Datasets:

raidium
/

MetricEval-BodyCT

license: other
license_name: raidium-metriceval-dua-2.1
license_link: LICENSE
gated: true
language:
  - en
task_categories:
  - text-ranking
  - summarization
annotations_creators:
  - expert-generated
size_categories:
  - n<1K
tags:
  - radiology
  - medical
  - clinical
  - ct
  - computed-tomography
  - report-generation
  - evaluation
  - nlg-evaluation
  - meta-evaluation
  - llm-as-a-judge
  - human-feedback
  - preference-ranking
  - inter-annotator-agreement
  - benchmark
  - expert-annotation
pretty_name: MetricEval-BodyCT
configs:
  - config_name: studies
    data_files:
      - split: test
        path: data/studies.parquet
    default: true
  - config_name: annotations
    data_files:
      - split: test
        path: data/annotations.parquet
extra_gated_heading: Access requires agreeing to the Data Use Agreement
extra_gated_description: Requests are reviewed manually; please allow a few business days.
extra_gated_prompt: >
  This dataset is for ACADEMIC AND RESEARCH USE only, under the Data Use
  Agreement in the LICENSE

  file. It contains de-identified radiology report text derived from Segmed's
  de-identified data

  pilot. By requesting access you agree that you will:

    1. use it solely for academic or research purposes on radiology report generation and its
       evaluation;
    2. make no attempt to re-identify any individual, provider, institution or site;
    3. not redistribute the dataset, or any derivative containing the report text, in whole or part;
    4. not use it to train models for clinical deployment, and not commercially exploit it, without
       a separate written agreement;
    5. cite the accompanying publication in any resulting work.

  Research inside a commercial organisation is permitted; commercial
  exploitation of the dataset is

  not. Access is personal and non-transferable, and may be revoked at any time.
  The data is provided

  "as is", with no warranty. It is not a medical device and must not be used to
  inform patient care.
extra_gated_fields:
  Full name: text
  Institution / affiliation: text
  Country: country
  Intended research use: text
  I will not attempt re-identification: checkbox
  I will not redistribute the report text: checkbox
  I agree to the Data Use Agreement: checkbox
extra_gated_button_content: Request access

MetricEval-BodyCT

This repository is a body CT benchmark for evaluating radiology report-generation metrics against radiologists' judgment.

It covers 100 CT studies (50 chest and 50 abdomen/pelvis), with three candidate reports each. Every candidate was independently annotated by multiple board-certified radiologists. The reference reports are de-identified radiology reports from multiple US centers, provided by Segmed and redistributed under the Data Use Agreement in LICENSE. The candidate reports are synthetic perturbations of those references: deliberately injected errors, false clinical statements, hallucinations, etc.

1. Evaluation protocol

Write your metric's scores to a CSV (300 rows, one per candidate), where higher must mean better. Then run:

python3 eval.py --scores your_metric_scores.csv --out results_your_metric.json --label YourMetric

An example input file, example_scores.csv, is included in this repository:

study_id,candidate_label,score
abdomen_0108ee4fedcc,a,0.617761
abdomen_0108ee4fedcc,b,0.717391
abdomen_0108ee4fedcc,c,0.745223
...

An example of the output:

  human target               agreement             95% CI   gamma  pairs
  -------------------------- --------- ------------------ ------- ------
  rank_accuracy                  0.611     [0.553, 0.671]  +0.222    594
  rank_answers_indication        0.592     [0.530, 0.656]  +0.184    554
  n_errors                       0.616     [0.548, 0.681]  +0.231    536
  n_significant                  0.611     [0.541, 0.680]  +0.223    368

Results are also persisted as a JSON file.

The four targets are different entry points into the radiologists' judgment:

target	what it is	inter-expert ceiling
`rank_accuracy`	primary — their ranking of the whole report's fidelity to the reference	0.833 [0.782, 0.881]
`rank_answers_indication`	their ranking of how well the impression answers the clinical indication	0.852 [0.809, 0.890]
`n_errors`	their total error count	0.901 [0.870, 0.930]
`n_significant`	their count of clinically significant errors	0.761 [0.712, 0.806]

2. Error records

Each rater, for each candidate, logged every error they found and then ranked the three candidates. annotations.errors is a nested column holding the individual errors; an empty list means the rater reviewed that candidate and confirmed it error-free.

field	description
`error_id`	stable id, so a specific error can be cited
`category`	one of the 9 categories — see Dataset statistics for the taxonomy
`significance`	`significant` \| `insignificant`
`anchor_line`	1-based line of `candidate_<label>` the error sits on, or `None`
`note`	optional free text, on 82 / 1582 errors (median 10 characters)

Because the published report text is the numbered view the radiologist saw, anchor_line resolves to a line by direct indexing:

line = study_row[f"candidate_{ann_row['candidate_label']}"].split("\n")[error["anchor_line"] - 1]

3. Dataset statistics

100 studies — 50 chest_ct, 50 abdomen_ct, one per patient
300 candidate reports — 3 per study
600 independent expert reads — 2 radiologists × 300 candidates
1582 error records
- mean: 2.64, median: 3, min:0, max:10
- 460 (29.1%) clinically significant
- 1495 (94.5%) localized to a line
- 82 (5.2%) carrying a free-text note
- 62 of the 600 reads confirmed error-free

The error taxonomy and the distribution of categories:

category	definition	n
`hallucinated`	finding asserted that is absent from the reference	903 (57.1%)
`wrong_certainty`	hedging error; over- or under-call of confidence	229 (14.5%)
`missed`	omission of a reference finding	135 (8.5%)
`wrong_severity`	wrong severity or extent	87 (5.5%)
`wrong_location`	wrong site or laterality	57 (3.6%)
`wrong_comparison`	false, missing, or wrong-direction comparison to a prior study	56 (3.5%)
`wrong_measurement`	wrong numeric measurement or count	52 (3.3%)
`wrong_characterization`	wrong descriptor or characterization	37 (2.3%)
`other`	catch-all; see `note`	26 (1.6%)

License

The report text was de-identified before it reached Raidium; removed identifiers appear in the text as segmed_* placeholders (segmed_DATE, segmed_NAME, segmed_FACILITY and similar), so strip or mask them if your metric is sensitive to out-of-vocabulary tokens.

Academic and research use only, under the Data Use Agreement in LICENSE. Research inside a company is fine; commercially exploiting the dataset is not. No redistribution, and no clinical-deployment training without a separate agreement. The reference reports are provided by Segmed and redistributed by Raidium, which does not own them.

Citation

@misc{corbiere2026radmatch,
  title         = {RadMatch: Auditable Radiology Report Evaluation via Finding-Level Matching},
  author        = {Corbi\`ere, Charles and Machado, L\'eo and Charley, Aubin and Callard, Baptiste
                   and Manceron, Pierre and Dancette, Corentin},
  year          = {2026},
  eprint        = {2609.01470},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2609.01470}
}