Cyst-X: A Multicenter MRI Benchmark for Pancreatic Cyst Risk Stratification and Cross-Institutional AI

Pancreatic cystic neoplasms, including Intraductal Papillary Mucinous Neoplasms (IPMNs), are increasingly detected due to advancements in imaging technologies and growing awareness. Early diagnosis and accurate characterization of IPMNs are essential for guiding appropriate clinical management and reducing unnecessary interventions.

To support research in this domain, we present Cyst-X, the largest multi-center pancreas MRI resource to date. It comprises 723 T1-weighted (T1W) and 738 T2-weighted (T2W) abdominal MRI scans from 764 distinct patients collected across seven international clinical centers:

• New York University Langone Health (NYU)
• Mayo Clinic Florida (MCF)
• Northwestern University (NU)
• Allegheny Health Network (AHN)
• Mayo Clinic Arizona (MCA)
• Istanbul University Faculty of Medicine (IU)
• Erasmus Medical Center (EMC)

📁 Dataset Directory Structure

Cyst-X/
├── classification/
│   ├── t1/
│   │   ├── images/                  # 723 preprocessed pancreatic ROIs
│   │   └── visualizations/
│   ├── t2/
│   │   ├── images/                  # 738 preprocessed pancreatic ROIs
│   │   └── visualizations/
│   ├── IPMN_labels_t1_total.xlsx
│   ├── IPMN_labels_t2_total.xlsx
│   ├── IPMN_labels_total.xlsx
│   ├── t1_split.csv                  # Five-fold split for T1W and all paired fusion methods
│   └── t2_split.csv
└── segmentation/
    ├── t1/
    │   ├── images/                  # 723 full-volume MRI scans
    │   ├── masks/                   # Corresponding pancreas masks
    │   └── visualizations/
    └── t2/
        ├── images/                  # 738 full-volume MRI scans
        ├── masks/                   # Corresponding pancreas masks
        └── visualizations/

The classification/ and segmentation/ directories are stored directly at the repository root. The classification ROIs were generated from N4-corrected images using the reference masks.

Use classification/t1_split.csv and classification/t2_split.csv for the current five-fold classification experiments. All paired T1W/T2W fusion experiments use classification/t1_split.csv, applying the same T1 patient fold assignment to both modalities.

Intermediate Cyst-X-RAS and Cyst-X-n4 directories are not included in this release. Preprocessing code is maintained in the GitHub code repository.

The current release (main, tagged v2) corrects image–mask orientation mismatches present in some v1 cases. The original release remains available under the v1 tag.

Image orientation

The classification dataset is released in RAS voxel orientation (Right, Anterior, Superior). The full-volume segmentation dataset retains its original acquisition orientation, with each mask aligned to its corresponding image; it has not been standardized to RAS. To convert the segmentation images and masks to RAS, use step1_reorientation_to_RAS.py.

Here, RAS follows the increasing-voxel-axis convention used by NiBabel and SimpleITK's orientation helpers. The supplied script uses MONAI Orientationd(axcodes="RAS") and writes NIfTI files with NiBabel. For a classification image, nib.aff2axcodes(nib.load(path).affine) returns ('R', 'A', 'S'); the corresponding SimpleITK check is sitk.DICOMOrientImageFilter.GetOrientationFromDirectionCosines(sitk.ReadImage(path).GetDirection()). ITK-SNAP may label the same RAS-oriented volume as LPI, because its orientation-code convention names the opposite ends of the axes. This is a difference in labeling, not an image/mask misalignment; see the ITK orientation conventions.

Citation

If you use Cyst-X in your research, please cite our paper:

@article{pan2025cyst,
  title={Cyst-X: A Multicenter MRI Benchmark for Pancreatic Cyst Risk Stratification and Cross-Institutional AI},
  author={Pan, Hongyi and Durak, Gorkem and Aktas, Halil Ertugrul and Hong, Ziliang and Wang, Zhixiang and Keles, Elif and Seyithanoglu, Deniz and Zhang, Zheyuan and Medetalibeyoglu, Alpay and Bejar, Andrea Mia and others},
  journal={arXiv preprint arXiv:2507.22017},
  year={2025}
}