top of page

How NLP Improves Insight Extraction From Radiology Reports

Radiology reports summarize findings from radiology studies and procedures. Due to the complexity of the field, most of the information is in textual format. Many health systems use reports following general guidelines as recommended by the American College of Radiology (ACR). These guidelines specify an overall format that includes sections on these topics (and in this order):

  • Indications: The overall clinical indications for an examination.

  • Procedures: A description of the procedures done as part of the examination, such as the administration of drugs or contrast agents.

  • Findings: From the exam and discussion of their significance. This section often includes information on comparisons to previous radiology studies if deemed appropriate by the radiologist.

  • Impressions: This section, sometimes referred to as the “conclusion” section, contains lists of summary statements indicating the conclusions from the examination. Recipients of the report tend to focus on this section, and some physicians have reported that this is the only section of a radiology report they commonly read.

Note that this is not a comprehensive list. The sections included in a radiology report can vary by radiology practice and by the modality of the exam. For example, for mammography reports, the Food and Drug Administration Quality Mammography Standards; Final Rule mandates that an “assessment” section be included in reports for mammograms. Similarly, ACR guidelines recommend including “recommendations” after the assessment section, and the ACR defines precise data elements for the representation of findings involving specific clinical characteristics.


Even with the standardization of vocabularies and coding terminology, substantial amounts of information are available from radiology reports. The use of Natural Language Processing (NLP) allows for the extraction of variables such as problems, location, size of lesions or tumors, etc., directly from the free text of the report. Tracking changes in tumor size/volume, for example, is an important part of assessing response to therapy.

Our clients, who were building extensive longitudinal disease registries, had a specific need to use NLP to extract information from the textual part of radiology reports to determine lung nodule sizes and volume, calculate estimated tumor mass, and to represent extracted data in the RadLex clinical vocabulary. RadLex, developed by the Radiological Society of North America, RadLex is, in their words, “a comprehensive set of radiology terms for use in radiology reporting, decision support, data mining, data registries, education and research,” and is an official “naming system” for HL7. Melax Tech was able to assist these clients by leveraging our extensive library of deep-learning-based clinical NLP pipelines in conjunction with our award-winning CLAMP NLP tool. The CLAMP tool uses an extensive variety of features specifically designed to allow clients to easily extract many common data elements and the relationships between them from clinical text. The tool also allows for data to be mapped to standard clinical vocabulary systems, such ICD-10, SNOMED-CT, and many other standard vocabularies, such as the RadLex terminology used in this solution.


In a closely related use case with a different client, our CLAMP NLP suite of tools was used to normalize references to and extract information about radiation treatment sites recorded as text in the radiation therapy delivery (record and verify) electronic medical record (MOSAIQ®, Elekta Care Management, Stockholm).


To learn more about our wide range of NLP solutions for various EHR document types, including radiology, radiation therapy, and pathology notes, and how they enhance research and billing functions, request a demo today!



For details on this use case, see Walker G, Soysal E, Xu H. Development of a Natural Language Processing Tool to Extract Radiation Treatment Sites. Cureus. 2019;11(10):e6010. Published 2019 Oct 28. doi:10.7759/cureus.6010.





bottom of page