PANORAMA - Public Training & Development Dataset

Contrast-enhanced (portal-venous) abdominal CT for pancreatic ductal adenocarcinoma (PDAC) detection, from the PANORAMA challenge organised by the Diagnostic Image Analysis Group (DIAG), Radboud UMC. This is the largest public PDAC CT dataset and the first public PDAC detection challenge.

Mirror of the official public training/development cohort only. The hidden validation and test sets used for the challenge leaderboard are not part of this release.

Contents (2238 cases)

Source (level)	Cases
Dutch centres (Radboud UMC + UMC Groningen)	1964
MSD Task07 Pancreas (level=MSD_dataset)	194
NIH Clinical Center Pancreas-CT (level=NIH_dataset)	80
Total	2238 (676 PDAC / 1562 non-PDAC)

images/                      2238 CT volumes,  {pid}_{sid}_0000.nii.gz
labels/manual/               482 expert-delineated masks (gold lesion tier)
labels/automatic/            1756 fully-automatic masks
clinical_information.csv      per-case metadata (+ .xlsx original)
data/                         preview parquet for the Dataset Viewer

Ground-truth annotation tiers

6-class voxel masks: 0 background, 1 PDAC lesion, 2 veins, 3 arteries, 4 pancreas parenchyma, 5 pancreatic duct, 6 common bile duct.

• labels/manual/ (482 cases) - GOLD STANDARD lesion reference. PDAC lesions manually delineated in ITK-SNAP by trained investigators under an expert radiologist (20+ yrs pancreatic-cancer experience). Use these for lesion benchmarking (lesion_gt_gold = True / annotation_tier = manual).
• labels/automatic/ (1756 cases) - weak/auto. Lesion (where present) and all anatomy classes (2-6, every case) were generated automatically by the Alves et al. (2021) algorithm. Treat as weak supervision.

Cross-dataset overlap (evaluation hazard)

This cohort embeds 194 cases from MSD Task07 Pancreas and 80 from the NIH Pancreas-CT collection. If you benchmark against Angelou0516/msd-pancreas or Angelou0516/pancreas-ct, exclude these via the level column (MSD_dataset / NIH_dataset) or the derived external_source column to avoid leakage. (No per-case ID map back to the original collections is published.)

License & citation

CC BY-NC 4.0 (non-commercial). Source: official DIAG/Radboud Zenodo records (imaging) + github.com/DIAGNijmegen/panorama_labels (annotations).

Alves, N., Schuurmans, M., Rutkowski, D., Yakar, D., Haldorsen, I., Liedenbaum, M., Molven, A., Vendittelli, P., Litjens, G., Hermans, J., & Huisman, H. (2024). The PANORAMA Study Protocol: Pancreatic Cancer Diagnosis - Radiologists Meet AI. Zenodo. https://doi.org/10.5281/zenodo.10599559