Skip to content
Open Imaging Index
MRI Prostate

PI-CAI

PI-CAI (Prostate Imaging: Cancer AI) Public Training and Development Dataset

Prostate MRI with labels and lesion outlines for clinically significant prostate cancer, released for a challenge comparing AI systems with radiologists at detecting it. 1,500 exams of 1,476 men from three Dutch centres, with clinical variables.

Go to dataset Edit Report

Overview

PI-CAI is a grand challenge on detecting clinically significant prostate cancer (csPCa, ISUP grade group 2 or higher) in prostate MRI, comparing AI systems with radiologists. Its Public Training and Development Dataset was released in 2022 by the Diagnostic Image Analysis Group at Radboud University Medical Center with partners in Groningen and Twente. This entry covers that public part only. The private training data and the hidden validation and testing cohorts are not released.

Composition

The public set holds 1500 MRI studies from 1476 patients, scanned between 2012 and 2021. 425 studies are labelled csPCa and 1075 contain benign tissue or indolent cancer. 328 cases come from the earlier ProstateX challenge, so the two should not be combined. Each study has an axial T2-weighted image, an axial high b-value diffusion image and an ADC map, and some also have sagittal or coronal T2-weighted images. No contrast-enhanced sequences are included. Images are MHA files, distributed as five zip archives named fold0 to fold4.

Acquisition

Exams were acquired at Radboud University Medical Center, University Medical Center Groningen (the clinical table lists this center as PCNN, Prostaat Centrum Noord-Nederland) and Ziekenhuisgroep Twente. The study protocol reports 1.5T and 3T scanners from Siemens and Philips with surface coils: five Siemens and two Philips scanners in the public set.

Annotations

Labels are maintained in the separate picai_labels repository. Positives are histologically confirmed ISUP 2 or higher; negatives are confirmed by histology (ISUP 1 or lower) or by MRI (PI-RADS 2 or lower), without follow-up. Lesions were delineated in ITK-SNAP by trained investigators supervised by expert radiologists. The original expert masks carry the ISUP grade, while the AI-derived and follow-up masks are binary. Expert masks cover 1295 cases; the remaining 205 positives first had only AI-derived masks and later received expert masks from a follow-up study. AI-derived whole-gland masks and a per-study clinical table (age, PSA, PSA density, prostate volume, biopsy type, PI-RADS and Gleason scores per lesion, center) are included.

Known limitations

  • Cases were selected by convenience sampling, so the case mix does not reflect routine care.
  • Some patients have several studies with different outcomes, each labelled on its own.
  • The README on Zenodo and the study protocol differ slightly on median age and lesion counts.
  • The whole-gland masks are automatic and contain known errors.

Cohort

Aggregate numbers from the sources below. Bars are relative to the 1,476 subjects.

Sex

  • Male 1,476 100%

p. 9, Item 22a in Saha et al. 2022, PI-CAI challenge study protocol (BIAS), Zenodo 6522364

Contrast / sequence

Studies

Groups can overlap

  • T2-weighted 1,500 100%
  • Diffusion-weighted 1,500 100%
  • ADC map 1,500 100%

Imaging Files in Zenodo record 6624726 v2.0, PI-CAI Public Training and Development Dataset, README.md

Country

Studies
  • Netherlands 1,500 100%

Dataset Characteristics in Zenodo record 6624726 v2.0, PI-CAI Public Training and Development Dataset, README.md

Age

Studies

mean 65.6 ± 7.2, median 66, range 35 to 92

0
200
400
600
770
30-4950-5960-6970-7980+

patient_age in picai_labels clinical_information/marksheet.csv (CC BY-NC 4.0), one row per study; Dataset Characteristics in Zenodo record 6624726 v2.0, PI-CAI Public Training and Development Dataset, README.md

Condition

Studies

Groups can overlap

  • Prostate cancer 653 44%

case_ISUP in picai_labels clinical_information/marksheet.csv (CC BY-NC 4.0), one row per study

Contrast combinations

Studies

How many studies have exactly each set of contrasts.

T2wdwiADCSubjects with exactly this set
1,500

Imaging Files in Zenodo record 6624726 v2.0, PI-CAI Public Training and Development Dataset, README.md

License and access

Our reading of the license, not legal advice. Before you use the data, read the original license and confirm that your use is allowed. We take no responsibility for how you use a dataset. Full disclaimer

Access
Open download

Download without an account

Access page

Creative Commons Attribution-NonCommercial 4.0 International

Use, share and adapt the data with credit, but only for non-commercial purposes.

Original license text Version read: 4.0 Checked 2026-10-07

What you can do

  • No
  • Not stated
  • Conditional
  • Yes
  • Yes

What you can share

  • Conditional
  • Conditional
  • Conditional

What you must do

  • Yes
  • Share alike No
  • No
  • No
  • Manuscript review No
  • Release code No
  • Return results No
  • Delete after use No

Limits

  • No
  • Location limits No

Citation

Saha A, Bosma JS, Twilt JJ, et al. Artificial intelligence and radiologists in prostate cancer detection on MRI (PI-CAI): an international, paired, non-inferiority, confirmatory study. Lancet Oncol 2024; 25: 879-887. https://doi.org/10.1016/S1470-2045(24)00220-1

Sources

Every number on this page comes from one of these documents. Each chart names the table or page it is taken from. The raw numbers are in stats.csv.