Radiology Report Extraction: Exam, Dates and Impression

Radiology report extraction pulls the facts that identify a study and its report: modality, body part, laterality, exam and signature dates, comparison studies, any stated assessment category and the report status. The findings and impression stay as verbatim text. Extraction records what the radiologist wrote and leaves every reading of the images to clinicians.

Radiology reports sit alongside the other chart documents covered in clinical document extraction by document type. This post goes section by section through one imaging report, showing which parts become fields and which stay as text.

Radiology report exam fields: modality, body part and laterality

Radiology report exam fields identify exactly which study the report describes. The core set is the modality (such as CT, MRI, ultrasound or X-ray), the body part, the laterality, whether contrast was used, and the exam name as printed. Together with the exam date, these fields separate one study from every other study in the chart.

The American College of Radiology (ACR) practice parameter for communication of diagnostic imaging findings, revised in 2025 (Resolution 9), suggests a report format that starts with demographics, including the "name or type of examination" and the "date of the examination". You can read it in the ACR Practice Parameter for Communication of Diagnostic Imaging Findings. The parameter suggests a format rather than fixing one, so exam names vary from one facility to the next.

Keep the exam name as printed in one field, then split out its parts in others:

• Exam name as printed: "MRI KNEE RIGHT WO CONTRAST" (example).
• Modality: MRI.
• Body part: knee.
• Laterality: right.
• Contrast: without.

Laterality causes more trouble than any other exam field. It can sit in the exam name, in the technique paragraph, or only in the findings. When the exam name and the body of the report give different sides, do not pick one. Keep both, with their sources, and send the report to a person.

"Bilateral" is a laterality value, not a missing one. So is "left and right" in a two-part exam. A blank laterality should mean that the report states no side, which is normal for many exams, such as a chest X-ray.

An accession number, when printed, is the best single key for a study. It is assigned by the imaging department and stays the same across the preliminary report, the final report and any addendum.

Which dates does a radiology report carry?

A radiology report can carry five or more dates: the exam date, the dictation date, the transcription date, the date the report was finalized and signed, and the dates of any comparison studies. The exam date is usually the one that dates the report as a clinical event. The others record when the paperwork happened.

The ACR practice parameter lists the date of the examination, the time when relevant, and, when applicable, the "date and time of dictation", the "date and time of transcription" and the "date and time report is finalized". A printed report may show all of them, labelled in different ways: "Exam date", "Dictated", "Transcribed", "Signed", "Final".

Extract each date into its own labelled field. A single "report date" field forces a choice that the page does not make for you. A scan done late on one day and signed after midnight carries two calendar dates, and both are correct for what they describe.

Comparison dates need special care. A line such as "Comparison: MRI right knee 2024-08-19" names an earlier study. That date belongs in a comparison field linked to the earlier study's description. It must never be read as the date of the current exam. Comparison dates look just like exam dates and often sit near the top of the report body.

Choosing which date to use as the encounter date for a whole page, when several dates compete, is covered in the guide to which date on a medical record page is the encounter date.

Findings and impression in radiology report extraction

The findings and impression sections of a radiology report hold the radiologist's interpretation, and extraction should keep both as verbatim text. The findings describe what the radiologist saw. The impression states the conclusion. Splitting either into fields means deciding what a sentence means clinically, which is interpretation, not data capture.

The ACR practice parameter says that, unless the report is brief, "each report should contain an 'impression' or 'conclusion.'" It adds that a specific diagnosis should be given when possible, a differential diagnosis rendered when appropriate, and follow-up studies suggested when appropriate. The parameter also notes that interpretations "may appear anywhere in a radiology report", which is why the findings are interpretive text too.

That mix of conclusions, alternatives and hedged suggestions is exactly what breaks field extraction. "Findings may represent a sprain; partial tear not excluded. Consider follow-up MRI in 6 weeks" contains a possibility, a negation of an exclusion and a suggestion. Any field such as "diagnosis: sprain" or "tear: no" would misstate it.

