DCA: Graph-Guided Deep Embedding Clustering for Brain Atlases

This repository contains the official implementation of our NeurIPS 2025 paper (5554, Poster) DCA: Graph-Guided Deep Embedding Clustering for Brain Atlases. The method integrates pretraining and spatial graph-based constraints to generate anatomically and functionally meaningful brain atlases.

DCA Usage

0. Dependencies

We recommend using Python == 3.9.21.

All required packages are listed in DCA/req_trim.txt.

1. Prepare Data

Use DCA/data/data_preparation.ipynb to prepare both fMRI and mask data: fMRI inputs should be resting-state scans normalized to the MNI152 space with at least 300 TRs (note that if TR ≠ 0.72 s, it is recommended to resample temporally to 0.72 s beforehand, as this step is not included in the notebook), and mask inputs should be the corresponding FreeSurfer aparc+aseg.nii.gz already registered in MNI152 space.

Place preprocessed 4D fMRI volumes in DCA/data/fmri/, we have placed a demo fMRI.

Place your ROI masks in DCA/data/mask/. This implementation supports customization for gray matter, white matter, and subcortex-specific atlases. We have placed a demo mask.

Ensure data/sub_test.txt contains the list of subject IDs (one per line), we have placed a demo text.

The pretrained model (swin_model_epoch_30.pth) is automatically loaded if present. We conduct pre-training using monai (https://monai.io/) and customize some functions through DCA/swin_unetr.py

2. Run DCA

python main.py

This will generate subject-level brain parcellations using the provided pretrained model. Results will be saved to results/demo/.

Command-line Options

You can customize key inference settings via arguments in main.py. The main options are:

• -k, --n_clusters: Number of parcels to generate (default: 100)
• -e, --epoch: Maximum training epochs (default: 8)
• -v, --vali: Whether to keep the best atlas based homogeneity (default: True)

Validation requires more computing resources. If --vali is set to False, we recommend using --epoch < 10 to avoid overfitting.

3. Results

Output can be found in data/results/demo

AtlaScore Usage

Downstream

For all the operation instructions, please see demo.ipynb.

Similarity

• Provide fMRI data shape=(x,y,z,t) and the atlas file shape=(x,y,z) to be evaluated, modify the paths in eva.py, and run the command python eva.py.
• To evaluate DCBC, you must provide the following files mapped to cortical surface vertices: fMRI data, vertex distance file, and parcellation file. Refer to Zhi et al. for more detailed methodological instructions.

Citations

If you find our work useful for your research, please consider citing our paper:

@article{wang2025dca,
  title={DCA: Graph-Guided Deep Embedding Clustering for Brain Atlases},
  author={Wang, Mo and Peng, Kaining and Tang, Jingsheng and Wen, Hongkai and Liu, Quanying},
  journal={arXiv preprint arXiv:2509.01426},
  year={2025}
}