Cyst-X: AI for Pancreatic Cyst Risk

• Cyst-X is an AI framework that applies pancreas-centered MRI analysis combining segmentation, radiomics, deep learning, and federated learning to predict IPMN malignancy risk.
• The framework achieves a binary AUC of approximately 0.82, significantly surpassing the Kyoto guidelines and expert radiologist performance on high-risk classification.
• It offers a publicly released, multicenter MRI dataset and benchmark for privacy-preserving, cross-site pancreatic imaging research, emphasizing robust domain generalization.

Cyst-X is an AI framework and a publicly released, first large-scale, multi-center pancreatic cysts MRI dataset for malignancy risk prediction in intraductal papillary mucinous neoplasms (IPMNs). It is designed around pancreas-centered MRI analysis rather than individual cyst segmentation, combining pancreas segmentation, radiomics, deep learning classification, and federated learning to estimate whether an IPMN is no risk, low risk, or high risk. In its principal binary setting, the framework attains an AUC of 0.82, significantly outperforming the Kyoto guidelines and expert radiologists, and it is explicitly positioned as a privacy-preserving benchmark for multicenter pancreatic MRI research (Pan et al., 29 Jul 2025).

1. Clinical problem and conceptual scope

Cyst-X addresses the clinical problem of risk stratification in pancreatic cystic lesions, especially IPMNs, which are among the few identifiable precursors to pancreatic ductal adenocarcinoma. The underlying motivation is that pancreatic cancer is projected to become the second leading cause of cancer death in Western countries by 2030, while current guideline-based management of pancreatic cysts remains limited by overtreatment, undertreatment, and inter-observer variability. The framework adopts MRI rather than CT because MRI offers superior soft-tissue contrast and sensitivity for duct communication, mural nodules, multifocal cysts, and isoattenuating pancreatic ductal adenocarcinoma, while avoiding ionizing radiation during surveillance (Pan et al., 29 Jul 2025).

Within the Cyst-X label space, cases are assigned to three categories: no risk or control, low-risk IPMN, and high-risk IPMN. High-risk cases are defined by histopathologically confirmed high-grade dysplasia or worse, including carcinoma in situ and invasive cancer, with MRI acquired within 6 months of the confirmatory procedure. Low-risk cases are either histologically confirmed low- or intermediate-grade dysplasia or presumed low-risk on imaging with at least 3 years of follow-up showing less than 2.5 mm growth and no worrisome features or high-risk stigmata. No-risk controls include a normal pancreas or benign cysts not associated with IPMN and lacking malignant potential. This definition makes Cyst-X primarily a malignancy prediction and surveillance problem rather than a generic lesion detection task (Pan et al., 29 Jul 2025).

A useful antecedent is an earlier CT-based framework that classified four pathologically confirmed pancreatic cyst subtypes without lesion pre-segmentation and achieved an overall accuracy of 72.8% on 206 patients, using DenseNet on the whole abnormal pancreas and saliency maps for explanation (Li et al., 2018). This suggests a methodological continuity between organ-level pancreatic cyst analysis on CT and the later MRI-based, multicenter, privacy-aware formulation embodied by Cyst-X.

2. Dataset composition, labels, and curation

Cyst-X comprises 764 unique individuals aged 18 years or older and 1,461 MRI volumes, specifically 723 T1-weighted scans and 738 T2-weighted scans collected across seven institutions: New York University Langone Health, Mayo Clinic Florida, Northwestern University, Allegheny Health Network, Mayo Clinic Arizona, Istanbul University Faculty of Medicine, and Erasmus Medical Center. The scanners span GE, Siemens, and Philips systems at both 1.5T and 3T. This multicenter composition is central to the framework’s emphasis on domain shift, generalization, and federated learning (Pan et al., 29 Jul 2025).

Each scan is associated with an expert-drawn pancreas segmentation mask and a risk label. DICOM images were de-identified at the source institutions and converted to NIfTI. Pancreas segmentation was performed by expert radiologists using ITK-SNAP, and all masks were reviewed by a senior abdominal radiologist to standardize quality. For classification, the dataset distribution is reported as 164 no-risk, 339 low-risk, and 149 high-risk T1-weighted scans, and 159 no-risk, 344 low-risk, and 152 high-risk T2-weighted scans. The dataset is described as IPMN-dominant, with 90–95% of cases belonging to that category (Pan et al., 29 Jul 2025).

