README.md

---
license:
  - cc-by-sa-4.0
  - cc-by-nc-4.0
  - cc-by-3.0
  - cc-by-4.0
pretty_name: FLIP mock trust data
tags:
  - medical-imaging
  - omop-cdm
  - dicom
  - synthetic
  - federated-learning
---

# FLIP mock trust data

Mock data for the dev and test trusts of [FLIP](https://github.com/londonaicentre/FLIP), the
Federated Learning Interoperability Platform. Everything here is either **synthetic** or **derived
from a public research dataset** under its licence — no patient data. Licensing is **per project**
(see the table and the per-project notes): the spleen- and cxr-derived content is **CC BY-SA 4.0**,
the prostate-derived content (`prostate_project`) is **CC BY-NC 4.0** and therefore non-commercial;
the pathology tables (`pathology_project`) are metadata derived from IDC's TCGA-BRCA collection
(**CC BY 3.0**) and its Pan-Cancer-Nuclei-Seg annotations (**CC BY 4.0**); the DICOM vocabulary bundle
keeps its own Apache 2.0.

## Versions

**There is exactly one copy of every artefact, on `main`, at the paths below. A data version is a
git tag on this repository** (`20260729`, `20260901`, …), and FLIP pins one in
`trust/.data_version`. Every consumer fetches
`https://huggingface.co/datasets/aicentreflip/trust-data/resolve/<version>/<path>`, so an old
version stays reachable at its tag forever and is never duplicated as a second directory or a
suffixed filename. Publishing a new version is one commit on `main` (the changed files) plus one tag:
`FLIP/trust/publish_trust_data.py`.

| tag | what changed |
| --- | --- |
| `20260729` | two-project (spleen, cxr) two-trust cut: OMOP tables, DICOM sets, vocab-free pgdata, Orthanc storage |
| `20260901` | `spleen_project` measurement rows carry the millimetre unit for SliceThickness (FLIP#1098); pgdata rebuilt |
| `20260902` | `prostate_project` added — PI-CAI fold 0, the first cohort published for the seed pipeline alone (no snapshot) |
| `20260903` | `prostate_project` DICOM re-cut with synthetic patient identities (name, birth date, referring physician, study description) — tables unchanged |
| `20260911` | `pathology_project` added — 24 IDC TCGA-BRCA whole-slide images described in OMOP; **tables and the slide manifest only**, the imaging stays on IDC (no `dicom/` set) |
| `20260917` | `brain_mri_project` added — 40 MSD Task01 brain tumour cases as four-series MR studies, **tables and metadata table only**; `spleen_project` re-cut on the same deterministic converter (every identity, accession and UID changed) and its `dicom/spleen_project.tar.gz` removed from `main` — both DICOM sets now regenerate locally from the public MSD archives (FLIP#1221) |
| `20261014` | `spleen_project` metadata table re-cut: the converter now stores every slice in DICOM display order (`ImageOrientationPatient` `[1,0,0,0,1,0]`, `ImagePositionPatient` moved with it), so the regenerated CT no longer renders rotated 180° in 2-D viewers (FLIP#1495). OMOP tables, identities and UIDs unchanged; `brain_mri_project`'s regenerated DICOMs are reoriented the same way, but nothing published for it changes |

The pre-tag layout — files with a version in their name or path (`trust<N>_pgdata_<v>.tar`,
`trust<N>_orthanc_data_<v>.tar`, `omop-csv/<v>/`, `dicom/<v>/`) — was removed from `main` on
2026-09-11, once FLIP v0.6.0 made every consumer fetch by tag. Those files remain reachable at the
tags whose trees carried them (`20260729` … `20260911`); nothing new is ever added at a versioned path.

## What is here

| path | contents | licence | source |
| --- | --- | --- | --- |
| `omop-csv/<project>/*.csv` | OMOP CDM 5.4 tables for each mock cohort, one canonical dataset; every row carries `source_trust` | per project | generated by the converters in `FLIP/fl-tutorials/datasets/` |
| `omop-csv/<project>/source/…` | the inputs each OMOP export was built from (DICOM metadata table; PI-CAI's marksheet; the pathology slide manifest) | per project | same |
| `dicom/<project>.tar.gz` | the DICOM instances behind each cohort, `<accession>/*.dcm`, one archive per project — only where the imaging has no public source to regenerate from (`cxr_project`, `prostate_project`). `pathology_project` has none (its slides are fetched from IDC); `spleen_project` and `brain_mri_project` have none from `20260917` on (regenerated locally from the MSD archives by FLIP's deterministic converter) | per project | see per-project notes below |
| `trust<N>/trust<N>_pgdata.tar` | vocab-free PostgreSQL data volumes, two-trust snapshot (spleen + cxr) — the pre-`20260917` spleen cut | CC BY-SA 4.0 | built from `omop-csv/` |
| `trust<N>/trust<N>_orthanc_data.tar` | Orthanc storage volumes, two-trust snapshot (spleen + cxr) — the pre-`20260917` spleen cut | CC BY-SA 4.0 | built from the DICOM sets |
| `omop-vocab/vocab_dicom_*.zip` | the DICOM vocabulary for OMOP | Apache 2.0 | [DICOM2OMOP](https://github.com/paulnagy/DICOM2OMOP) |

