PanSegData — Multi-center Abdominal MRI Pancreas Segmentation

767 abdominal MRI volumes (385 T1W + 382 T2W) with manual pancreas masks, collected across five US institutions between March 2004 and November 2022 and released by the Machine & Hybrid Intelligence Lab at Northwestern University. This is the MRI half of the PanSegNet study, and it is one of the very few large, multi-center, manually annotated pancreas MRI resources in existence.

The pancreas is among the hardest abdominal organs to segment — small, soft-tissue-isointense, and highly variable in shape and position. Here the foreground is on average 0.44 % of voxels.

⚠️ What this mirror contains — read first

MRI only. There is no CT here. The paper is titled "Large-scale multi-center CT and MRI segmentation of pancreas", but only the MRI is the authors' own releasable data. The paper's 1,350 CT scans are third-party public datasets (AbdomenCT-1K 1000, AMOS 200, WORD 120, BTCV 30) used only for benchmarking — they are not part of PanSegData and are not redistributed here. If you expected 2,117 volumes, you get 767.

Do not join t1 and t2 by filename. The two modalities were renumbered independently, so the same released id means different patients in 259 of the 370 ids that exist in both. Use the patient_uid column. See Cross-modality identity.

Do not reorient the masks by their own affine. 66 of the 385 T1 masks carry a sign-flipped direction matrix. See Mask affine defect.

Dataset Details

Field	Value
Modality	Abdominal MRI — T1-weighted (385) and T2-weighted (382)
Body part	Abdomen — pancreas, single foreground class
Labels	0 background · 1 pancreas (verified: the only values present in all 767 masks)
Patients	405 distinct, of whom 362 were imaged in both modalities
Centers	5 — NYU, Mayo Clinic Florida, Northwestern, Allegheny Health Network, Mayo Clinic Arizona
Acquisition window	March 2004 – November 2022
Format	.nii.gz (NIfTI-1), nnU-Net v1 raw layout upstream
Geometry	not resampled, not cropped, not co-registered — native per-case
In-plane	0.53 – 1.80 mm · 166 distinct shapes (T1), 165 (T2)
Slice thickness	1.8 – 10.4 mm — strongly anisotropic, typical 2D-acquired abdominal MRI
Slices per volume	20 – 148 (T1), 20 – 102 (T2)
Foreground	0.44 % of voxels; present on only ~43.5 % of slices
Empty masks	none — every one of the 767 volumes has annotated pancreas
License	CC BY-NC 4.0 — redistribution explicitly permitted, non-commercial only
Paper	Zhang et al., Medical Image Analysis 99:103382 (2025) · doi:10.1016/j.media.2024.103382
Official source	https://osf.io/kysnj/ (author-controlled; md5s verified against the OSF API)

Per-center composition (matches paper Table 2 cell-for-cell)

Center	Code	T1W	T2W	Split
NYU Medical Center	NYU	162	162	train
Mayo Clinic Florida	MCF	151	150	train
Northwestern University	NWU	30	19	test
Allegheny Health Network	AHN	17	28	test
Mayo Clinic Arizona	MCA	25	23	test
Total		385	382

Splits — center-based, and why it has to be

The official release has no split file: all 767 volumes ship under imagesTr/ + labelsTr/. This mirror materialises the paper's own protocol — centers 1–2 (NYU, Mayo FL) as the internal cohort, centers 3–5 (NWU, AHN, MCA) held out as the external test set:

Config	train	test	total
t1	313	72	385
t2	312	70	382

This is not merely convention. Center is a property of the patient, so a center-based split is automatically patient-disjoint across modalities as well. A random split would put some of the 362 dual-modality patients in t1/train and their other scan in t2/test. The center column is preserved so you can re-derive any other split, including the paper's 5-fold CV over the internal cohort.

⚠️ Cross-modality identity — the single biggest trap

The released ids (NYU_0001, MCF_0042, …) were assigned independently per modality. Each excluded scan shifts every subsequent number, and the two modalities excluded different scans (64 from T1, 58 from T2). Measured on this mirror:

released ids present in both t1 and t2	370
…that refer to the same patient	111
…that refer to a different patient	259

Mismatches by center: NYU 158, MCF 48, NWU 19, AHN 17, MCA 17. Example: AHN_0001 is original AHN05 in T1 but ahn_02 in T2.

Use patient_uid. It is derived from the authors' official T1-name_mapping.json / T2-name_mapping.json by folding case, separators and leading zeros (AHN05 and ahn_05 → ahn5). Verified globally unique: no original id is used by more than one center, and within a modality every patient contributes exactly one scan.

