Radiology Report Tools

What it does

Radiology Report Tools work on report text rather than pixels. The bare model ID writes an English narrative from MedRun structured findings: findings, impression and an EHR-ready layout, where a validator rejects any number or label that is not present in the findings. This is the same report service behind the opt-in draft report of every MedRun model.

Two more tools serve physicians who paste an existing report. The analysis tool turns free text into structured data, lists incidental findings with the follow-up a guideline recommends, checks the report for errors and suggests codes. The rewrite tool produces an impression from a findings section, a patient-friendly version or a translation, keeping measurements, laterality and terminology intact. A fourth tool scores a generated report against a reference for developers who validate report generators.

Intended use

Support for reporting work, not for image interpretation: a structured record of a free-text report for registries and audits, a second look for laterality slips and missed follow-ups, a plain-language version the physician can hand to the patient. Recommendations come from cited guidelines.

Who it is for

Radiologists and other physicians who write or receive reports, quality and audit teams in hospitals, coders, and developers who build reporting systems or evaluate report generation.

Accepted input

• Report text: plain text, DICOM SR, HL7 v2 ORU messages or a FHIR DiagnosticReport, up to about 20 pages.
• Structured findings: the MedRun result JSON of any model, for the narrative.
• DICOM header (optional, error check): patient sex, body part and laterality to compare with the text.
• Generated and reference reports: for the evaluation tool.

Requirements

Requirement	Why
One report per request	Codes, follow-up items and errors are tied to one examination
Machine-readable text, not a scanned image	Every extracted value keeps a link to its source text span
Rewrite mode (impression, patient-friendly or translation); a target language for translation	Selects which rewrite the run returns

The input language is detected automatically.

Optional context

The examination type and indication in clinical_context give the follow-up and error checks the context a report alone may leave out.

The result

Section	Content
Structured data	RadElement common data element values with source spans, plus RadLex and SNOMED CT codes
Incidental findings and follow-up	Incidental findings with the recommended test and interval and the guideline cited
Error check	Laterality errors, sex or organ mismatches, internal contradictions, inconsistent numbers and impression omissions
Coding suggestions	ICD-10-CM and CPT codes, each with its evidence span
Impression	Ranked, actionable impression with recommendations, written from the findings section
Patient-friendly version	Plain-language explanation at a target reading level that adds no new facts, with a readability score
Translation	The report in another language, with a check that measurements, laterality and terms survived

The first four sections are returned by the analysis tool, the last three by the rewrite tool according to its mode.

Standards

• Terminologies: RSNA RadElement common data elements, RSNA RadReport templates, RadLex, SNOMED CT, LOINC/RSNA Radiology Playbook, ICD-10-CM and CPT.
• Follow-up guidance: ACR white papers on incidental findings; Fleischner Society 2017 guideline for pulmonary nodules.
• Formats: FHIR DiagnosticReport (status preliminary for generated drafts) with Observation resources; DICOM SR text items; JSON with cde[], codes[], findings[] and recommendations[].

Related tools

• universal/report/narrative: findings and impression from MedRun structured findings, every number validated (the bare model ID).
• universal/report/analyze: structured data, incidental findings and follow-up, error check and coding in one run.
• universal/report/rewrite: impression, patient-friendly version or translation, chosen with mode.
• universal/report/evaluate: clinically significant error counts and similarity scores of a generated report against a reference, for developers.