Extracting Findings and Impressions From a Radiology Report

A radiology report is not prose with a summary at the end. It is two different kinds of claim in two labelled sections, and the difference between “there is a 9 mm nodule in the right upper lobe” and “findings are most consistent with a benign process” is the entire information content of the document.

The sections and what each asserts

Reports vary by institution but the canonical section set is stable enough to key on:

• Exam / Procedure — what was performed, with modality, body part and contrast status. Contrast status is a separate fact and is frequently the thing a downstream query filters on.
• Clinical history or Indication — why the study was ordered. This is a statement about the referrer’s question, not a diagnosis, and extracting it as one is a common and consequential error: “indication: rule out pulmonary embolism” does not mean the patient has one.
• Technique — how the study was acquired. Rarely extracted, occasionally decisive, because a limited or non-contrast technique bounds what the study could have shown.
• Comparison — which prior studies were reviewed. The dates here are the reference points for every change statement in the body.
• Findings — observations, organ system by organ system, including everything normal. This section is descriptive.
• Impression — the radiologist’s interpretation, usually numbered, usually short. This section is conclusive, and it is what the ordering clinician reads.

The first practical task is section detection, and it is less trivial than it sounds. Headings appear in upper case, in title case, with and without colons, sometimes as IMPRESSION: inline at the start of a paragraph rather than on its own line. Some reports have no Findings heading at all and run straight into an impression. Detect by heading where a heading exists, record which heading text matched, and mark the section as inferred where it did not — because a downstream consumer treating an inferred impression as a labelled one is inheriting your guess without knowing it.

What merging them destroys

Ask a model to “extract the findings from this report” and it will return a merged summary. Three specific things are lost, and none of them are recoverable afterwards.

The distinction between observed and concluded. A finding is something visible in the images. An impression is an inference from findings plus history plus prior studies. A dataset that cannot tell them apart cannot answer “what did the imaging show” separately from “what did the radiologist think”, and those are different questions with different reliability.

The normal statements. Findings sections state that things are normal, at length. Summarisation discards them as uninformative, but a documented negative — “no pneumothorax” — is a positive assertion that the structure was examined and was normal, which is materially different from the structure not being mentioned. The absence of a comment about an organ can mean it was outside the field of view.

The numbering. Impressions are numbered because the numbers are referenced — in follow-up recommendations, in tumour board discussion, in the next report’s comparison. Merging into a paragraph removes the addressability. Keep the impression as an ordered array with the original numbers preserved as printed, including the case where they are lettered or bulleted instead.

The schema follows: sections as a keyed object, findings as organ-system-tagged statements, impression as an ordered list of items, each item with its verbatim text and its own extracted attributes — the case for that shape over a flattened one is nested and flat extraction schemas. This is a structure-preserving extraction rather than a summarisation, and if you do want a summary as well, produce it as a second, clearly derived field — the reasoning for keeping the two separate is the same as in extract, then reason.

Certainty is part of the finding

Radiological language is deliberately graded, and the grade is information. “Consistent with”, “compatible with”, “suspicious for”, “cannot exclude”, “probable”, “possible” and “no evidence of” sit at different points on a scale that the radiologist chose on purpose. An extraction that outputs a finding as a bare label has flattened a chosen hedge into an assertion.

So each extracted item wants a certainty attribute alongside the finding, populated from the phrase actually used and with the phrase retained. Do not attempt to map hedges onto numeric probabilities: there is no published mapping that holds across radiologists, and inventing one converts a qualitative judgement into a false quantity. An enumerated ordinal with the verbatim phrase attached is honest; “0.7 probability” is not.

Change statements need the same treatment. “Increased in size compared to the prior study” is a relation between two studies, and it is meaningless without the comparison date from the Comparison section. Extract change as its own typed field — increased, decreased, stable, new, resolved — with a pointer to the comparison study, rather than as an adjective inside the finding text.

The same negation machinery that the diagnosis code page describes applies here with more force, because a radiology findings section has a higher density of negated statements than almost any other clinical text. A pipeline that gets negation wrong will report the exact opposite of the report’s content, fluently.

Incidental findings live in one section only

An incidental finding is something noticed that was not what the study was looking for. It appears in Findings by definition, and it may or may not be carried into the Impression. This asymmetry is the strongest practical argument for extracting the two sections separately.

An organisation running a follow-up or care-coordination programme typically wants exactly these: the observation in the body of a report that carries a recommendation for further imaging. If your extraction only keeps the impression, you have systematically discarded the population you were trying to find, and the failure is invisible because the impressions you did extract are all correct.

Recommendations deserve their own extracted object regardless of which section they appear in: the recommended action, the interval if one is stated, and the section it came from. They are frequently phrased conditionally (“if clinically warranted”), and the condition is part of the recommendation, not decoration on it. What an organisation does with a recommendation is a clinical and operational matter well outside the scope of extraction; getting it out of the document intact is the part this page is about.

Where a controlled vocabulary exists

Some report types carry a structured assessment category, and where one exists it is far more extractable than the prose around it. The American College of Radiology publishes reporting and data systems — BI-RADS for breast imaging, and equivalents for lung, liver and other domains — whose assessment categories are a fixed, numbered scale printed explicitly in the report.

A BI-RADS assessment, for instance, is an integer category with defined meanings running from an incomplete assessment through negative, benign, probably benign, suspicious, highly suggestive of malignancy and known biopsy-proven malignancy, with the suspicious category subdivided. Extract the category as an integer plus its subdivision letter where present, plus the verbatim string, plus the edition if the report states one — the categories are defined by an edition of the published atlas and the definitions have been revised between editions.

Where no controlled vocabulary exists — which is most reports — resist building one by asking a model to classify findings into categories you invented. The verbatim text with structure preserved is a better asset than a lossy classification into a taxonomy nobody else uses, and it can be classified later when you know what the question is.