Causal3D-Net: A Causal Learning-Driven Pancreatic Cancer Diagnosis Model for Multi-Center Contrast-Enhanced CT Data

English | 中文

This is the official repository of Causal3D-Net, a dedicated model for pancreatic cancer diagnosis.

The model is based on causal learning theory, enabling it to identify causal features and confounding features, and ultimately use reliable causal features as the basis for diagnosis, thereby achieving stable diagnostic learning.

The figure below illustrates the model architecture proposed in the paper:

⚙️ Install

conda env create -f environment.yaml

🚀 Usage

All experiments can be launched with the following command:

python -m src.main <mode> [options]

Segmentation Training

python -m src.main seg \
    --train /path/to/dataset_for_train.xlsx \
    --test /path/to/dataset_for_test.xlsx \
    --cuda 0 \
    --outdir ./results/segmentation

Arguments:

• –train : Path to the training dataset (Excel file).
• –test : Path to the testing dataset (Excel file).
• –cuda : GPU index to use (default: 0).
• –outdir : Directory to save results (default: ./results).

Causal3DNet Training

python -m src.main causal \
    --train /path/to/dataset_for_train.xlsx \
    --test /path/to/dataset_for_test.xlsx \
    --indi 1 \
    --cent 1 \
    --orth 1 \
    --cuda 0 \
    --outdir ./results/causal3dnet \
    --weight ./pretrained/best_model.pth

Arguments:

• –train : Path to the training dataset (Excel file).
• –test : Path to the testing dataset (Excel file).
• –indi : Whether to use the individual branch (default: enabled).
• –cent : Whether to use the center branch (default: enabled).
• –orth : Whether to use orthogonal constraint (default: enabled).
• –cuda : GPU index to use (default: 0).
• –outdir : Directory to save results (default: ./results).
• –weight : Path to the pretrained segmentation weight file (default: ./best_model.pth).

📝 Notes

• Use -m to run modules so that Python correctly resolves package imports.
• Make sure your dataset is prepared in Excel format (.xlsx), containing proper training and testing splits.
• GPU index (–cuda) should be set according to your hardware.

Example of Dataset Excel File

image_path	mask_path	cancer	center	cluster
Center01Img00002_private.npy	Center01Mask00002_private.npy	1	0	0
Center01Img00003_private.npy	Center01Mask00003_private.npy	1	0	0
…	…	…	…	…

📖 Citation

If you find this repository helpful in your research, please consider citing our work:

@unpublished{huang2026causal3dnet,
  title  = {Causal3D-Net: A Causal Learning–Driven Pancreatic Cancer Diagnosis Model for Multi-Center Contrast-Enhanced CT Data},
  author = {Huang, Denan and others},
  note   = {Manuscript under review}
}

🤝 Collaborating Institutions