Data standard

Every dataset is one folder in datasets/. The folder name is the dataset id.

datasets/brats-2021/
  dataset.yaml   what the dataset is, where to get it, which license applies
  README.md      the description, in your own words
  stats.csv      numbers about the cohort, one per row

Licenses live in licenses/, one file per license, so that many datasets can share one breakdown. Allowed values for every field live in vocab/. The site, the checker (bun run validate) and the JSON Schemas in schema/ all read these same files.

dataset.yaml

Field	Required	Notes
id	yes	Lowercase letters, digits and dashes. Equals the folder name.
name	yes	Short name, at most 60 characters.
full_name		The long name, if there is one.
summary	yes	50 to 320 characters. Shown in search results and search engines.
homepage	yes	The official page of the dataset.
doi		Bare DOI, e.g. 10.7937/K9/TCIA.2015.LO9QL9SX.
year	yes	Year of the first public release. updated holds the year of the latest release.
version, status		status is active, superseded or retired.
creators	yes	Groups or people who made the dataset.
species		Defaults to human.
modalities	yes	vocab/modality.yaml (DICOM modality codes).
contrasts, tracers		vocab/contrast.yaml, vocab/tracer.yaml.
anatomy	yes	vocab/anatomy.yaml.
conditions, tasks, formats, vendors		vocab/condition.yaml, vocab/task.yaml, vocab/format.yaml, vocab/vendor.yaml.
countries		ISO 3166-1 alpha-2 codes.
field_strengths		Tesla, e.g. [1.5, 3].
size_gb		Download size.
access	yes	type from vocab/access.yaml and the url where access starts.
licenses	yes	One entry per license. With several licenses, applies_to says which part of the data each covers.
license_combine		all (default) when each license covers a different part of the data, any when the same data is offered under alternative licenses. The site combines answers with the most restrictive or the friendliest rule accordingly.
citation		text and/or bibtex, as the creators ask to be cited.
sources	yes	One entry per document that stats.csv rows point to. The key names the document, e.g. baid2021 or tcia-brats2021, never its kind like paper. kind is paper, website, data or computed.
related		Ids of related datasets in this index.
verified	yes	The date someone last checked the entry against its sources, and their GitHub handle in by.

Values in the vocabulary lists (contrasts, conditions, ...) are filled in automatically from stats.csv, so a value that appears there does not have to be repeated in dataset.yaml.

stats.csv

A table with six columns. Each row is one number.

measure,by,value,source,where,note
subjects,,1251,baid2021,Table 1,
subjects,sex=female,512,baid2021,Table 2,
subjects,contrast=FLAIR,1251,baid2021,Section 2.1,
subjects,sex=female;age=60-69,118,baid2021,Table 2,
age_mean,,58.9,baid2021,p. 4,

Column	Meaning
measure	What is counted or summarized, from vocab/measures.yaml: subjects, studies, scans, images, slides, or an age statistic such as age_mean.
by	The breakdown. Empty means the whole dataset. Otherwise dimension=value pairs joined with ;. Dimensions are listed in vocab/dimensions.yaml.
value	The number. Write ~600 when the source only gives an approximate number.
source	The key of one document under sources in dataset.yaml.
where	Where in that document the number is: Table 2, Figure 3, p. 4, Section 2.1. Empty only when the document is a short web page.
note	Optional. Quote the column in CSV if it contains a comma.

Dimensions

Dimension	Values	How several values combine in a filter
modality	vocab/modality.yaml	all (subject has every selected modality)
contrast	vocab/contrast.yaml	all
contrast_set	contrasts joined with + in alphabetical order, e.g. FLAIR+T1w+T2w	the exact set a subject has
tracer	vocab/tracer.yaml	all
anatomy, condition, vendor, view	their vocabulary files	any
sex	female, male, other, unknown	any
age	bins in completed years: 60-69, 90+, 0-0.5	any
field_strength	tesla, e.g. 3	any
country	ISO 3166-1 alpha-2	any
split	train, validation, test	any

sex, age, contrast_set, country and split are partitions: every subject has exactly one value, so their counts add up to at most the total. The other dimensions can overlap. A subject can have several contrasts or conditions.

Three levels of detail

The site answers questions like "datasets with at least 500 subjects that have both T1w and FLAIR and are aged 60 to 80". How exact the answer is depends on what stats.csv contains.

