CoolFace
Datasetpublic

MedOtter/PI-CAI

PI-CAI: Prostate Imaging - Cancer AI Challenge (Public Training & Development) The PI-CAI Public Training & Development dataset contains 1,500 biparametric MRI (bpMRI) studies from 1,476 patients acquired at four Dutch centers (RUMC, ZGT, PCNN, UMCG) between 2012 and 2021. The challenge targets clinically significant prostate cancer (csPCa, ISUP ≥ 2) detection and segmentation. Dataset Summary Field Details Modality Biparametric MRI:… See the full description on the dataset page: https://huggingface.co/datasets/MedOtter/PI-CAI.

sourceHugging Facecc-by-nc-4.0updated 5mo agoView on Hugging Face
0likes2.1kdownloads
Dataset Card

PI-CAI: Prostate Imaging - Cancer AI Challenge (Public Training & Development)

The PI-CAI Public Training & Development dataset contains 1,500 biparametric MRI (bpMRI) studies from 1,476 patients acquired at four Dutch centers (RUMC, ZGT, PCNN, UMCG) between 2012 and 2021. The challenge targets clinically significant prostate cancer (csPCa, ISUP ≥ 2) detection and segmentation.

Dataset Summary

FieldDetails
ModalityBiparametric MRI: axial/coronal/sagittal T2W + axial high-b-value (≥ 1000 s/mm²) DWI + axial ADC
Body PartProstate (whole gland + csPCa lesions)
Cases1,500 studies / 1,476 patients
Class composition1,075 benign (ISUP ≤ 1) / 425 csPCa-positive (ISUP ≥ 2)
Total Size~26 GB (images) + 139 MB (labels)
LicenseCC BY-NC 4.0

Data Structure

images/
  {patient_id}/
    {patient_id}_{study_id}_t2w.mha   # axial T2W
    {patient_id}_{study_id}_cor.mha   # coronal T2W (occasionally missing)
    {patient_id}_{study_id}_sag.mha   # sagittal T2W (occasionally missing)
    {patient_id}_{study_id}_adc.mha   # axial ADC
    {patient_id}_{study_id}_hbv.mha   # axial high-b-value DWI
labels/
  csPCa_lesion_delineations/
    human_expert/
      original/         # native-resolution expert csPCa lesion masks (1,295 cases)
      resampled/        # SAME masks resampled to T2W axial geometry (1,295 cases)
      Pooch25/          # 2025 expert annotations for the 205 originally-only-AI-labeled cases
    AI/
      Bosma22a/         # AI-derived csPCa lesion masks (binary, 1,500 cases)
  anatomical_delineations/
    whole_gland/
      AI/Bosma22b/      # AI whole-prostate-gland masks (binary, 1,500 cases) — only gland labels
      AI/Guerbet23/     # alternate AI gland masks
    zonal_pz_tz/
      AI/HeviAI23/      # AI zonal masks (peripheral zone / transition zone)
      AI/Yuan23/        # alternate AI zonal masks
  clinical_information/
    marksheet.csv       # per-case clinical info (PSA, ISUP, center, etc.)
  additional_resources/
    ProstateX-mapping.json
marksheet.csv           # convenience copy at root

Recommended Ground Truth

For csPCa lesion segmentation, use:

labels/csPCa_lesion_delineations/human_expert/resampled/   # 1,295 cases
labels/csPCa_lesion_delineations/human_expert/Pooch25/     # +205 cases (= 1,500 total expert masks)

These are produced by trained investigators under three expert radiologists' supervision, carry granular multi-class ISUP labels (0/2/3/4/5), and are aligned to T2W axial geometry. The PI-CAI organizers explicitly recommend them over the AI-derived Bosma22a masks, which exist primarily as a complementary semi-supervised label source for the originally unannotated 205 positive cases (now superseded by Pooch25 for those cases).

For whole-prostate-gland segmentation, only AI-derived masks are available (anatomical_delineations/whole_gland/AI/Bosma22b/); the README in the picai_labels repo cautions these "can be susceptible to errors or faulty segmentations."

Important Caveats

  • —Modalities are NOT co-registered. T2W (axial/cor/sag), DWI, and ADC are acquired at different resolutions and not co-registered across sequences (only 54 / 9,107 cases were manually co-registered in the full corpus). Models must handle this.
  • —ADC absolute intensities are NOT standardized across centers — do not treat as universal.
  • —Splits: The 1,500 public cases are physically distributed across 5 cross-validation folds (fold0–fold4) in the original Zenodo release; here they are merged into a single images/ directory. There is no fixed train/val/test split.

Citation

bibtex
@article{saha2024picai,
  title   = {Artificial intelligence and radiologists in prostate cancer detection on MRI (PI-CAI): an international, paired, non-inferiority, confirmatory study},
  author  = {Saha, Anindo and Bosma, Joeran S and Twilt, Jasper J and van Ginneken, Bram and Bjartell, Anders and Padhani, Anwar R and others},
  journal = {The Lancet Oncology},
  volume  = {25},
  number  = {7},
  pages   = {879--887},
  year    = {2024},
  doi     = {10.1016/S1470-2045(24)00220-1}
}

@dataset{picai_public_training_2022,
  title     = {The PI-CAI Challenge: Public Training and Development Dataset},
  author    = {Saha, Anindo and Twilt, Jasper J and Bosma, Joeran S and others},
  year      = 2022,
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.6624726},
  url       = {https://doi.org/10.5281/zenodo.6624726}
}

Sources

  • —Original imaging: https://zenodo.org/records/6624726 (Zenodo, CC BY-NC 4.0)
  • —Original labels: https://github.com/DIAGNijmegen/picai_labels
  • —Challenge: https://pi-cai.grand-challenge.org/
  • —Preprocessing toolkit: https://pypi.org/project/picai-prep/
  • —Baseline code: https://github.com/DIAGNijmegen/picai_baseline
MedOtter/PI-CAI · CoolFace