README.md

metadata

license: cc-by-4.0
pretty_name: SlicePoint AAA CT
language:
  - en
tags:
  - medical
  - ct
  - abdominal-aortic-aneurysm
  - segmentation
  - nifti
  - distribution-shift
  - totalsegmentator
size_categories:
  - n<1K
task_categories:
  - image-segmentation
  - image-classification
configs:
  - config_name: cases
    data_files: tables/cases.parquet
  - config_name: features
    data_files:
      - split: all
        path: cohorts/all/features.parquet
      - split: vendor_train
        path: cohorts/vendor_train/features.parquet
      - split: site_retrospective
        path: cohorts/site_retrospective/features.parquet
  - config_name: labels
    data_files:
      - split: all
        path: cohorts/all/labels.parquet
      - split: vendor_train
        path: cohorts/vendor_train/labels.parquet
      - split: site_retrospective
        path: cohorts/site_retrospective/labels.parquet
  - config_name: predictions
    data_files:
      - split: all
        path: cohorts/all/predictions.parquet
      - split: vendor_train
        path: cohorts/vendor_train/predictions.parquet
      - split: site_retrospective
        path: cohorts/site_retrospective/predictions.parquet
  - config_name: embeddings
    data_files:
      - split: all
        path: cohorts/all/embeddings.parquet
      - split: vendor_train
        path: cohorts/vendor_train/embeddings.parquet
      - split: site_retrospective
        path: cohorts/site_retrospective/embeddings.parquet

SlicePoint AAA CT

567 abdominal CT scans from the TotalSegmentator dataset, with ground-truth masks, aortic measurements, abdominal aortic aneurysm (AAA) labels, a baseline model's predictions and image embeddings. The AAA labels are not radiologist reads: they come from a 30 mm threshold on the aortic diameter measured on the ground-truth aorta masks that ship with the TotalSegmentator dataset. The baseline predictions apply the same rule to masks from the public TotalSegmentator model. The scans were collected through University Hospital Basel but come from 10 coded imaging institutes and 15 scanner models (Siemens, Philips, GE); the institute column in tables/cases.parquet says which. It is the canonical dataset of SlicePoint, an open-source tool that checks whether an imaging AI model fits a hospital's patients and scanners before it is switched on.

Not for clinical use. The AAA labels come from a diameter rule applied to segmentation masks, not from radiology reports or clinical diagnoses.

What is in it

Path	What
images/<id>.nii.gz	CT volume, 1.5 mm isotropic, int16 HU (567 files)
masks/<structure>/<id>.nii.gz	Ground-truth binary masks: aorta, kidney_left, kidney_right, liver, spleen, iliac_artery_left, iliac_artery_right
tables/cases.parquet	Every scan: age, sex, institute (coded a to j), scanner, kVp, study type, pathology, voxel spacing, eligible and exclusion_reason
tables/pipeline/	All computed measurements: four aortic diameter methods (outer wall / lumen, axial / centerline-orthogonal) on the ground-truth (_ref) and model (_pred) masks, organ volumes, Dice per structure, Merlin embeddings
cohorts/<name>/	The SlicePoint input contract (features, labels, predictions, embeddings parquet, joined on case_id) for three cohorts below
vendor_datasheet/	Vendor datasheets (summary statistics only, in the SlicePoint VendorDatasheet format): aaa-baseline-v0.json summarises vendor_train, as a vendor might disclose it; totalsegmentator-v2-train-all.json and totalsegmentator-v2-train-abdominal.json describe the TotalSegmentator model's real training data (all 1,082 training scans, every body region, and the 481 of them in this dataset), computed from the source metadata
manifest.json	Counts and sha256 of every table

case_id is TotalSegmentator's own image id (s0001...), so every row traces back to the source dataset. The source has no patient id; each scan is treated as one patient.

Not recorded: race and ethnicity. The source has no race, ethnicity, ancestry or residence fields (Swiss hospital records generally do not hold them), so this dataset cannot show shifts or subgroup differences along those lines. The population is that of a Swiss university hospital system.

Cohorts

Cohort	Scans	AAA	Use
all	555	60	Every eligible, labelled scan
vendor_train	333	36	Stands in for the data a vendor trained on
site_retrospective	222	24	Stands in for a hospital's retrospective pull ("the last 6 to 9 months of abdominal CTs")

The two halves are a random split stratified on AAA and sex (not on institute), so they come from the same population: a fair check should not find a meaningful difference. In practice a hospital never sees the vendor's rows, only something like vendor_datasheet/. For sites that do differ (other scanners, thick slices, different labelling rules, repeat patients), see slicepoint-hospitals.

12 scans are not in the cohorts; tables/cases.parquet gives each reason (5 fail the eligibility rule, 6 have no measurable infrarenal diameter, 1 is a thorax CT). Scans the baseline model fails on (for example Dice 0) are kept: a real cohort contains them.