1. Totals per value. subjects,contrast=T1w,... and subjects,contrast=FLAIR,.... The site can only give a range for subjects with both, using the Fréchet bounds: at least T1w + FLAIR − total, at most the smaller of the two.
2. Contrast combinations. subjects,contrast_set=FLAIR+T1w,... for every combination that occurs. Answers about contrasts become exact. This is the most useful extra for machine learning and usually easy to get.
3. Cross tables. Rows with several dimensions, like sex=female;age=60-69. Answers for those combinations become exact.

Add whatever the sources report. Every row narrows the range shown to users.

Checks

bun run validate runs on every pull request. It checks that:

• every value is in its vocabulary, every source key exists and no row is repeated
• source keys name a document rather than a kind (paper, website are rejected) and no two keys share a url
• no count is larger than its total, partition counts do not add up to more than the total
• no cross-table row is larger than one of its single-dimension rows
• a contrast count is at least the sum of the contrast sets that contain it
• age statistics are in order (minimum ≤ median ≤ maximum)

Where numbers may come from

The index publishes aggregate numbers only, never data about individual subjects. A number may come from:

• the dataset's own documentation, website or data descriptor paper (kind: paper or website)
• metadata files that anyone can download under a public license without registering or signing anything, e.g. a participants.tsv under CC0, CC BY or CC BY-NC (kind: computed, with the file in url and the columns used in where)

A number may not be computed from data you received under a data use agreement, registration terms or any other contract. Most of those agreements forbid passing on information derived from the data, even counts.

Counts that you compute yourself for a combination of two or more dimensions are left out when they are below 10, because small cells can identify people. The checker enforces this for computed sources. Numbers copied from a publication stay as published.

licenses/*.yaml

Each license file answers the same questions, listed in vocab/license-rules.yaml. Every answer is yes, no, conditional (yes, with a condition stated in the note) or unspecified (the text is silent).

id: LicenseRef-Example-DUA # SPDX id, or LicenseRef-<name> for a custom agreement
name: Example Data Use Agreement
url: https://example.org/dua.pdf
version: 'v2, March 2024'
summary: Plain-language summary in one to three sentences.
purpose: research # any | research | noncommercial | noncommercial_research | health_research | approved_project
rules:
  commercial_use:
    value: 'no'
    quote: The data may be used for non-commercial research purposes only.
  model_training:
    value: unspecified
  # ... every rule from vocab/license-rules.yaml
commercial_license:
  available: 'yes'
  url: https://example.org/commercial
verified:
  date: '2026-01-31'
  by: your-github-handle

Rules of thumb:

• Quote the sentence an answer is based on. If a quote comes from somewhere else than the license text, add its source url.
• unspecified is a real answer. Do not turn silence into yes or no.
• A "no" for a duty (no signed agreement, no ethics approval) needs no quote. The evidence is that the text does not ask for it.
• Where the provider sells a commercial license, record it under commercial_license.

The site combines the answers of all licenses of a dataset: the most restrictive one per rule when every license applies to some part, the friendliest one when they are alternatives (license_combine: any). Rules map to GA4GH Data Use Ontology codes where one exists.

Datasets without published terms

Some datasets are described in a paper but have no license and no data use terms: they are shared on request at the data holder's discretion (access.type: on_request) or not shared at all (access.type: not_shared). They use licenses/LicenseRef-No-Published-Terms.yaml, where every rule is unspecified and the purpose is unspecified. Use it only after checking the paper's data availability statement, the dataset website and the repository for a license. A plain "all rights reserved" or a data availability sentence without terms also counts as no published terms.

Vocabularies

Every file in vocab/ lists terms with an id, a label and optional description, synonyms, parent (a broader term in the same file) and mappings to other vocabularies.

File	Based on
modality.yaml	DICOM Modality (0008,0060)
contrast.yaml	BIDS suffixes for MR, local terms for CT phases, stains and ultrasound modes
anatomy.yaml	UBERON
condition.yaml	MONDO, HPO for findings
sex.yaml	BIDS participants.tsv levels
vendor.yaml	normalized names, raw DICOM Manufacturer strings as synonyms

SNOMED CT and ICD-O terms are not embedded, because their licenses restrict redistribution.

To add a term, prefer the upstream code: if BIDS, MONDO or UBERON has the concept, use it and add the mapping. Look up ids with bun run lookup "glioma" mondo. Term ids are never deleted. A replaced term keeps its id and gets a description that points to the new one.