annotations.parquet (the radiologist labels) is licensed CC-BY-4.0.
studies.parquet contains de-identified radiology report text derived from Segmed's
de-identified data pilot. It is restricted. By requesting access you agree that you will:
1. use it solely for non-commercial research on radiology report generation and its evaluation;
2. make no attempt to re-identify any individual, provider, institution, or site;
3. not redistribute the report text, or any derivative containing it, in whole or in part;
4. not use it to train or fine-tune models for clinical deployment without a separate agreement;
5. cite this dataset in any resulting publication.

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.

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 output

  endpoint                      tau_b             95% CI  blocks
  -------------------------- -------- ------------------ -------
  n_errors                     -0.361   [-0.432, -0.291]      97
  n_significant                -0.213   [-0.288, -0.142]      72
  rank_accuracy                -0.310   [-0.371, -0.252]      99
  rank_answers_indication      -0.209   [-0.275, -0.145]      94
  cat_hallucinated             -0.212   [-0.281, -0.142]      97
  cat_missed                   -0.172   [-0.308, -0.038]      33
  cat_other                    -0.075   [-0.387, +0.245]       6
  cat_wrong_certainty          -0.051   [-0.148, +0.050]      40
  cat_wrong_characterization   -0.270   [-0.453, -0.095]      13
  cat_wrong_comparison         -0.312   [-0.540, +0.069]       8
  cat_wrong_location           -0.369   [-0.494, -0.231]      20
  cat_wrong_measurement        -0.235   [-0.401, -0.054]      18
  cat_wrong_severity           -0.317   [-0.435, -0.197]      29

Results are also persisted as a JSON file.

Each endpoint is a different entry point into the radiologists' judgment:

endpoint	what it is	inter-rater ceiling (95% CI)
`n_errors`	primary — their total error count	-0.694 [-0.751, -0.634]
`n_significant`	their count of clinically significant errors	-0.471 [-0.564, -0.383]
`rank_accuracy`	their ranking of the whole report's fidelity to the reference	-0.645 [-0.746, -0.538]
`rank_answers_indication`	their ranking of how well the impression answers the clinical indication	-0.701 [-0.774, -0.618]
`cat_hallucinated`	findings asserted by the candidate that the reference does not support	-0.821 [-0.867, -0.773]
`cat_missed`	findings present in the reference and absent from the candidate	-0.690 [-0.820, -0.532]
`cat_wrong_location`	a real finding placed at the wrong anatomic site	-0.890 [-0.982, -0.786]
`cat_wrong_severity`	a real finding graded more or less severe than the reference	-0.879 [-0.975, -0.765]
`cat_wrong_characterization`	a real finding described with the wrong morphology	-0.922 [-1.000, -0.769]
`cat_wrong_measurement`	a size, count or distance that disagrees with the reference	-0.874 [-0.978, -0.700]
`cat_wrong_certainty`	hedging that over- or under-states how definite the reference is	-0.658 [-0.792, -0.517]
`cat_wrong_comparison`	a change from a prior study invented, dropped or reversed	-1.000 [-1.000, -1.000]
`cat_other`	an error the rater logged outside the categories above	too few blocks

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%)

Citation

@misc{radmatch2026,
  title        = {RadMatch: Auditable Radiology Report Evaluation via Finding-Level Matching},
  author       = {Corbi\`ere, Charles and Machado, L\'eo and Charley, Aubin and Caillard, Baptiste
                  and Le Floch, Korentin and Manceron, Pierre and Dancette, Corentin},
  year         = {2026},
}

Downloads last month: 58

Total file size:

328 kB