A central feature of the curation protocol is that individual cysts are not separately delineated. The region of interest is the entire pancreas, and the task is to learn risk from pancreas-centered MRI rather than from explicit lesion masks. This suggests that Cyst-X should not be understood as a cyst segmentation benchmark in the narrow sense. Instead, it is a pancreas-level representation learning resource in which cysts, nodules, ductal changes, and surrounding parenchymal patterns are implicitly encoded by the model’s pancreas ROI and the associated malignancy label (Pan et al., 29 Jul 2025).

3. Imaging heterogeneity and preprocessing design

The MRI data span 2004–2024 and exhibit substantial heterogeneity in slice thickness, in-plane resolution, acquisition parameters, image quality, and center-specific protocol choices. To characterize this heterogeneity, the authors use MRQy to compute 21 image-quality indicators, including intensity statistics, contrast metrics, entropy, and signal-to-noise-related quantities, and then reduce these indicators with UMAP. Three normalization schemes are explicitly reported: $$x'_{\text{minmax}} = \frac{x - x_{\min}}{x_{\max} - x_{\min}},$$

$x'_{\text{whiten}} = \frac{x}{\sigma},$

$x'_{\text{zscore}} = \frac{x - \mu}{\sigma}.$

The resulting UMAP plots show distinct clustering by center and a strong influence of slice thickness on image-quality clusters, directly documenting the domain shifts that the learning system must absorb (Pan et al., 29 Jul 2025).

The preprocessing pipeline for classification is pancreas-centered. Volumes undergo resizing, bias field correction, and intensity normalization, after which the pancreas segmentation is used to crop a 3D ROI resized to $96 \times 96 \times 96$ voxels for the DenseNet classifiers. In the segmentation pipeline, the entire pancreas is segmented first and only then forwarded to downstream risk prediction. Because individual cysts are not segmented, the model must infer malignancy from distributed pancreatic structure, including cyst morphology, mural nodules, duct caliber, and diffuse gland-level patterns, rather than from an isolated lesion crop (Pan et al., 29 Jul 2025).

This design has a methodological implication. The dominant nuisance factor is not merely lesion appearance but cross-site acquisition drift. By explicitly quantifying center clustering and slice-thickness effects, Cyst-X turns heterogeneity into a benchmark dimension rather than treating it as incidental noise. A plausible implication is that methods performing well on Cyst-X are being tested not only on pancreatic cyst risk prediction, but also on MRI domain generalization under realistic multi-institutional variation.

4. Computational architecture

Cyst-X is organized into pancreas segmentation and pancreas-centered risk classification. For segmentation, the primary centralized model is PanSegNet, which is built atop nnU-Net and augments the bottleneck with linear self-attention. Given input features $\mathbf{X} \in \mathbb{R}^{N \times d}$, queries, keys, and values are defined as

$$\mathbf{Q} = \mathbf{X}\mathbf{W}_Q + \mathbf{b}_Q,\quad \mathbf{K} = \mathbf{X}\mathbf{W}_K + \mathbf{b}_K,\quad \mathbf{V} = \mathbf{X}\mathbf{W}_V + \mathbf{b}_V.$$

Standard self-attention is written as

$$\text{sim}(\mathbf{q}, \mathbf{k}) = \mathrm{SoftMax}\left(\frac{\mathbf{q}\mathbf{k}^T}{\sqrt{d}}\right),$$

whereas the linearized form uses

$$\text{sim}'(\mathbf{q}, \mathbf{k}) = \phi(\mathbf{q})\,\rho(\mathbf{k})^T,$$

reducing complexity from $O(dN^2)$ to $O(d^2N)$. This choice is aimed at capturing global pancreas context efficiently in MRI segmentation (Pan et al., 29 Jul 2025).

For federated segmentation experiments, the framework uses Swin-UNETR instead of PanSegNet because nnU-Net self-configures from global dataset statistics, which is incompatible with the federated setting where no site observes the full data distribution. For risk classification, Cyst-X includes two branches. The first is radiomics-based, with 763 base 3D features expanded to 3,052 by multi-scale statistics and 447 base 2D features expanded to 1,788, including intensity descriptors, Sobel-derived gradients, Laws’ texture energy, GLCM features, and CoLlAGe features, followed by mRMR feature selection and a random forest classifier run over 50 iterations. The second, and central, branch is a 3D DenseNet-121 operating on pancreas ROIs of size $x'_{\text{whiten}} = \frac{x}{\sigma},$0, trained for both 3-class and 2-class tasks (Pan et al., 29 Jul 2025).

