Cyst‑X Dataset: Overview & Benchmarks

• Cyst‑X is a reused label referring to diverse cyst imaging collections across pancreatic CT, retinal OCT, and multicenter pancreatic MRI.
• Each variant targets specific tasks with unique supervision: 3D lesion segmentation for CT, dual-grader cyst segmentation for OCT, and risk prediction with pancreas segmentation for MRI.
• Benchmark results use metrics like Dice coefficients to highlight the challenges in different modalities, emphasizing the need for precise task and anatomy specification.

The designation Cyst‑X Dataset does not denote a single, universally fixed benchmark across the arXiv literature. Instead, it is used for multiple cyst-related imaging resources in different subfields. In pancreatic imaging, later literature and project materials refer to a 2017 abdominal CT collection for pancreas and pancreatic cyst segmentation as Cyst‑X (Zhou et al., 2017). In ophthalmic imaging, the OCT community uses Cyst‑X or OPTIMA Cyst‑X for the MICCAI 2015 OPTIMA cyst segmentation challenge dataset for intraretinal cysts (Gopinath et al., 2017, Dharmaratnakar et al., 12 Apr 2026). In 2025, Cyst‑X became the formal name of a publicly released multicenter pancreatic MRI dataset for intraductal papillary mucinous neoplasm (IPMN) malignancy prediction and federated learning (Pan et al., 29 Jul 2025). A practical implication is that references to “Cyst‑X” are underspecified unless the imaging modality, anatomy, and task are stated explicitly.

1. Terminology and scope

In the literature summarized here, Cyst‑X is a reused label rather than a unique canonical dataset name. The ambiguity is historically understandable: earlier authors introduced cyst-focused datasets without formal names, and later papers or project pages attached the shorthand retrospectively; the 2025 pancreatic MRI release then formalized Cyst‑X as an official dataset name (Zhou et al., 2017, Pan et al., 29 Jul 2025).

Usage in literature	Modality and domain	Core content
Later-named pancreatic Cyst‑X	Contrast-enhanced abdominal CT	131 pathological samples with pancreas and cyst masks
OPTIMA / OCSC / Cyst‑X	Multi-vendor SD-OCT	30 OCT volumes with dual-grader retinal cyst annotations
Formal Cyst‑X release	Multicenter pancreatic MRI	1,461 MRI volumes from 764 patients with pancreas masks and IPMN risk labels

This naming overlap matters because the three resources differ not only in modality but also in supervision granularity, target anatomy, and benchmark objective. The CT resource is a 3D lesion segmentation dataset; the OCT resource is a multi-vendor retinal cyst segmentation challenge; the MRI resource is primarily a pancreas-segmentation and malignancy-risk-classification benchmark (Zhou et al., 2017, Gopinath et al., 2017, Pan et al., 29 Jul 2025).

2. The pancreatic CT resource later referred to as Cyst‑X

The 2017 paper on pancreatic cyst segmentation introduced a new dataset with 131 pathological samples, described as 131 contrast-enhanced abdominal CT volumes with cystic pancreata (Zhou et al., 2017). The paper itself did not assign a formal dataset name, but later literature and the project webpage often referred to it as the Cyst‑X dataset. Each volume has in-plane resolution 512 × 512, slice count $L \in [358, 1121]$, and 0.5 mm – 1.0 mm slice thickness. Every case is manually labeled with both pancreas segmentation $\mathbf{P}^\star$ and cyst segmentation $\mathbf{C}^\star$, giving voxel-level 3D supervision for organ and lesion.

The dataset encodes a strong anatomical prior: in 121 out of 131 cases, more than 95% of cyst voxels lie within the pancreas volume,

$$\frac{|\mathcal{P}^\star \cap \mathcal{C}^\star|}{|\mathcal{C}^\star|} > 95\%.$$

This relation underpins the paper’s two-stage design, in which pancreas localization guides cyst segmentation. The class imbalance is severe: the pancreas occupies <1% of voxels in a CT volume, while the cyst fraction is often much smaller than 0.1% and can be approximately 0.0015% of the entire volume, or about 1.5% of the pancreas.

The benchmark protocol uses 4 fixed folds with 4-fold cross-validation. The model slices each 3D CT along coronal, sagittal, and axial directions and trains a 2D FCN-8s based on VGG-16, initialized from Pascal VOC segmentation weights. For cyst segmentation, the input is transformed by masking around the pancreas prediction with a distance threshold $t = 15$ voxels. The principal metric is the Dice-Sørensen Coefficient (DSC), reported per 3D volume.