No licensed OHDSI vocabularies (SNOMED CT, LOINC, …) are on this dataset; FLIP loads those at seed
time from a source each deployment licenses itself.

## Per-project provenance and licence

### `spleen_project`

CT volumes from the **Medical Segmentation Decathlon, Task09 Spleen**
(Memorial Sloan Kettering Cancer Center; Antonelli et al., *Nature Communications* 2022,
[medicaldecathlon.com](http://medicaldecathlon.com/)), converted from NIfTI to DICOM with
**synthetic patient identities** — names, NHS-style numbers, dates, sex, accession numbers and UIDs
are all generated; none belongs to a real person. The pixel data is MSD's.

**Tables and the metadata table only, from `20260917`.** MSD is open data, so the DICOM set is not
re-hosted: FLIP regenerates it from the Task09 archive with a deterministic converter
(`fl-tutorials/datasets/spleen/`, on `datasets/utils/dicom_writer.py` — every UID, date and identity a
pure function of the case id, so the regenerated tree is byte-identical anywhere and matches these
tables exactly; `make -C fl-tutorials verify-spleen-dicom` checks that both ways) and seeds a trust
from the local tree. That cut replaced the plastimatch-era converter, so every spleen identity,
accession and UID changed at `20260917`; the earlier set (41 studies, 3,650 instances) stays at the
tags up to `20260911` as `dicom/spleen_project.tar.gz`, and the two-trust volume snapshots still
carry it.

**Licence: CC BY-SA 4.0**, as MSD Task09 is (`dataset.json`: `"licence": "CC-BY-SA 4.0"`).
Attribution to the Medical Segmentation Decathlon and MSKCC is required by anything that reuses the
spleen-derived content here, and derivatives must carry the same licence.

### `brain_mri_project`

**Tables and the metadata table only — no imaging.** Forty cases from the **Medical Segmentation
Decathlon, Task01 Brain Tumour** (BraTS 2016/17; Antonelli et al. 2022) — the first 40 training
cases in natural order — each described as one MR study of four series (FLAIR, T1w, T1Gd, T2w) with
**synthetic patient identities** as the other projects carry, split 20/20 across two trusts
(`source_trust` round-robin over the cases). The DICOM set is regenerated locally from the public
Task01 archive by the same deterministic converter (`fl-tutorials/datasets/brain_mri/`), so what is
published is the small thing that makes a run reproducible: `source/dicom_metadata.csv`, one row per
series, naming every case, identity, accession and UID the regenerated tree must contain, and the six
OMOP tables derived from it (`person`, `visit_occurrence`, `procedure_occurrence` — LOINC 24587-8
*MR Brain WO and W contrast IV* — `image_occurrence` per series with SNOMED *Brain structure*,
`image_feature` and `measurement` for the DICOM attributes). The tumour labels are not in OMOP.

**Licence: CC BY-SA 4.0**, as MSD Task01 is. Attribution to the Medical Segmentation Decathlon and
the BraTS challenge is required by anything that reuses the brain-derived content here, and
derivatives must carry the same licence.

### `cxr_project`

Chest radiographs produced by a **generative model** trained by the London AI Centre for Value Based
Healthcare, with a synthetic radiology report per image (the `conditioning` / `pathologies` columns of
its metadata table) and synthetic patient identities as above. No real radiograph is reproduced.

**Licence: CC BY-SA 4.0.**

### `prostate_project`

Biparametric prostate MRI (axial T2W, ADC, high-b-value DWI — three series per study) from
**fold 0 of the PI-CAI Public Training and Development Dataset** (Saha, Twilt, Bosma, van Ginneken
et al.; [Zenodo record 6624726](https://zenodo.org/records/6624726)): 300 studies of 295 patients
from three Dutch centers (RUMC, ZGT, PCNN), converted from the released `.mha` to DICOM with
deterministic UIDs, PI-CAI's own anonymised identifiers (`PatientID` = PI-CAI `patient_id`,
`AccessionNumber` = `<patient_id>_<study_id>`) and, from tag `20260903`, **synthetic patient
identities** as the other projects carry — a name, a birth date consistent with PI-CAI's recorded
age, a referring physician and a study description, all generated deterministically from the
PI-CAI ids; none belongs to a real person. The clinical marksheet (PSA, PSA density, prostate
volume, ISUP grade group, csPCa, PI-RADS) is carried into OMOP as measurements and observations.
`source_trust` is one contributing center per trust: ZGT → 1 (76 studies), PCNN → 2 (69), RUMC → 3
(155; FLIP's two-trust dev stack loads sources 1 and 2, and RUMC waits for a third trust).

