A foundation model utilizing chest CT volumes and radiology reports for supervised-level zero-shot detection of abnormalities
Paper • 2403.17834 • Published • 5
Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains radiology report text derived from CT-RATE and MR-RATE, both released under CC-BY-NC-SA-4.0 with gated access. By requesting access you confirm you have accepted the terms of the source datasets and will use this data for non-commercial research only.
Log in or Sign Up to review the conditions and access this dataset content.
Error-injected radiology reports (CT-RATE / MR-RATE)
Reference reports paired with deliberately corrupted versions, where every injected error is labelled by construction. Built to develop and validate claim-level evaluation metrics for radiology report generation, where the standard lexical metrics fail badly: a single constant "all normal" string scores BLEU-1 48.3 / ROUGE-L 49.7 against real CT-RATE references, beating published state of the art, while asserting no finding at all.
Contents
| file | examples | injections | errors/example |
|---|---|---|---|
ct.jsonl |
2,000 | 3,395 | 1.7 |
mr.jsonl |
2,000 | 3,311 | 1.66 |
Each record is self-contained:
{
"source_id": "...", // volume / study id in the source dataset
"modality": "ct" | "mr",
"split": "clean" | "single" | "multi" | "stylistic",
"reference": "...", // the original report
"candidate": "...", // the corrupted version
"n_errors": 3,
"categories": ["negation_error", "..."],
"injections": [
{"category": "...", "category_id": 3, "significant": true,
"original": "...", "replacement": "...", "detail": "..."}
]
}
Splits
| split | meaning | correct behaviour |
|---|---|---|
clean |
no injection | metric must report zero errors |
single |
exactly one error | clean per-category attribution |
multi |
1-6 errors | the realistic case |
stylistic |
meaning-preserving paraphrase only | metric must report zero errors |
clean and stylistic are the false-positive controls: anything reported there is
a metric penalising a correct report.
Error taxonomy (12 categories)
- omission of finding · 2. fabrication of finding · 3. negation error ·
- laterality error · 5. location error · 6. severity error ·
- size / measurement error · 8. descriptive / morphology error ·
- modality-unobservable (claims something the study cannot physically show) ·
- out-of-field (anatomy outside the acquisition) · 11. comparison error ·
- stylistic variation (inverse test — must score zero penalty)
Categories 9 and 10 were added for multi-modality work and are not in the standard
ReXVal/GREEN taxonomy. Each error carries a significant flag: an omitted
normality statement is insignificant, an omitted finding is not.
Provenance
- generated
2026-08-25 20:07:11 - extractor:
llm - seed:
17 - source split: train
- code commit:
nogit
Source reports are drawn from the training splits of each dataset; validation and test are reserved for final evaluation.
Note the corpus is extractor-dependent at generation time — omission and negation injections choose targets from parsed claims, so a different extractor would corrupt different sentences. The labels are ground truth regardless, but the extractor used is recorded above.
Licence and attribution
CC-BY-NC-SA-4.0, inherited from the source datasets. Please cite:
- CT-RATE — Hamamci et al., Developing Generalist Foundation Models from a Multimodal Dataset for 3D Computed Tomography (arXiv:2403.17834)
- MR-RATE — Forithmus / MR-RATE (HuggingFace)
Research use only. Not for clinical use.
- Downloads last month
- 8