The reported cyst-segmentation numbers established an early fully automatic baseline: 60.46 ± 31.37 DSC without deep supervision, 63.44 ± 27.71 with deep supervision, and 77.92 ± 12.83 when using a ground-truth cyst bounding box as an oracle region of interest (Zhou et al., 2017). The paper also states that the method missed 8 cyst cases with deep supervision and 16 without it. Historically, this dataset was presented as, to the authors’ knowledge, the largest set for pancreatic cyst segmentation at the time.

3. The OCT intraretinal cyst challenge dataset often called Cyst‑X

In retinal OCT, Cyst‑X typically refers to the MICCAI 2015 OPTIMA Cyst Segmentation Challenge (OCSC) dataset (Gopinath et al., 2017). Later literature explicitly equates the challenge data with Cyst‑X or OPTIMA / Cyst‑X, and describes it as the first publicly available novel cyst segmentation challenge dataset (Dharmaratnakar et al., 12 Apr 2026). The dataset contains 30 OCT volumes total, split into 15 training volumes and 15 testing volumes, acquired from 4 different SD‑OCT vendors: Cirrus, Nidek, Spectralis, and Topcon (Gopinath et al., 2017).

The OCT resource is heterogeneous. The 2017 challenge-oriented description reports lateral resolution ranging from $496 \times 512$ to $1024 \times 512$ pixels and 5 to 200 B‑scans per volume (Gopinath et al., 2017). Another later paper, focusing on implementation on the same challenge data, states that Spectralis examples consist of four OCT volumes, each with 49 frames, and that after converting all 3D volumes to 2D slices the authors obtained 1676 training images and 909 testing images (Dharmaratnakar et al., 12 Apr 2026). These are not contradictory; they reflect different views of the same multi-vendor challenge corpus.

Annotation is dual-grader and pixel-wise. The challenge provides manual cyst segmentations from Grader 1 and Grader 2, and different papers use different label combinations. One study trained with the intersection $G1 \cap G2$ and reported performance against G1, G2, and $G1 \cap G2$ (Gopinath et al., 2017). Another trained on the union of the two graders to increase positive samples and evaluated against Grader 1, Grader 2, and the intersection of Grader 1 & 2 (Dharmaratnakar et al., 12 Apr 2026). This grader structure makes annotation variability part of the benchmark itself.

The dataset became a focal point for methodological progression. A domain-knowledge-assisted pipeline using TV denoising, ILM–RPE restriction, MSER candidates, and a Random Forest with 50 trees achieved a mean Dice coefficient of 0.3893 with standard deviation 0.2987 on the training set under leave-one-out validation, with particularly weak detection for small cysts (Gopinath et al., 2016). A later selective-enhancement method based on Generalized Motion Patterns and a CNN achieved mean Dice 0.71 on the OCSC test set in masked evaluation against $G1 \cap G2$, outperforming published challenge participants (Gopinath et al., 2017). A 2026 ResNet18 patchwise classifier on the same public challenge resource reported overall mean Dice 0.8255 against Grader 1, with vendor-level means 0.888 for Cirrus, 0.8217 for Nidek, 0.8276 for Spectralis, and 0.7639 for Topcon, alongside very high precision and reduced sensitivity on noisy Topcon scans (Dharmaratnakar et al., 12 Apr 2026).

4. The formal 2025 Cyst‑X multicenter pancreatic MRI dataset

The 2025 paper “Cyst‑X: AI-Powered Pancreatic Cancer Risk Prediction from Multicenter MRI in Centralized and Federated Learning” introduced the first resource in this group for which Cyst‑X is the formal dataset name (Pan et al., 29 Jul 2025). It is described as the first large-scale, multi-center pancreatic cysts MRI dataset, publicly released to support IPMN malignancy prediction, pancreas segmentation, and federated learning.

Cyst‑X comprises 764 unique patients and 1,461 total MRI volumes, including 723 T1-weighted and 738 T2-weighted scans, collected from 7 institutions: New York University Langone Health, Mayo Clinic Florida, Northwestern University, Allegheny Health Network, Mayo Clinic Arizona, Istanbul University, Istanbul Faculty of Medicine, and Erasmus Medical Center (Pan et al., 29 Jul 2025). The study period spans March 2004 – June 2024. Vendors include GE, Siemens, and Philips, with 1.5T and 3T field strengths represented.

Each scan includes the MRI volume, a ground-truth pancreas segmentation mask, and a risk/malignancy label. The label space is three-class: no risk, low risk, and high risk. High-risk IPMN denotes high-grade dysplasia or worse, including carcinoma in situ and invasive carcinoma; low-risk IPMN includes histologically confirmed low- or intermediate-grade dysplasia or lesions without progression over $\mathbf{P}^\star$0 years of follow-up; and no-risk cases are normal pancreas or benign cysts not associated with IPMN (Pan et al., 29 Jul 2025). The paper emphasizes that the dataset predominantly contains images of IPMNs (90–95%).

