A study published in the Journal of Digital Imaging in late June looked into how useful artificial intelligence (AI) is in helping healthcare providers extract useful information from radiology reports in real time. According to the report’s authors, their study is unique because it honed in on providing additional, useful context that’s frequently omitted during other information extraction (IE) tasks. Amazingly, the study found that machine learning (ML) is able to help healthcare professionals extract all relevant facts from radiology reports.
“Here, we develop a prototype system to extract all useful information in abdominopelvic radiology reports (findings, recommendations, clinical history, procedures, imaging indications and limitations, etc.), in the form of complete, contextualized facts.” lead author Jackson M. Steinkamp writes. He continues: “We construct an information schema to capture the bulk of information in reports, develop real-time ML models to extract this information, and demonstrate the feasibility and performance of the system.”
The information schema used was developed from a sample of 120 pelvic and abdominal radiology reports sourced from one institution between 2013 and 2018. Their dataset featured 48 MRIs, 22 ultrasounds, and 50 CT scans. Every bit of information was linked to a specific part of the document using customized labeling software. They also developed a two-part neural network to extract useful information on an as-needed basis. Ultimately, the team’s schema labeled over 5,000 facts and 15,000 distinct bits of information. This meant that over 86% of the radiology reports’ text corresponded to at least one labeled fact.
After labeling about 50 reports, according to the study’s authors, they no longer even needed to add any new facts. To them, this suggested that they achieved relative content saturation within the reports they were using as source material.
One of the study’s key findings was that it worked so well that it was able to zero in on facts it hadn’t before “seen.” This showed to the research team that the system wasn’t limited by any “pre-specified vocabularies and ontologies.” The team did, however, concede that their system sometimes missed rare information types, and cannot process certain types of “implied” information that isn’t directly referenced in the text it reads.
Despite the system’s shortcomings, the research team views their work as promising in the movement towards IE tasks in radiology. They plan for future research to incorporate increasingly sophisticated language models, as well as the creation of downstream applications.