Intended use

A test harness for checking how an imaging AI model behaves at a new site, not a leaderboard. The ground-truth masks are independent of any one model, so you can train your own classifiers or segmenters (AAA, enlarged kidneys, organ volumes) on one part of the data, apply them to another, and see how shifts in scanners, slice thickness, phase, age or case mix, or hidden subgroups, change their results. The aim is to show the difference between a model's results on clean development data and on a site's own data, not to build the best model.

How the values were made

• Ground-truth masks: the segmentations distributed with TotalSegmentator v2.0.1, unchanged. The dataset authors made them with model help and expert review and correction; they are not output of the SlicePoint pipeline.
• Labels: max_aortic_diameter_mm is the maximum infrarenal outer-wall diameter, measured perpendicular to the centerline of the ground-truth aorta mask. aaa is diameter >= 30 mm; aaa_size_class is none, 30-54mm or >=55mm. These are measurement-derived labels, not radiologist reads.
• Predictions (baseline model): the output of an AI model, not a second ground truth. We ran the public TotalSegmentator model (v2.18.0, total task) on each CT and applied the same measurement and 30 mm rule to its aorta mask. It stands in for the kind of nnU-Net-based segmentation found in commercial imaging AI products. Null means the model gave no answer. _ref columns use the ground-truth masks, _pred columns the model's.
• The baseline model has seen most of these scans. TotalSegmentator was trained on this dataset: of the 567 scans, its own split (totalsegmentator_split in tables/cases.parquet) marks 491 as training, 21 as validation and 55 as test. So _pred results are not an independent test of the model. Here it scores about the same on its test scans as on its training scans (aorta Dice 0.968 vs 0.967; AAA agreement 98% vs 99%), but the test scans include only 4 AAA cases, too few to rule out a difference. For a fair baseline, use the test scans; if you train your own model on part of this data, compare it with TotalSegmentator on scans neither model was trained on.
• Features: scanner and acquisition fields come from the TotalSegmentator metadata (v3.0.0 adds slice thickness). Contrast phase is predicted. study_description and protocol_name are synthetic: the source has none, so they are drawn from a vocabulary of realistic orderables. HU and organ volumes come from the masks.
• Embeddings: 2048-dimensional image embeddings from the Merlin CT encoder.
• Images: unchanged voxels and affines; four free-text NIfTI header fields emptied.

Download

hf download slicepoint/slicepoint-aaa-ct --repo-type dataset --include "cohorts/*" "vendor_datasheet/*"   # tables only, ~20 MB
hf download slicepoint/slicepoint-aaa-ct --repo-type dataset --include "masks/aorta/*"                    # one structure
hf download slicepoint/slicepoint-aaa-ct --repo-type dataset                                              # everything, ~17 GB

from datasets import load_dataset
labels = load_dataset("slicepoint/slicepoint-aaa-ct", "labels", split="site_retrospective")

Test a site against a vendor datasheet (summary statistics only, no vendor rows). The two TotalSegmentator datasheets show why a datasheet's population matters: the abdominal site_retrospective cohort is covered by the all-regions datasheet but distributed differently (thinner slices there), and matches the abdominal one:

slicepoint datasheet check --data cohorts/site_retrospective --site-id site \
    --datasheet vendor_datasheet/totalsegmentator-v2-train-all.json --scope-broader
slicepoint datasheet check --data cohorts/site_retrospective --site-id site \
    --datasheet vendor_datasheet/totalsegmentator-v2-train-abdominal.json

Run SlicePoint on the demo pair:

slicepoint assess --data cohorts/site_retrospective --site-id site \
    --reference-dir cohorts/vendor_train --reference-site-id vendor --html report.html

Source, license and citation

Built from the TotalSegmentator dataset v2.0.1 (Wasserthal J. et al., University Hospital Basel, doi:10.5281/zenodo.10047292), with acquisition details from v3.0.0 (doi:10.5281/zenodo.22688904), licensed CC BY 4.0. This dataset is a modified version (subset, header fields emptied, derived tables added) by Asher Informatics PBC and is released under CC BY 4.0. Segmentations by the TotalSegmentator total model (Apache-2.0). Embeddings by Merlin (MIT).

Please cite TotalSegmentator:

@article{wasserthal2023totalsegmentator,
  title   = {TotalSegmentator: Robust Segmentation of 104 Anatomic Structures in CT Images},
  author  = {Wasserthal, Jakob and Breit, Hanns-Christian and Meyer, Manfred T. and others},
  journal = {Radiology: Artificial Intelligence},
  volume  = {5}, number = {5}, year = {2023},
  doi     = {10.1148/ryai.230024}
}

The scans came out of the 2026 Nucleate Pittsburgh hackathon, where SlicePoint started.