The binary and three-class DenseNet-121 models are trained with AdamW, initial learning rate 0.001 decayed by a factor of 0.1 every 30 epochs, batch size 16, and 100 epochs. The loss is standard cross-entropy,

$x'_{\text{whiten}} = \frac{x}{\sigma},$1

DenseNet-121 is compared with 3D ResNet-34, ShuffleNet-V2, MobileNet-V2, and EfficientNet-B0, and is reported to offer a favorable balance of performance and computational load, with 18.31G MACs and 11.25M parameters. Multisequence fusion is implemented in two ways: feature fusion concatenates T1-weighted and T2-weighted encodings before a shared classifier, whereas probability fusion combines separate classification heads by weighted averaging of predicted probabilities (Pan et al., 29 Jul 2025).

5. Centralized and federated learning formulation

A defining property of Cyst-X is that it is built not only for centralized model training but also for federated learning across seven institutions. The federated objective is

$x'_{\text{whiten}} = \frac{x}{\sigma},$2

where $x'_{\text{whiten}} = \frac{x}{\sigma},$3 is the local objective at site $x'_{\text{whiten}} = \frac{x}{\sigma},$4. In FedAvg, the global model is updated by weighted averaging,

$x'_{\text{whiten}} = \frac{x}{\sigma},$5

In FedProx, each local objective is augmented by a proximal term,

$x'_{\text{whiten}} = \frac{x}{\sigma},$6

to mitigate client drift under heterogeneity (Pan et al., 29 Jul 2025).

The segmentation results underscore that federated MRI pancreas segmentation remains difficult. Centralized Swin-UNETR reaches Dice $x'_{\text{whiten}} = \frac{x}{\sigma},$7 on T1-weighted scans and $x'_{\text{whiten}} = \frac{x}{\sigma},$8 on T2-weighted scans, whereas federated Swin-UNETR with FedAvg drops to $x'_{\text{whiten}} = \frac{x}{\sigma},$9 and $x'_{\text{zscore}} = \frac{x - \mu}{\sigma}.$0, respectively. PanSegNet, evaluated only in the centralized regime, is stronger still, with Dice 86.81% on T1-weighted scans and 89.62% on T2-weighted scans. This establishes that anatomical localization remains a significant vulnerability in cross-site learning (Pan et al., 29 Jul 2025).

The classification task is more robust to federation. In the global binary high-risk versus non-high-risk setting, centralized DenseNet-121 reaches ACC $x'_{\text{zscore}} = \frac{x - \mu}{\sigma}.$1, AUC $x'_{\text{zscore}} = \frac{x - \mu}{\sigma}.$2 on T1-weighted scans, and ACC $x'_{\text{zscore}} = \frac{x - \mu}{\sigma}.$3, AUC $x'_{\text{zscore}} = \frac{x - \mu}{\sigma}.$4 on T2-weighted scans. FedAvg yields ACC $x'_{\text{zscore}} = \frac{x - \mu}{\sigma}.$5, AUC $x'_{\text{zscore}} = \frac{x - \mu}{\sigma}.$6 on T1-weighted scans and ACC $x'_{\text{zscore}} = \frac{x - \mu}{\sigma}.$7, AUC $x'_{\text{zscore}} = \frac{x - \mu}{\sigma}.$8 on T2-weighted scans. FedProx with $x'_{\text{zscore}} = \frac{x - \mu}{\sigma}.$9 produces the best reported federated T2-weighted AUC, $96 \times 96 \times 96$0. The drop from centralized to federated classification is therefore small, on the order of 1–2 AUC points, whereas segmentation degrades more substantially (Pan et al., 29 Jul 2025).

These results support a specific interpretation of Cyst-X’s technical contribution. The framework shows that privacy-preserving, multicenter MRI training is substantially more viable for risk classification than for pancreas segmentation. This suggests that future gains may depend less on the classifier backbone than on stronger cross-site anatomical localization and heterogeneity-aware segmentation.