Unlike the CT and OCT resources, this MRI benchmark does not release cyst-only lesion masks. The segmentation task is whole-pancreas segmentation, and the primary downstream prediction task is malignancy risk classification from the pancreas ROI. The dataset is distributed in NIfTI format, converted from de-identified DICOM, and is accompanied by a code repository implementing PanSegNet, Swin‑UNETR, DenseNet‑121, and federated-learning baselines. The paper gives the public links https://osf.io/74vfs/ for the dataset and https://github.com/NUBagciLab/Cyst-X for code (Pan et al., 29 Jul 2025).

The benchmark results are correspondingly broader than segmentation alone. PanSegNet reached Dice $\mathbf{P}^\star$1 on T1W and $\mathbf{P}^\star$2 on T2W. For binary high-risk versus not-high-risk IPMN classification, DenseNet‑121 achieved AUC $\mathbf{P}^\star$3 on T1W and AUC $\mathbf{P}^\star$4 on T2W, exceeding the reported Kyoto guidelines (AUC ≈ 0.75) (Pan et al., 29 Jul 2025). The same paper also showed that federated classification performed close to centralized training, whereas federated segmentation remained materially harder.

5. Labels, tasks, and benchmark conventions across the three usages

The three Cyst‑X usages define substantially different machine-learning problems. The commonality is cyst-centered medical imaging; the divergence lies in supervision level and endpoint.

Resource	Supervision structure	Primary endpoint
Pancreatic CT	3D pancreas mask + 3D cyst mask	Cyst segmentation DSC
Retinal OCT	Dual-grader pixel-wise cyst masks	Cyst segmentation Dice, precision, sensitivity
Pancreatic MRI	Whole-pancreas mask + scan-level risk label	Pancreas Dice; IPMN ACC/AUC

For the CT and OCT variants, the dominant overlap metric is the Dice-Sørensen Coefficient,

$\mathbf{P}^\star$5

reported per 3D volume in pancreatic CT and in challenge-style slice/volume aggregates in OCT (Zhou et al., 2017, Gopinath et al., 2017). OCT studies additionally report precision and sensitivity, and sometimes distinguish masked from unmasked evaluation in the central macular region (Gopinath et al., 2017, Dharmaratnakar et al., 12 Apr 2026). The MRI release broadens the evaluation vocabulary to Dice, Jaccard, precision, recall, HD95, ASSD, ACC, and AUC, reflecting its hybrid segmentation-plus-classification design (Pan et al., 29 Jul 2025).

This suggests that “Cyst‑X” is best understood not as a single benchmark family with one invariant protocol, but as a label applied to datasets that foreground cyst-related pathology while varying sharply in anatomical target, task formulation, and annotation granularity.

6. Misattributions and adjacent resources

Several cyst-related papers do not define a dataset called Cyst‑X, even when the topic might invite that assumption. The dental radiology paper “Radious” uses “our own dataset” from Valiasr Hospital in Tehran, with 963 OPG X-rays, 514 periapical X-rays, 3673 bitewing X-rays, and 466 annotated images, but it explicitly does not introduce or mention any dataset called Cyst‑X (Mashayekhi et al., 2023). Its dataset includes labels such as radicular cyst, periapical cyst, and cyst, but remains unnamed and private.

Similarly, the kidney organoid video paper on Organoid Tracker does not introduce a dataset called Cyst‑X. It presents an analysis platform and demonstrates it on one PKD mutant video and one wild-type video, with a 7-frame time-lapse bright-field microscopy video used in the case study, but no named external benchmark is released under that title (Huang et al., 14 Sep 2025). The breast spectral mammography study on cyst fluid attenuation also does not define a formal dataset named Cyst‑X; it reports measurements on 50 breast cyst fluid samples and 50 water samples, which function more as a well-characterized measurement cohort than a packaged public dataset (Fredenberg et al., 2021).

A recurring misconception, therefore, is to treat Cyst‑X as a generic shorthand for any cyst-imaging collection. The literature here does not support that usage. In practice, the term refers either to the later-named pancreatic CT set, to the OCT OPTIMA challenge corpus, or to the formal 2025 multicenter pancreatic MRI release; outside those contexts, it should be regarded as a misattribution unless a paper states otherwise (Zhou et al., 2017, Gopinath et al., 2017, Pan et al., 29 Jul 2025).