Request access to SiteShift
SiteShift contains de-identified clinical images and is released under a data-use agreement. Requests are reviewed manually. Access will open on publication; requests made before then are not processed.
By requesting access you agree to: (1) use the data for non-commercial research only; (2) make no attempt to re-identify any patient or the contributing institution; (3) not redistribute the data or any part of it; (4) cite the SiteShift paper in any work that uses the data; (5) hold, or obtain, approval from your institution's ethics board for this use where required.
Log in or Sign Up to review the conditions and access this dataset content.
SiteShift: A Cross-Site Benchmark for Prostate Tumour Segmentation
Status: under review. The data will be released here on publication. Access requests submitted before then are not processed.
SiteShift tests whether prostate tumour segmentation models trained on public data work at a hospital they were not trained for. It contains the first public prostate H&E segmentation data from Central Asia, with whole-tumour outlines on both needle biopsies and large (whole-mount) sections.
Contents
| Patients | 47 |
| Slides (H&E) | 194 |
| Needle biopsies | 89 slides, 25 patients |
| Large sections | 105 slides, 22 patients |
| Tiles | 449,623 tiles of 512 × 512 px at 0.5 µm/px |
| Annotation | Region-level whole-tumour outlines (QuPath), one pathologist |
| Splits | Patient-level 5-fold cross-validation |
The outlines include the stroma between tumour glands and follow the site's own annotation convention; they should be treated as that convention rather than as exact ground truth.
Benchmark tracks
The labels of one cross-validation fold are withheld and used to score a public leaderboard. Evaluation code, the patient-level protocol and the Detection–Delineation Attribution (DDA) metric are released with the data.
Planned layout
slides/ whole-slide images
annotations/ tumour outlines (GeoJSON)
splits/ patient-level folds
metadata.csv patient, slide, specimen type, fold
Citation
Citation will be added on publication.
- Downloads last month
- 391