Cyst-X: A Multi-Center MRI Benchmark and Federated Learning Framework for Malignancy-Risk Stratification of Pancreatic Cystic Neoplasm

Official implementation of Cyst-X, an end-to-end multi-center pipeline integrating state-of-the-art pancreas segmentation, federated optimization, and classical radiomics pipelines for automated malignancy risk stratification of intraductal papillary mucinous neoplasms (IPMNs).

---

📌 Overview

Pancreatic cancer is projected to be the second-deadliest cancer by 2030, making early detection critical. IPMNs are key cancer precursors, but current consensus guidelines struggle to stratify malignancy risk accurately.

Cyst-X addresses this structural bottleneck by providing:

1. The Largest Multi-Center Pancreas MRI Resource: 1,461 abdominal MRI scans from 764 patients across seven international medical centers with expert annotations and pathology-anchored ground truth.
2. Advanced Pancreas Segmentation: Integrating PanSegNet (a linear self-attention transformer-based backbone) for high-fidelity region of interest (ROI) extraction.
3. Privacy-Preserving Federated Learning: Distributed risk classification training utilizing FedAvg and FedProx without exchanging raw clinical images.

📂 Repository Structure

Cyst-X/
├── Classification/          # Pipelines and frameworks for deep risk modeling
├── Segmentation/            # Pipeline and engines for pancreas ROI localization
├── MRQy/                    # MRQy analysis
└── README.md                # Main repository documentation hub

📦 Dataset Access & Download

The Cyst-X dataset includes raw/preprocessed NIfTI volumes, expert pancreas segmentation masks, center-wise stratified splits, and clinical metadata. Please download the dataset via HuggingFace.:

First, please make sure that huggingface_hub is installed:

pip install huggingface_hub

Then, using Python and the huggingface_hub library:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="phy710/Cyst-X",
    repo_type="dataset",
    local_dir="/dataset"
)

You can also access and download the dataset mirror directly from the Open Science Framework (OSF). However, uploading massive amounts of files to OSF is difficult. So we compressed the dataset into several zip files. Please organize the files according to the expected directory layout.

Expected Directory Layout

Once downloaded, ensure your dataset directory matches the following structure:

/dataset/
├── data splits/
│   ├── classification/
│   │   ├── 2-class/
│   │   │   ├── t1.csv                          # 4-fold CV splits for T1W (High-Risk vs. No/Low-Risk; stratified center-wise)
│   │   │   ├── t2.csv                          # 4-fold CV splits for T2W (High-Risk vs. No/Low-Risk; stratified center-wise)
│   │   │   └── two_inputs.csv                  # 4-fold CV splits for paired (T1W+T2W) multi-modal fusion; stratified center-wise
│   │   └── 3-class/
│   │       ├── t1.csv                          # 5-fold CV splits for T1W (No-Risk vs. Low-Risk vs. High-Risk; pooled across all centers)
│   │       └── t2.csv                          # 5-fold CV splits for T2W (No-Risk vs. Low-Risk vs. High-Risk; pooled across all centers)
│   └── pansegnet segmentation/
│       ├── Task210_PancreasT1MRI_splits.csv    # 5-fold cross-validation splits for T1W pancreas segmentation
│       └── Task211_PancreasT2MRI_splits.csv    # 5-fold cross-validation splits for T2W pancreas segmentation
├── IPMN_Classification/                        # IPMN classification dataset
│   ├── t1/                                     # Cropped 3D pancreatic ROIs for single-input T1W (.nii.gz)
│   ├── t2/                                     # Cropped 3D pancreatic ROIs for single-input T2W (.nii.gz)
│   ├── IPMN_labels_t1_total.xlsx               # Comprehensive labels and metadata for all T1W scans
│   ├── IPMN_labels_t2_total.xlsx               # Comprehensive labels and metadata for all T2W scans
│   └── IPMN_labels_total.xlsx                  # Master cohort metadata and paired (T1W+T2W) ground truth labels
└── IPMN_images_masks/                          # Pancreas segmentation dataset
    ├── t1/
    │   ├── images/                             # Full-volume T1W NIfTI scans (.nii.gz)
    │   └── masks/                              # Full-volume ground-truth pancreas masks (.nii.gz)
    └── t2/
        ├── images/                             # Full-volume T2W NIfTI scans (.nii.gz)
        └── masks/                              # Full-volume ground-truth pancreas masks (.nii.gz)

📊 Mapping Between Reported Table/Figure and the Corresponding Script

Table 1 and Supplementary Tables A2--A3: /Segmentation/PanSegNet/Task210_PancreasT1MRI/get_results.sh and /Segmentation/PanSegNet/Task211_PancreasT2MRI/get_results.sh.

Table 2:

(1) Three-class classification: Classification/Deep Learning/internal/3-class/centralized/fold_test.py;

(2) Two-class classification: Classification/Deep Learning/internal/2-class/xxx/fold_test.py and Classification/Radiomics, where xxx is centralized, fedavg, or fedprox. Alternatively, you can find the results in Supplementary Tables A9 and A10 at row Global.

Figure 3: Classification/tsne/tsne.py and Classification/Deep Learning/internal/2-class/centralized/tsne.py.

Table 3: Classification/vs_radiologists/run.sh for calibrated results comparison, Classification/vs_radiologists/run_uncalibrated.sh for uncalibrated results comparison, Classification/vs_radiologists/run_histology.sh for calibrated histology results comparison, Classification/vs_radiologists/run_histology_uncalibrated.sh for uncalibrated histology results comparison.

Figure 4: Classification/results_calibration/analysis_internal.sh.

Figure 5: MRQy/MRQy.py.

Supplementary Tables A9, A10: Classification/results_calibration/analysis_internal.sh.

Supplementary Tables A11, A12: Classification/results_calibration/analysis_external.sh.

Supplementary Table A13: Classification/results_calibration/analysis_fusion.sh.

To download the pre-trained models:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="phy710/Cyst-X",
    repo_type="model",
    local_dir="./"
)

Then put the models into the folders of the corresponding tasks.

📝 Citation

If you use this dataset in your research, please cite our paper:

@article{pan2025cyst,
  title={Cyst-X: A Multi-Center MRI Benchmark and Federated Learning Framework for Malignancy-Risk Stratification of Pancreatic Cystic Neoplasm},
  author={Pan, Hongyi and Durak, Gorkem and Keles, Elif and Hong, Ziliang and Seyithanoglu, Deniz and Zhang, Zheyuan and Medetalibeyoglu, Alpay and Aktas, Halil Ertugrul and Bejar, Andrea Mia and Taktak, Yavuz and others},
  journal={arXiv preprint arXiv:2507.22017},
  year={2025}
}