A note for anyone comparing against other write-ups: AHN is joinable across modalities. T1 spells it AHN05 and T2 spells it ahn_05; that is a formatting difference, not a missing key. All 17 T1 AHN patients are present in T2. Patients imaged in both modalities, by center: NYU 161, MCF 150, NWU 19, AHN 17, MCA 15 = 362.

metadata/patient_crossref.csv gives the full patient_uid ↔ (t1 id, t2 id) table.

Unreconciled count

The paper states "767 scans from 499 adult participants". Tracing all 767 released scans back through the official mapping files yields 405 distinct original patient ids, not 499. The mapping files list 889 original scans of which 122 are marked "Not included". The most plausible reading is that 499 counts the full collected cohort while 405 is what was actually released, but the paper attributes 499 directly to the 767, so this is flagged rather than silently resolved. Nothing downstream depends on it; patient_uid is derived from the released scans only.

⚠️ Mask affine defect — 66 T1 cases

In 66 of 385 T1 cases the mask's affine disagrees with the image's. The images are LPS while the masks read RAS (28 cases) or RAI (38 cases) — all of them in the MCF center. T2 is unaffected (0 cases).

The affines differ only in the sign of the diagonal, with identical zooms and identical array shapes:

MCF_0001   image affine diag = [-1.484, -1.484,  7.7]     zooms (1.484, 1.484, 7.7)
           mask  affine diag = [+1.484, +1.484,  7.7]     zooms (1.484, 1.484, 7.7)

That is the signature of a mask written without direction cosines, not of genuinely reordered voxel data. Measured here to confirm it, using the fact that MRI background outside the body is ≈0 — a correctly aligned pancreas mask must not sit on air:

Alignment	mean fraction of mask voxels below the body/air threshold
Raw voxel grid (ignore the mask affine)	0.109 ✅
Reorient each by its own affine	0.312 ❌
Control: concordant cases, raw	0.248 mean / 0.085 median

Raw voxel alignment wins in 64 of 66 cases. (The 2 exceptions, MCF_0052 and MCF_0107, score >0.69 under both alignments — the heuristic simply fails on them; the control's p90 of 0.85 shows such values occur in perfectly concordant cases too. They are not counter-evidence.)

What to do: take geometry from the image, and apply the same transform to the mask. Never call nib.as_closest_canonical() — or any reorientation — on the mask independently. Doing so is the natural way to handle this dataset's mixed LPI/LPS/RAS orientations, and it silently corrupts 17 % of the T1 set.

The headers are deliberately not patched, so this mirror stays byte-identical to the official release. The affected cases are flagged by mask_affine_unreliable: true in the jsonl.

Other geometry gotchas

• Mixed orientation across cases: T1 is LPI 223 / LPS 143 / RAS 19; T2 is LPI 251 / LPS 131. Reorient to a canonical frame using the image affine or L/R and S/I flip silently between cases.
• Mixed image dtype: T1 float32 330, int16 35, uint16 19, float64 1; T2 int16 362, uint16 19, float32 1. Masks are uint16 (766) / uint8 (1). Do not assume a dtype; do not assume a fixed intensity range (no upstream normalisation).
• Sparse foreground: 0.44 % of voxels, and only ~43.5 % of slices carry any pancreas (min 10.8 %). A slice sampler with a small budget can easily draw only empty slices.
• Anisotropy: through-plane spacing is 1.8–10.4 mm against 0.53–1.80 mm in-plane. Any 3D model or physical-unit metric must read real spacing from the header.

Ground truth — single tier, 100 % manual

"Five radiologists (one per center) manually segmented the pancreas on axial T1W and T2W MRI scans using ITK-SNAP. A senior radiologist double-checked the annotations for quality and consistency." — paper §3.3

There is exactly one mask per volume and no alternative rater tier is distributed. No AI-assisted, semi-automatic, or pseudo-labelled masks anywhere — PanSegNet is trained on these labels and was never used to produce them. Annotation took ≈25 min per scan, following a protocol agreed among the radiologists beforehand.

Human ceiling (the authors' own sub-studies; the repeat annotations are not released):

	T1W	T2W
Inter-observer Dice (50 scans)	0.8014	0.8058
Inter-observer Cohen's κ	0.624	0.638
Intra-observer Dice (20 scans, 4-week washout)	0.960	0.936

Inter-observer agreement is only moderate by the authors' own description — a model at Dice ≈0.80 is already at the human agreement ceiling.

✅ Cross-dataset overlap: none

Unusually clean for a pancreas dataset. The 767 MRI volumes are newly collected, IRB-approved private hospital data, not curated from any public archive. They therefore cannot overlap NIH Pancreas-CT, MSD Task07_Pancreas, AbdomenCT-1K, FLARE22/23, AMOS, WORD, BTCV, or PANORAMA — every one of those is CT, and the AMOS MRI subset is a disjoint public cohort.

