fidopt Interactive view
Prostate cancer on biparametric MR, segmentation

PI-CAI prostate cancer lesions

Biparametric prostate MR from one of three Dutch centres: T2, ADC and high b-value diffusion as three channels on the T2's grid, with the clinically significant cancers drawn by people (PI-CAI).

30 test casesprostateimage MRresearch only
Look at the images
Read before comparing (12)
  • Pool. PI-CAI's public set holds 1,500 studies of 1,476 men, 2012 to 2021, from Radboud, Groningen (PCNN) and Twente, 425 with a clinically significant cancer. The 1,075 benign or indolent studies are not cases, and the challenge's hidden test set of 1,000 is not public.
  • Held out. Every study of the 327 patients ProstateX-mapping.json names, since PROSTATEx_Seg draws from them, 111 of them with a lesion; and 10845_1000861 and 11154_1001177, whose lesion map is not on the T2's grid.
  • Channels. The diffusion images, on a grid of about 2 mm, are resampled onto the T2 in scanner coordinates by B-spline, as picai_prep does: the ADC dips below zero at edges, and both are 0 outside their field of view. No unit is given.
  • Motion. The T2 and diffusion images are separate acquisitions aligned only by scanner coordinates. A rigid fit of the ADC to the T2 around the gland moves the gland's centre by 0.8 to 9.8 mm on the 20 cases built first (median 2.7 mm), over 6 mm on 10354_1000360, 10636_1000652 and 10760_1000776. The channels stay as the challenge's methods got them.
  • Truth. Groningen's 109 studies carry PI-CAI's own maps, drawn in ITK-SNAP by trained investigators or a resident under expert radiologists, partly on the diffusion grid, and resampled by the provider onto the T2. Radboud's 125 and Twente's 80 carry Pooch et al.'s (2025), drawn by hand on the T2's grid. One reader a study; every lesion is label 1, whatever its grade.
  • Pooch maps. How they were drawn and checked is in a preprint the corpus could not read; its abstract says only that the 205 AI-annotated scans were manually annotated, and the repository calls them expert-derived.
  • Unlabelled lesions. Only clinically significant cancers are drawn: an ISUP grade group 1 lesion, a PI-RADS 3 lesion found benign or a cancer not seen on MR has no label, and a method that segments one is charged a false positive.
  • Lesions. 0.08 to 44 mL on the thirty test cases (median 1.0 mL), in one or two pieces. Groningen's maps are blocky on the T2 and lie partly outside the gland (57% inside the AI gland mask on 10486_1000494 and 10808_1000824); Pooch et al.'s follow the T2's edges.
  • Checks. On the 20 cases built first the lesion is darker on the ADC than the rest of the gland on 19, and brighter on the high b-value image on 18, as a significant cancer reads.
  • Overlap. Neither the 20 T2 volumes built first nor the 38 added later match the corpus's prostate MR images (PROMISE12, ProstateISBI2013, PROSTATEx_Seg, QINProstateRepeat, ProstateMRUS, MSD_Prostate) or the decathlon's 48 prostate studies: 69 and 108 pairs of one size compared voxel for voxel, best correlations 0.47 and 0.48.
  • Outside numbers. The challenge ranks lesion detection and diagnosis on its hidden test set, so no outside Dice compares. The test cases are public with their labels: a network trained on PI-CAI's public set has seen them, and Pooch et al.'s nnU-Net was trained on these maps.
  • Licence. CC BY-NC 4.0: non-commercial, scored locally, apart from the official leaderboards.

No method has a result on every case of this dataset yet.

Case by case

Click a score to look at that result: the image, the truth in green and the method's map in red. Underlined scores have their image; for the others you see the case without the result.

About the data

Test cases
30 test studies drawn with seed 2026 by centre among the 284 outside development (11 Radboud, 11 Groningen, 8 Twente), and 28 development studies, the last of the same seeded order from the centre with the most left (22 Radboud, 6 Groningen). The pool is the 312 studies, one a patient, whose human lesion map holds a clinically significant cancer, less PROSTATEx patients and two maps off the T2's grid.
Truth
contours by experts every clinically significant prostate cancer (ISUP grade group 2 or more) of the study, one label whatever its grade: on the Groningen (PCNN) studies the PI-CAI delineations, drawn in ITK-SNAP by one of 10 trained investigators or a radiology resident under one of 3 expert radiologists, mostly on the diffusion images' grid, and resampled by the provider onto the T2; on the Radboud and Twente studies the delineations drawn by hand for Pooch et al. (2025) on the T2's grid. The AI-derived lesion and gland maps of the release are not used.
Licence
CC BY-NC 4.0
28 cases for tuning, never tested

10636_1000652, 10811_1000827, 11283_1001306, 10365_1000371, 10318_1000324, 10760_1000776, 11346_1001369, 10660_1000676, 10565_1000577, 10050_1000050, 11055_1001076, 10651_1000667, 10799_1000815, 10679_1000695, 10560_1000572, 10545_1000556, 10471_1000479, 11208_1001231, 11350_1001373, 10274_1000279, 10140_1000142, 10304_1000310, 10717_1000733, 10490_1000499, 10763_1000779, 10970_1000989, 10369_1000375, 11447_1001471