Natural language processing automatically extracts valuable measurement information from radiology reports to help providers standardize follow-up care for incidentally detected pancreatic lesions.
The widespread use of abdominal CT and MRI has led radiologists to discover more, small pancreatic cystic lesions. But this hasn’t coincided with well-established management approaches for small PCLs, researchers explained Dec. 22 in Radiology: Artificial Intelligence.
So, they developed an NLP tool to scour more than 400,000 free-text reports for new insights. The process automatically spotted patients with PCLs and extracted measurements that corresponded to radiologist assessments.
The Stanford team says other organizations can likely apply the algorithm to their own reports and records, which may help develop diagnostic benchmarks.
“Robust NLP tools provide a potential opportunity to mine these historical archives across multiple institutions to provide
evidence-based recommendations based on natural history as assessed on serial imaging examinations,” Rikiya Yamashita, with the Department of Biomedical Data Science at Stanford University and co-authors wrote.
The NLP tool was trained to identify PCLs, and judged on a sample of 1,000 free-text CT and MRI reports generated between January 1991 and July 2019. It spotted lesions in 2.7% of CT reports and 10.2% of MRI documents, achieving near perfect agreement with radiologists.
For part two of the study, Yamashita et al. had the system grab PCL measurements from the same reports. NLP notched an overall accuracy of 0.96 when compared to radiologists’ annotations.
Using these measurements, the team found that 7 of 9 pancreatic ductal adenocarcinomas were reported in patients with PCLs ranging from 5-25 mm. Three of those remained stable throughout the follow-up period.
Given this finding, they argued that the American College of Radiology’s recently released baseline size group and interval growth recommendations “may not be sufficient” for determining the risk of smaller pancreatic cystic lesions.
Going forward, however, more research will be needed to confirm the results of their findings, the authors concluded.
You can read the entire study here.
Matt joined Chicago’s TriMed team in 2018 covering all areas of health imaging after two years reporting on the hospital field. He holds a bachelor’s in English from UIC, and enjoys a good cup of coffee and an interesting documentary.