slicepoint/slicepoint-hospitals

SlicePoint Hospitals

Seven simulated hospitals for studying dataset shift and non-IID federated learning on abdominal CT, with the answer keys. Each hospital is a sample of the SlicePoint AAA CT scans, and all but two have a problem planted on purpose: the kind of difference that makes an AI model behave differently at a new site. They were built for the 2026 Nucleate Pittsburgh hackathon, where teams had to find the problems. The game is over, so the keys are published.

Not for clinical use. AAA labels come from a 30 mm diameter rule on the ground-truth aorta masks that ship with TotalSegmentator (at hospital_e, a 25 mm lumen rule), not from radiologist reads. Predictions apply the same rule to the public TotalSegmentator model's masks; that model was trained on this dataset and has seen most of these scans, so its results are not an independent test (see the SlicePoint AAA CT card; crosswalk.parquet links each case to its scan there, and that card's tables/cases.parquet has the totalsegmentator_split).

The hospitals

Hospital	Scans (patients)	AAA	What is different	How clearly it shows
hospital_a	100 (100)	9	Nothing. A clean random sample; the reference the others are compared with	none
hospital_b	80 (40 people under 60 IDs)	7	Patients scanned again and again, some under new IDs; 2 patients reappear at hospital_e	strong (repeats), weak (cross-site)
hospital_c	80 (80)	7	Every scan thick-slice (5 mm) and venous phase: same patients, scanned differently	strong on acquisition, weak on Dice
hospital_d	80 (80)	20	Far more aneurysms, and older, more male, more arterial phase, all moving together	strong; the factors are confounded
hospital_e	80 (80)	7 by its own rule	Labels measure the inside of the vessel and call AAA at 25 mm: same images, different truth	strong
hospital_f	80 (80)	7	Nothing. Tests for false alarms	none
hospital_g	80 (80)	7	The exam name gives away the answer (a shortcut), plus 2 hospital_b patients under new IDs	weak

keys/<hospital>.json holds the full recipe of each hospital: the planted findings (waldos: kind, severity, description, the evidence a correct finding cites), the targets, what was realised, and one row per case (source scan, repeat or cross-site copy, new patient id or not). keys/practice/ has the reference measures given to teams for hospital_a to hospital_e, with the chance levels two clean samples would show.

crosswalk.parquet maps every case to its scan in SlicePoint AAA CT (canonical_image_id), its true patient, and whether its image was changed: hospital_c ships 5 mm re-makes, and the repeat scans at hospital_b and hospital_g have added noise and a few slices dropped, so no file is a byte copy of another.

Layout

Path	What
hospital_<x>/images/<case_id>.nii.gz	CT volume (1.5 mm; hospital_c 1.5 x 1.5 x 5 mm)
hospital_<x>/masks/<case_id>_aorta.nii.gz	Ground-truth aorta mask on the same grid
hospital_<x>/{features,labels,predictions,embeddings}.parquet	The SlicePoint input contract, joined on case_id
keys/, crosswalk.parquet	Answer keys, as above
manifest.json	The practice release manifest: row counts and sha256 of the hospital_a to hospital_e tables as first released (they are unchanged)

Column meanings are the same as in SlicePoint AAA CT. Race and ethnicity are not recorded in the source, so no hospital differs along those lines and none can be tested for them. study_description and protocol_name are synthetic; at hospital_g they carry the shortcut.

Download and use

hf download slicepoint/slicepoint-hospitals --repo-type dataset --exclude "*/images/*" "*/masks/*"   # tables + keys, ~10 MB
hf download slicepoint/slicepoint-hospitals --repo-type dataset --include "hospital_c/*"             # one site with images

slicepoint assess --data hospital_b --site-id hospital_b \
    --reference-dir hospital_a --reference-site-id hospital_a --html report.html

For federated-learning experiments, treat each hospital_<x> folder as one client; crosswalk.parquet tells you where clients share patients.

Source, license and citation

Derived from the TotalSegmentator dataset v2.0.1 (Wasserthal J. et al., University Hospital Basel, doi:10.5281/zenodo.10047292; acquisition details from v3.0.0, doi:10.5281/zenodo.22688904), licensed CC BY 4.0. Modified by Asher Informatics PBC (sampled, perturbed as described above, re-identified, derived tables added) and released under CC BY 4.0. Segmentations by the TotalSegmentator total model (Apache-2.0). Embeddings by Merlin (MIT). Please cite TotalSegmentator (Radiology: AI 2023, doi:10.1148/ryai.230024).