Radiologist-reviewed radiology report data
AI-drafted radiology reports and imaging AI findings, reviewed by radiologists against the images themselves. They mark what the model missed, what it invented, left and right swapped, wrong comparisons with prior studies and impressions the findings don’t support, then write the corrected sentence. Images are de-identified and reports are synthetic, so no patient can be identified.
Reviewed by certified medical professionals
Specification
- Studies
- X-ray, CT and MRI of the chest, head, abdomen and spine, with prior studies for comparison
- AI outputs
- Draft reports from report-generation models and findings from detection models
- Labels
- Missed and invented findings, laterality errors, wrong comparison with priors, wrong measurements and the clinical significance of each error
- Quality check
- Reviewed by radiologists whose credentials Avenzoar has verified; significant errors confirmed by a second radiologist
- Privacy
- De-identified images and synthetic reports: no names, dates or other patient identifiers
- Format
- DICOM or PNG images; reports and error labels as JSON or CSV, each error tied to its sentence
Example of the data
What you see, and the file you get.
| Missed | Small left pneumothorax | significant |
|---|---|---|
| Wrong side | Nodule in the right lower lobe | should be left |
| Prior | “Unchanged since prior” | grew 4 → 7 mm |
| study_id | modality | draft_sentence | error_type | severity | correction |
|---|---|---|---|---|---|
| st_0001 | chest_xray | No pneumothorax. | missed_finding | significant | Small left apical pneumothorax. |
| st_0002 | ct_chest | A 7 mm nodule in the right lower lobe. | laterality | significant | A 7 mm nodule in the left lower lobe. |
| st_0002 | ct_chest | Unchanged from the prior study. | wrong_comparison | significant | Grown from 4 mm to 7 mm since the prior study. |
| st_0003 | chest_xray | Small right pleural effusion. | invented_finding | minor | No pleural effusion. |
| st_0004 | mri_brain | No acute intracranial abnormality. | |||
| st_0005 | ct_abdomen | Impression: acute appendicitis. | unsupported_impression | significant | Impression: normal appendix, no acute abnormality. |
Sample rows. Also as JSON or Excel.
Contact us to get the dataset
Tell us what you need. We walk you through the data, its formats and how it is checked on a short call.
What it is used for
Report generation models
Train and fine-tune on drafts with radiologists’ corrections, sentence by sentence.
Hallucination and laterality checks
Measure how often a model invents findings or swaps left and right.
Imaging AI validation
Check detection models’ findings against radiologists’ reads before and after deployment.
Clinical benchmarks
Compare models and releases on the same studies, with every error graded by clinical significance.
Questions
Is there real patient data in it?
No patient identifiers. Images are de-identified and reports are synthetic or AI-drafted, so there are no names, dates or record numbers in the data.
Who reviews the reports?
Radiologists whose credentials Avenzoar has verified. They read the images, mark every error in the draft and grade how much it matters clinically.
Can I see samples first?
Yes. Request samples with the form on this page and we will send them with the full specification.
Tell us what your model needs.
Request samples and the full specification. We reply within a couple of business days.
Request samples