What you can safely capture from these sections:

• The full findings text, verbatim, as one field.
• The full impression text, verbatim, as one field, keeping its numbering if it has any.
• The impression item count, when items are numbered, which helps a reviewer check nothing was dropped.
• Stated communication records, such as "Discussed with Dr. Grant at 14:05", for which the ACR parameter gives the date and time, the method and the name of the person reached as an example of documentation.

Keep line breaks and numbering in the verbatim fields. A numbered impression list that runs together into one paragraph is harder for a reader to check against the original.

Should coded categories like BI-RADS be extracted from a radiology report?

Coded assessment categories should be extracted from a radiology report only when the report states one, and then exactly as written, with both the words and any number printed beside them. A category must never be inferred from the findings. If the report gives no category, the field stays empty and says so.

Mammography is the clearest case. The federal mammography rule, 21 CFR 900.12 on the eCFR reader (last amended March 10, 2023), requires each mammography report to include an "overall final assessment of findings" in one of a fixed set of categories. These include "Negative", "Benign", "Probably Benign", "Suspicious" and "Highly Suggestive of Malignancy". The rule adds that "the assessment statement is only the word or phrase within the quotation marks."

The same rule sets incomplete categories, such as "Incomplete: Need additional imaging evaluation", and four fixed statements of breast density. The rule does not use the name BI-RADS. When a report prints a label and a number beside the required words, the label and number are part of what was written.

So the capture rules are simple:

1. Store the statement exactly as printed, including the number if one appears: "BI-RADS 1: Negative" (example).
2. Store breast density separately, using the report's exact sentence.
3. Never derive a category from the findings or the impression wording.
4. Flag a mismatch, such as a number and a word that do not belong together, for review rather than correcting it.

When a report states a category from any other named scale, apply the same rule: capture what is written, infer nothing.

Preliminary reads and addenda that change a radiology report

A preliminary radiology read is an early report issued before the final one, and an addendum is text added after a report is final. Either can change what the report says. Extraction must keep every version as a separate record with its status, date, time and author, and must never let a later version silently replace an earlier one.

The ACR practice parameter says a preliminary report "very likely will contain limited or incomplete information". It says preliminary reports should be "appropriately labeled as a preliminary report, distinct from the final report, and archived because clinical decisions may have been based on the preliminary report." It also says that documentation of communicating any discrepancy "should be incorporated into the final report."

Status labels give each version its place. The HL7 FHIR R4 DiagnosticReportStatus code system (version 4.0.1) offers a useful vocabulary: "preliminary" for verified early results, "final" for a complete report verified by an authorized person, and "amended", "corrected" and "appended" for changes after final. "Appended" means new content was added and "the existing content is unchanged."

For each version of a report, capture:

• The status as printed: "PRELIMINARY", "FINAL", "ADDENDUM".
• The date and time of that version.
• The author or signer of that version. The hospital radiology rule, 42 CFR 482.26 on the eCFR reader (last amended August 4, 1986), says the practitioner who performs radiology services must sign reports of their interpretations.
• The accession number, which ties the versions together.
• The full text of that version, verbatim.

When a final report and its preliminary read differ, do not compare them yourself and record "changed". Keep both texts and let a clinician read them. Emergency visits often produce a preliminary read during the visit and a final report later. The post on emergency department record extraction covers how those visits are assembled.

Radiology report sections and the fields each yields

This table maps each section of a fictional radiology report to the fields it yields and whether the text stays verbatim. The patient, staff, dates and numbers are examples.

Only the header, the dates, the status and the signer are true fields here. Everything that interprets the images stays as text.

Lab and pathology reports follow different layouts, with result tables, units and reference ranges. Those are covered in the lab report posts. Dated records with one entry per event, such as vaccines, work differently again, as the post on immunization record extraction explains.

Radiology reports are not named among the document types in Schema Care's supported documents section, so read this table as a method, not a feature description. This post gives no medical advice. It covers only how to carry a report's words and dates into data without changing them.