The segmentation masks that pair with these studies — whole gland (Bosma22b) and zonal PZ/TZ
(HeviAI23) from [`picai_labels`](https://github.com/DIAGNijmegen/picai_labels) — are not on this
dataset; FLIP's tutorial downloads them from that repository and uploads them into a project's
XNAT as data enrichment.

**Licence: CC BY-NC 4.0** — PI-CAI's images and labels are both released under
[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/), so unlike the rest of this dataset
the prostate-derived content is **non-commercial**. Reuse must credit the PI-CAI challenge (see
Citation) and may not be commercial; anyone using FLIP's mock data commercially must leave
`prostate_project` out.

### `pathology_project`

**Tables and a manifest only — no imaging.** Twenty-four haematoxylin-and-eosin whole-slide images
from the **TCGA-BRCA** collection of the [NCI Imaging Data Commons](https://portal.imaging.datacommons.cancer.gov/)
(DICOM Slide Microscopy), one slide per patient, twelve from each of two TCGA tissue source sites
(A8, A7 → `source_trust` 1 and 2). The slides and their Pan-Cancer-Nuclei-Seg nuclei annotations
(DICOM Microscopy Bulk Simple Annotations) are **not re-hosted here**: FLIP's tutorial downloads
them from IDC's public buckets at run time (`make -C fl-tutorials download-idc-pathology-data`), and
the annotations reach a trust's XNAT by data enrichment. What is published is the small thing that
makes a run reproducible — `source/manifest.csv`, the lockfile naming every selected slide (its UIDs,
site and the IDC index version the selection was resolved against, since IDC issues versioned
releases and series come and go) — and the four OMOP tables derived from it. The OMOP rows are
metadata: TCGA barcodes as `person_source_value`/`accession_id`, study and series UIDs, IDC's
recorded study dates, and the Slide microscopy modality concept. TCGA pathology DICOM is
de-identified, so demographics are OMOP's "No matching concept" (0) and `year_of_birth` is the
deliberately implausible sentinel 1900 rather than an invented value.

**Licence: CC BY 3.0 for the TCGA-BRCA-derived metadata and CC BY 4.0 for the annotation-derived
columns**, as IDC records for the collection and the `pan_cancer_nuclei_seg_dicom` analysis result
respectively (the `license_short_name` column of the IDC index). Attribution to TCGA / the NCI
Imaging Data Commons and, for the annotations, to Hou et al. (see Citation) is required.

## Provenance

The code that produced every table here is in the FLIP repository under
`fl-tutorials/datasets/`, and each project's OMOP export is verified to reproduce byte-for-byte from
its published inputs before it is tagged (`make -C fl-tutorials reproduce-<project>-omop`, the gate
`fl-tutorials/datasets/utils/verify_omop_tables.py`; the run output is recorded in the FLIP pull
request that published the tag). The DICOM sets are the original converter outputs, checked against
the published tables before publishing (`trust/orthanc/publish_dicom.py`): every accession number,
study UID and patient ID present, both ways.

## Citation

If you use the spleen-derived content, cite the Medical Segmentation Decathlon:

> Antonelli, M., Reinke, A., Bakas, S. et al. The Medical Segmentation Decathlon.
> *Nat Commun* 13, 4128 (2022). https://doi.org/10.1038/s41467-022-30695-9

If you use the prostate-derived content, cite the PI-CAI challenge and its public dataset:

> Saha, A., Twilt, J. J., Bosma, J. S., van Ginneken, B. et al. The PI-CAI Challenge: Public Training
> and Development Dataset (2022). Zenodo. https://doi.org/10.5281/zenodo.6624726

If you use the pathology-derived content, cite the NCI Imaging Data Commons and the nuclei annotations:

> Fedorov, A., Longabaugh, W. J. R., Pot, D. et al. NCI Imaging Data Commons. *Cancer Research* 81(16),
> 4188–4193 (2021). https://doi.org/10.1158/0008-5472.CAN-21-0950

> Hou, L., Gupta, R., Van Arnam, J. S. et al. Dataset of segmented nuclei in hematoxylin and eosin
> stained histopathology images of ten cancer types. *Sci Data* 7, 185 (2020).
> https://doi.org/10.1038/s41597-020-0528-1