Draft. Review. Settle.
An ordinary report drafts in about a minute and reviews in a few.
- 1
Draft
Local models read each new report and draft its answers. The draft is where a review starts, never an answer by itself.
- 2
Review
Tap a line of the report and its answers come up. Correct, next. Or fix what's wrong and save.
- 3
Settle
One radiologist per report, or a blind panel of two to five. Agreement settles it; otherwise an adjudicator decides.
- 4
Export
A finished review becomes a new version of the expected answers, with who changed what. Export the set as files, or every review as JSON lines.
What a reviewer checks.
Each report is split into units: a sentence of the findings, a line of the impression. Every unit has its answers.
- UnitsWhat each part of the report says: a positive finding, a negative one, a recommendation, or none.
- FindingsThe words quoted from the report, with side, size, change since the last study, and whether it's a key finding.
- RecommendationsThe action, the study, the interval and any condition: CT chest in 6–12 months, if high risk.
- Follow-upFor the whole report: recommended, conditional, not recommended or not stated.
One reader, or a blind panel.
Set per report set.
- One radiologist
- Whoever opens a report holds it for ten minutes. Anyone can review it again on purpose; the earlier review is kept and both names are credited.
- Blind panel
- Two to five reviewers each work alone and never see the others' answers. When they all agree, or most of three or more do, it's settled.
- Adjudicator
- When they don't, an adjudicator sees each reviewer's changes and notes and picks the final answers in one click.
- Calibration set
- Reviewers start from blank answers and are scored against the settled ones, which they never change. A way to see how a new reader compares.
Numerous guidelines, versioned.
The working references the models draft and write against. Read and edit them; every save is a numbered version with who, when and why, and any version can be restored.
- Fleischner
- Lung-RADS
- BI-RADS
- TI-RADS
- LI-RADS
- PI-RADS
- O-RADS
- Bosniak
- Deauville
- AAA surveillance
- Adrenal incidentaloma
- Pancreatic cyst
- Thyroid nodule
See how well reviewers agree.
On the Agreement page, administrators and adjudicators see each set at a glance. Kappa corrects for the agreement that chance alone would give: 0 is chance, 1 is perfect.
Across the panel
Fleiss' kappa over unit labels and over each report's follow-up.
Pair by pair
For every two reviewers: how often their reports came out identical, Cohen's kappa, and how far their findings and recommendations overlap.
Time per report
Measured by the server from opening to saving, and carried in every export.
Who it's for.
Radiologists
Review at a desk or on a phone, the report above and the answers below. Your corrections become the reference.
Researchers
Report sets with provenance, versions and each reviewer's changes, ready to score a report reader against.
Documentation teams
Building something that reads radiology reports? See what radiologists say it should find, report by report.
Local models only.
Every draft and every written report comes from a model running on RadReportX's own machine.
- Private by design: report text never goes to a cloud AI service.
- Every report keeps its provenance: who added it, when, and from which file.
- Drafting and writing run on local models.
Review your first report.
RadReportX is in a small beta. Enter your invite code to open it, then sign in with the account we send you.