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

This repo stores the model weights. Please visit our GitHub repo for more information:

https://github.com/NUBagciLab/Cyst-X/

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.

The two-class classification models in v1 used four-fold cross-validation; in v2, these models are standardized to five-fold cross-validation.

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}
}