The overlap risk in the paper lives entirely in the CT half, which is not shipped here: AbdomenCT-1K is itself curated from 12 centers including NIH and MSD, so it transitively contains both NIH Pancreas-CT and MSD Task07_Pancreas. Keep that in mind if you ever pair this MRI set with those CT benchmarks.

Structure

t1/train/images/NYU_0001_0000.nii.gz     # 313 T1W volumes (NYU + MCF)
t1/train/masks/NYU_0001.nii.gz           # 313 masks, same voxel grid
t1/test/images/  t1/test/masks/          #  72 T1W volumes (NWU + AHN + MCA)
t2/train/...  t2/test/...                # 312 / 70 T2W volumes

t1_train.jsonl  t1_test.jsonl            # per-case metadata
t2_train.jsonl  t2_test.jsonl

metadata/T1-name_mapping.json            # official, verbatim from OSF
metadata/T2-name_mapping.json
metadata/t2_info_osf.xlsx                # official T2 demographics (408 rows)
metadata/patient_crossref.csv            # patient_uid <-> t1 id <-> t2 id  (derived here)
metadata/demographics.csv                # xlsx joined onto released ids via patient_uid

data/*.parquet                           # display-only preview, see below
README.md  LICENSE.txt

Filenames are the official ones, including the nnU-Net _0000 channel suffix on images.

jsonl columns

Column	Meaning
case_id	released id, e.g. "NYU_0001"
center, center_name	"NYU" … / full institution name
modality	"T1W" or "T2W"
image, mask	repo-relative paths
split	"train" (NYU+MCF) or "test" (NWU+AHN+MCA)
patient_uid	cross-modality patient key — group on this, never on case_id
original_name	the pre-anonymisation filename from the official mapping
paired_case_id	the same patient's id in the other modality, or null
has_other_modality	whether this patient also appears in the other config
n_slices, shape_xyz	geometry
spacing_xyz, slice_thickness_mm	real spacing from the header
axcodes_image, axcodes_mask	orientation as declared by each file
mask_affine_unreliable	true for the 66 sign-flipped T1 masks
image_dtype, mask_dtype	not constant — see gotchas
intensity_min, intensity_max	per case, never renormalised
label_values	always [0, 1]
fg_voxels, fg_fraction	pancreas volume in voxels / fraction
n_fg_slices, fg_slice_fraction	how many slices carry pancreas
demographics	{gender, age, mri_brand, magnet_strength} where known, else null

The parquet preview layer is display-only

data/*.parquet exists so the HF Dataset Viewer can render this dataset. Each row holds one slice of one volume as PNG: image (grayscale, percentile-windowed), mask, and overlay.

The preview renders the slice with the LARGEST pancreas area, not the middle slice. With foreground on only ~43.5 % of slices, a middle-slice preview would show an empty mask for over half the dataset. slice_index records which slice was rendered.

Do not train or evaluate on the preview. Its intensities are percentile-windowed to 8-bit for display and it holds one slice per volume. The real data is the byte-identical .nii.gz at the repo root.

Source & Citation

• Official data: https://osf.io/kysnj/ — CC BY-NC 4.0, anonymous download, no registration.
• Code (PanSegNet): https://github.com/NUBagciLab/PaNSegNet
• The medsam-datasetlist entry records the license as "GPL 3.0"; that is wrong — see LICENSE.txt in this repo for the evidence.

@article{zhang2025pansegnet,
  author  = {Zhang, Zheyuan and Keles, Elif and Durak, Gorkem and Taktak, Yavuz and
             Susladkar, Onkar and Gorade, Vandan and Jha, Debesh and
             Ormeci, Asli C. and Medetalibeyoglu, Alpaslan and Yao, Lanhong and
             Wang, Bin and Isler, Ilkin Sevgi and Peng, Linkai and Pan, Hongyi and
             Vendrami, Camila L. and Bourhani, Amir and Velichko, Yury and
             Gong, Boqing and Spampinato, Concetto and Pyrros, Ayis and
             Tiwari, Pallavi and Klatte, Derk C. F. and Engels, Megan and
             Hoogenboom, Sanne and Bolan, Candice W. and Agarunov, Emil and
             Harfouch, Nassier and Huang, Chenchan and Bruno, Marco J. and
             Schoots, Ivo and Keswani, Rajesh N. and Miller, Frank H. and
             Gonda, Tamas and Yazici, Cemal and Tirkes, Temel and
             Turkbey, Baris and Wallace, Michael B. and Bagci, Ulas},
  title   = {Large-scale multi-center CT and MRI segmentation of pancreas with deep learning},
  journal = {Medical Image Analysis},
  volume  = {99},
  pages   = {103382},
  year    = {2025},
  doi     = {10.1016/j.media.2024.103382}
}