input images should be chest CT volumes in NIfTI format. Before inference, please ensure that:

The image orientation is correct.
The CT image is not corrupted or truncated.
The image spacing and metadata are properly preserved during DICOM-to-NIfTI conversion.
The input file follows nnU-Net naming conventions.
The image intensity is preserved in Hounsfield units whenever possible.

A typical DICOM-to-NIfTI conversion can be performed using dcm2niix:

dcm2niix -z y -o ./nifti_output ./dicom_folder

Then rename the converted file according to the nnU-Net format:

case001_0000.nii.gz

Inference

Run prediction using nnU-Net v2:

nnUNetv2_predict \
  -i imagesTs \
  -o predictions \
  -d DatasetXXX_BreastSeg \
  -c 3d_fullres \
  -f all \
  -chk checkpoint_best.pth

Please replace DatasetXXX_BreastSeg with the actual dataset identifier used during training.

If the model was exported and installed correctly, the corresponding dataset name and configuration should be recognized by nnU-Net automatically.

Output

The predicted segmentation masks will be saved in the output folder:

predictions/
├── case001.nii.gz
├── case002.nii.gz
├── case003.nii.gz
└── ...

Each output file is a segmentation mask corresponding to the input CT image.

Label Definition

The segmentation labels should be interpreted according to the label definition used during model training.

Example label structure:

Label value	Anatomical structure
0	Background
1	Breast tissue

If left and right breasts were labeled separately during training, the label definition may be:

Label value	Anatomical structure
0	Background
1	Left breast
2	Right breast

Please confirm the final label definition according to the training dataset.

Example Workflow

# 1. Activate environment
conda activate nnunet_breastseg

# 2. Install pretrained model
nnUNetv2_install_pretrained_model_from_zip nnunet_CT_breastseg_best.zip

# 3. Prepare input files
mkdir -p imagesTs
cp case001_0000.nii.gz imagesTs/

# 4. Run inference
nnUNetv2_predict \
  -i imagesTs \
  -o predictions \
  -d DatasetXXX_BreastSeg \
  -c 3d_fullres \
  -f all \
  -chk checkpoint_best.pth

Visualization

The predicted masks can be visualized using medical image viewers such as:

ITK-SNAP
3D Slicer
MITK Workbench
FSLeyes

For example, in ITK-SNAP, open the CT image as the main image and load the predicted mask as the segmentation overlay.

Model Sharing

The model file can be shared through platforms such as:

GitHub Releases
Hugging Face Model Hub
Zenodo
Institutional data repositories

For long-term academic sharing and citation, Zenodo is recommended because it can provide a DOI.

Citation

If you use this model in your research, please cite the corresponding publication or repository:

@misc{nnunet_ct_breastseg,
  title = {nnU-Net Model for Bilateral Breast Segmentation on Chest CT},
  author = {Your Name},
  year = {2026},
  note = {Pretrained nnU-Net model for bilateral breast segmentation from chest CT}
}

Disclaimer

This model is provided for research purposes only. It is not intended for clinical diagnosis, treatment planning, or direct patient care. The model output should be reviewed and validated by qualified experts before being used in any clinical or research analysis.

License

Please specify the license before releasing the model.

Recommended options include:

MIT License for code
CC BY-NC 4.0 for research-only model weights
Custom institutional research-use license

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