structuring-radiology-reports

# Structuring radiology reports A radiology report is prose, but its *meaning* is structured: a **technique**, a **comparison**, a list of **findings** (each with anatomy, laterality, and a measurement), and an **impression** that may carry an **assessment category** (BI-RADS, Lung-RADS) and a **follow-up recommendation**. This skill turns the narrative into that structure so findings are trackable — especially **incidental findings** that need downstream follow-up. OpenMed extracts the anatomy, disease/finding, and measurement spans on-device; this skill organizes them into sectioned, coded findings. It is **decision-support, not a diagnostic device** — every structured finding must be attributable back to its source sentence for radiologist review. ## When to use - You have a CT/MRI/X-ray/US/mammography report and need `{technique, comparison, findings[], impression}` with measurements and laterality. - You must capture **BI-RADS** (breast) or **Lung-RADS** (lung screening) assessment categories and the recommended action. - You need to **track incidental findings** and the follow-up interval/modality the report recommends. - You are mapping findings toward **RadLex** terms or a **DICOM-SR** structured report. ## Quick start ```python import openmed report = ( "TECHNIQUE: CT chest without contrast.\n" "COMPARISON: CT 2023-11-02.\n" "FINDINGS: A 8 mm solid nodule is noted in the right upper lobe, " "unchanged. No pleural effusion.\n" "IMPRESSION: 8 mm right upper lobe nodule, stable. Lung-RADS 2. " "Recommend annual low-dose CT screening." ) # 1) De-identify the report on-device first (synthetic example shown). deid = openmed.deidentify(report, policy="hipaa_safe_harbor") text = deid.deidentified_text # 2) Run NER for anatomy / finding / measurement spans. ents = openmed.analyze_text( text, model_name="anatomy_detection_superclinical", # Anatomy category output_format="dict", )["entities"] # 3) Split sections by header, then attach entities + measurements per finding. import re SECTION = re.compile(r"(?im)^(TECHNIQUE|COMPARISON|FINDINGS|IMPRESSION)\s*:") sections, last, name = {}, 0, None for m in SECTION.finditer(text): if name: sections[name] = text[last:m.start()].strip() name, last = m.group(1).upper(), m.end() if name: sections[name] = text[last:].strip() structured = { "technique": sections.get("TECHNIQUE"), "comparison": sections.get("COMPARISON"), "findings": _split_findings(sections.get("FINDINGS", "")), # one per sentence "impression": sections.get("IMPRESSION"), "measurements": re.findall(r"\b\d+(?:\.\d+)?\s?(?:mm|cm)\b", text), "laterality": sorted({w for w in ("right", "left", "bilateral") if re.search(rf"\b{w}\b", text, re.I)}), "assessment": (re.search(r"\b(?:BI-RADS|Lung-RADS)\s*\d[A-C]?\b", text, re.I) or [None])[0] if re.search(r"RADS", text, re.I) else None, "follow_up": _extract_followup(sections.get("IMPRESSION", "")), } ``` `_split_findings` / `_extract_followup` are your sentence splitter and a recommendation matcher ("recommend …", "follow-up in N months"); keep each finding tied to its source sentence offsets. ## Workflow 1. **De-identify first.** `openmed.deidentify(report, policy=...)`; structure from `deidentified_text`. Patient name, MRN, accession, and dates go before anything is stored or shared. 2. **Split sections** by the standard headers (TECHNIQUE, COMPARISON, FINDINGS, IMPRESSION; also HISTORY/INDICATION). Reports vary — fall back to position if headers are missing. 3. **Run `analyze_text`** for anatomy and finding entities; capture measurements ("8 mm", "1.2 cm") and laterality ("right", "left", "bilateral") near each finding. 4. **Build one structured finding per observation**: `{anatomy, finding, laterality, measurement, change_vs_prior, source_offsets}`. "Unchanged", "stable", "increased", "new" capture temporal change against the comparison. 5. **Pull the assessment category** (BI-RADS 0-6, Lung-RADS 1-4X) from the impression and the **recommended follow-up** (modality + interval). 6. **Flag incidental findings** — findings unrelated to the exam indication — and route them to a follow-up tracker so they aren't lost. 7. **Map toward RadLex / DICOM-SR** if you need coded interoperability, and surface the whole structure to a radiologist for verification. ## Hand-off to / from OpenMed OpenMed's `analyze_text` returns a `dict`; `result["entities"]` items carry `text`, `label`, `confidence`, `start`, `end`. - **From** `extracting-clinical-entities`: Anatomy and Disease/finding entities populate each structured finding; keep offsets so every field traces to a source sentence. - **From** `extracting-lab-tables` / OCR: if the report is a scan, OCR it first (`openmed.multimodal.ocr.ocr`), then run NER on the recognized text. - **From** `segmenting-clinical-sections`: reuse section detection if your reports don't use canonical headers. - **To** `building-patient-timelines`: dated findings + change-vs-prior feed a longitudinal view (e.g. nodule size over time). - **To** `extracting-dicom-metadata`: pair the structured findings with the study's DICOM metadata when assembling a DICOM-SR object. - **De-identify** with `deidentifying-clinical-text` (`openmed.deidentify`) before any export. Everything runs **on-device**. ## Edge cases & gotchas - **Negation and uncertainty change meaning.** "No pleural effusion" and "cannot exclude metastasis" are findings *about* absence/uncertainty — don't record them as positive findings. Use `resolving-clinical-context` (`openmed.clinical`) for negation/hedging before asserting a finding. - **Laterality errors are clinically dangerous.** "Right" vs "left" must bind to the correct finding; a misattributed side can drive wrong-site decisions. Tie laterality to the nearest anatomy span by offset, not document-wide. - **Measurements need their unit and axis.** "8 mm" vs "0.8 cm" are equal; a bare "8" is ambiguous. Capture the unit; for masses, capture all reported dimensions ("2.1 x 1.4 cm"), not just the first. - **Assessment categories have controlled value sets.** BI-RADS 0-6 and Lung-RADS 1, 2, 3, 4A, 4B, 4X each map to a defined management action — don't invent or round categories; pull the literal value from the impression. - **Incidental findings get lost.** A renal cyst mentioned in a chest CT is the classic missed follow-up. Explicitly separate incidental from indication-related findings and push incidentals to a tracker. - **The impression is the actionable summary**, but findings may contain detail the impression omits — structure both, and prefer the impression for follow-up/assessment. - **Decision-support disclaimer.** This is not a diagnostic medical device; it organizes text a radiologist authored. Every structured field must be reviewable against its source. Do not auto-act on a derived category or follow-up without clinician sign-off. ## Standards & references - RadLex (RSNA radiology lexicon): https://radlex.org/ - DICOM Structured Reporting (PS3.16 templates): https://www.dicomstandard.org/ - ACR BI-RADS Atlas: https://www.acr.org/Clinical-Resources/Reporting-and-Data-Systems/Bi-Rads - ACR Lung-RADS: https://www.acr.org/Clinical-Resources/Reporting-and-Data-Systems/Lung-Rads - RSNA Radiology Reporting templates: https://www.rsna.org/practice-tools/data-tools-and-standards/radreport-template-library - ACR Incidental Findings white papers: https://www.acr.org/Clinical-Resources/Incidental-Findings - HL7 FHIR R4 DiagnosticReport (imaging): https://hl7.org/fhir/R4/diagnosticreport.html