How a health system automatically integrates AI results into radiology reports

An academic health system is sharing insights into how it automatically integrates artificial intelligence results into radiology reports. 

The U.S. Food and Drug Administration has now cleared more than 1,000 clinical AI applications, with radiology accounting for over 70% of the total. However, diagnostic imaging providers have sometimes struggled to integrate this technology into physicians’ regular daily workflows, experts wrote Monday. 

The University of Pennsylvania in Philadelphia has aimed to tackle this conundrum, developing a deep learning-based image analysis tool and AI “orchestrator.” Dubbed “AInsights,” the platform helps monitor and coordinate information exchange between clinical and AI systems, experts detailed in the Journal of Imaging Informatics in Medicine.

“For these advances to improve patient care, it is essential that the output from AI models be integrated easily and effectively into the key work product of diagnostic radiology, namely the radiology report,” Charles E. Kahn Jr., MD, MS, professor and vice chair of radiology at the Perelman School of Medicine, and colleagues wrote Jan. 27. “This article describes an innovative approach that utilizes [common data elements] to embed measurements derived from AI models directly into radiology reports in clinical practice,” they added later.

Such common data elements are units of information that consist of a specific, precise question and the corresponding set of allowable answers, the authors noted. Using this predefined value set is a “fundamental aspect of CDEs,” allowing them to facilitate the uniform collection and exchange of data, “ensuring consistency and interoperability across diverse healthcare systems.” In radiology, these common data elements provide a standardized approach for collecting and reporting info extracted from imaging exams—such as defining patient demographics, measurements, qualitative assessments and diagnostic interpretations. 

“By providing a clear framework for documentation, CDEs help maintain consistency and enable efficient data analysis, thereby supporting evidence-based clinical decision-making and advancements in patient care,” the authors noted. “Development, curation, and translation of CDEs into clinical practice has been a key informatics initiative in radiology.”

At Penn, imaging experts have utilized deep learning technology for anatomic segmentation of CT. Their initial focus has been around volumetry of the liver and spleen, along with opportunistic detection of hepatic steatosis in patients undergoing noncontrast exams of the abdomen and pelvis. To capture measurements generated by their AInSights platform, Kahn and colleagues defined a set of common data elements. These were then submitted to the American College of Radiology and RSNA, which have collaboratively produced the RadElement.org resource to aid imaging departments in these endeavors. 

“The development of CDEs involves a rigorous process that includes consensus-building among experts in relevant fields,” the authors noted. “This process ensures that each element is clearly defined, clinically relevant, and widely acceptable. For radiology reports specifically, CDEs have been pivotal in advancing structured reporting—a format which enhances report clarity by systematically organizing information according to these predefined elements.”

After revisions, the final common data elements were published in March 2023 under the title “Morphometric CT Quantification of Liver, Spleen, and Abdominal Fat.” With Penn’s implementation, AInSights generated various image series, such as image overlays, transmitting them to the health system’s PACS. The AI model was able to successfully segment liver and spleen in noncontrast CT exams, generating measurements of the two organs’ volume and attenuation. Values were automatically incorporated successfully into reports. Between May 2023 and February 2024, the system analyzed a total of 3,920 exams, of which 339 (or 8.7%) were positive for hepatic steatosis. 

“As artificial intelligence systems for image analysis have demonstrated increasingly powerful performance, they play an increasing role in clinical medicine in domains such as radiology,” the authors advised. “It is critical that the data generated by such systems can be integrated readily with imaging reports, with imaging data, and with the EHR. The use of CDEs in AI-based clinical workflows allows for this data integration. The utilization of CDEs in radiology reporting not only optimizes clinical workflows but also opens new avenues for data-driven research, quality improvement initiatives, and benchmarking of diagnostic practices.”

You can read much more of the technical details about the investigation in the official journal of the Society for Imaging Informatics in Medicine here. The study is open access and does not require a subscription or sign in.