6. Empirical performance against guidelines and radiologists

Cyst-X reports both radiomics and deep learning baselines. In binary high-risk versus non-high-risk classification, radiomics reaches AUC 72.93% on T1-weighted 2D features, 75.67% on T2-weighted 2D features, 77.98% on T1-weighted 3D features, and 75.75% on T2-weighted 3D features. For three-class classification with pooled data, DenseNet-121 reaches ACC $96 \times 96 \times 96$1, AUC $96 \times 96 \times 96$2 on T1-weighted scans and ACC $96 \times 96 \times 96$3, AUC $96 \times 96 \times 96$4 on T2-weighted scans. In the main binary clinical comparator, centralized DenseNet-121 on T2-weighted scans reaches AUC $96 \times 96 \times 96$5, which is the source of the headline AUC $96 \times 96 \times 96$6 reported for the framework (Pan et al., 29 Jul 2025).

The direct comparison with the Kyoto guidelines is one of the most consequential results. The Kyoto guidelines achieve AUC $96 \times 96 \times 96$7, whereas DenseNet-121 on T2-weighted MRI achieves AUC 0.82, with the improvement reported as statistically significant at $96 \times 96 \times 96$8. The high-risk sensitivity difference is larger still: DenseNet-121 reaches 87.8%, whereas Kyoto reaches 64.1%, with $96 \times 96 \times 96$9. This is clinically important because missed high-risk IPMNs are the more consequential error mode (Pan et al., 29 Jul 2025).

Three abdominal radiologists, each with 4–7 years of experience and more than 2,000 abdominal MRI reads, independently scored 629 cases using T1-weighted plus T2-weighted MRI and Kyoto-style criteria, but without clinical information, histology, or prior imaging. Their average no/low-risk accuracy was 93.91%, while their average high-risk accuracy was 46.01%. Reader-level high-risk accuracy ranged from 64.08% for the most aggressive reader to 32.39% for the most conservative. DenseNet-121 achieved no/low-risk accuracy of 91.38% on T1-weighted MRI alone, 86.11% on T2-weighted MRI alone, 95.69% under feature fusion, and 94.66% under probability fusion; for high-risk classification the corresponding values were 40.85%, 53.96%, 43.66%, and 47.18%. Probability-fusion DenseNet therefore slightly exceeded the average radiologist on high-risk classification while feature fusion exceeded the average radiologist on no/low-risk classification (Pan et al., 29 Jul 2025).

The link between anatomical localization quality and final risk prediction is also explicit. For T2-weighted three-class classification, AUC is 81.09% with radiologist ROIs, 74.18% with PanSegNet ROIs, and 69.63% with Swin-UNETR ROIs. In binary T2-weighted classification, the corresponding AUCs are 82.37%, 77.01%, and 68.92%. This establishes that segmentation quality is not a secondary issue: it directly determines downstream malignancy discrimination (Pan et al., 29 Jul 2025).

7. Interpretability, limitations, and broader significance

Cyst-X incorporates Grad-CAM and Information Bottleneck Attribution as interpretability tools. Grad-CAM highlights image regions by backpropagating class gradients into feature maps, whereas Information Bottleneck Attribution learns a perturbation mask that removes information from intermediate features while minimally affecting the prediction. In example cases, both methods concentrate attention on cyst walls, mural nodules, ductal segments with dilation, and solid components within or adjacent to cysts. These saliency patterns align with established clinical markers, while also suggesting that the network may exploit more distributed textural heterogeneity and multifocal gland-level patterns than current guideline rules explicitly encode (Pan et al., 29 Jul 2025).

The framework’s limitations are equally important. The study is retrospective, the MRI protocols evolved over two decades, and the multicenter cohort still requires broader demographic validation. Not all low-risk labels are histologically confirmed; some rely on at least 3 years of stable imaging follow-up, introducing a degree of label noise. Segmentation degrades materially in the federated setting. The federated design keeps raw data local, but it does not incorporate differential privacy or secure aggregation, so formal privacy guarantees are not claimed. These are not peripheral caveats: they define the frontier for subsequent work on clinically deployable multicenter pancreatic MRI models (Pan et al., 29 Jul 2025).

Cyst-X is also significant as an open research resource. The anonymized MRI scans, pancreas masks, and risk labels are publicly released through OSF, and the segmentation and classification code is available through GitHub. In practical terms, this makes Cyst-X a benchmark simultaneously for pancreas MRI segmentation, IPMN risk classification, domain generalization, and federated learning. A plausible implication is that its long-term importance may extend beyond IPMN risk prediction: by centering the pancreas rather than the manually delineated cyst, it provides a testbed for representation learning under weak lesion supervision, where risk is determined by distributed anatomical evidence rather than by a single segmented target (Pan et al., 29 Jul 2025).