Datasets:

YongchengYAO
/

PI-CAI-Lite

📝 Landmark annotations, visualization figures and the benchmark plan files live in 🔥MedVision🔥, where you can load the complete images and annotations from dataset configs.

Relation to the source dataset


In the source	1500 cases, each with T2W + ADC + HBV sequences; all 1500 now carry a human-expert csPCa delineation
Excluded here	the ADC and HBV sequences, and the cases whose expert mask is empty (no csPCa lesion to measure)
In this repo	425 `Images` + 425 `Masks`

Only cases with a non-empty human-expert lesion mask are kept.

picai_labels publishes expert csPCa delineations for all 1500 cases, in two disjoint folders: human_expert/resampled/ (1295 cases, already resampled onto the axial T2W grid) and human_expert/Pooch25/ (205 cases, added 2025-07-01 by Pooch et al., 2025 for the positives that previously carried only an AI mask). Both are used here.

Of those 1500, only a minority contain a delineated clinically significant lesion — the rest have an all-zero expert mask. A case with no lesion yields no measurement for the Tumor-Lesion-Size task, so empty-mask volumes are not redistributed.

Only the T2W sequence is included. PI-CAI is biparametric (T2W + ADC + HBV) and, per the upstream README, the original annotations were drawn at the resolution of the T2W, ADC or DWI/HBV image depending on the annotator — which is why the T2W-resampled delineations are the ones with an exact image/mask correspondence.

Why -Lite? The suffix marks this as a derived redistribution rather than a copy of the source. These are preprocessed volumes — every case has been format-converted, geometry-normalised and reoriented to RAS+ — and for some sources cases or modalities are excluded as well (see the table above). Use it to reproduce MedVision, not as a substitute for the original release. See Preprocessing below for exactly what was changed.

Preprocessing

T2W images converted from .mha to nii.gz; masks aligned to the T2W grid.
Masks encode the ISUP grade as the voxel value ({2,3,4,5}, with no label 1); they are binarized to {0,1} so label 1 means 'lesion'.
Images and masks standardized to RAS+ orientation.

Segmentation Labels

labels_map = {
    "1": "clinically significant prostate cancer lesion"
}

Landmarks

landmarks_map = {
    "P1": "most right/anterior/superior endpoint of the major axis",
    "P2": "most left/superior/inferior endpoint of the major axis",
    "P3": "most right/anterior/superior endpoint of the minor axis",
    "P4": "most left/superior/inferior endpoint of the minor axis"
}

News

[25 Jul, 2026] Initial release. This dataset is integrated into 🔥MedVision🔥, where you can use these config names to load data in python:
- PI-CAI_BoxSize_Task01_Axial_Test
- PI-CAI_BoxSize_Task01_Axial_Train
- PI-CAI_BoxSize_Task01_Coronal_Test
- PI-CAI_BoxSize_Task01_Coronal_Train
- PI-CAI_BoxSize_Task01_Sagittal_Test
- PI-CAI_BoxSize_Task01_Sagittal_Train
- PI-CAI_MaskSize_Task01_Axial_Test
- PI-CAI_MaskSize_Task01_Axial_Train
- PI-CAI_MaskSize_Task01_Coronal_Test
- PI-CAI_MaskSize_Task01_Coronal_Train
- PI-CAI_MaskSize_Task01_Sagittal_Test
- PI-CAI_MaskSize_Task01_Sagittal_Train
- PI-CAI_TumorLesionSize_Task01_Axial_Test
- PI-CAI_TumorLesionSize_Task01_Axial_Train

Data Usage Agreement

By using the dataset, you agree to the terms as follow.

You must comply with the original CC BY-NC 4.0 license terms of the source dataset.
You are recommended to refer to the source of this dataset in any publication: https://huggingface.co/datasets/YongchengYAO/PI-CAI-Lite
You must cite the original publication(s):
- https://doi.org/10.5281/zenodo.6667655
- https://doi.org/10.1101/2025.05.13.25327456

Official Release

For more information, please go to the official site: https://pi-cai.grand-challenge.org/

Download from Huggingface

# python
from huggingface_hub import snapshot_download
snapshot_download(repo_id="YongchengYAO/PI-CAI-Lite", repo_type='dataset', local_dir="/your/local/folder")

Downloads last month: 17