Generative AI increases efficiency, quality of radiology reports

Radiologists can harness generative AI to increase the efficiency and quality of their reports, according to new research published Tuesday. 

Prompt, accurate image interpretation is crucial to patient care, with chest radiographs—the most common radiologic exam—requiring considerable time and expertise, experts note. Researchers aimed to use generative artificial intelligence to aid in this task, sharing their work in RSNA’s Radiology. 

They used a set of 758 chest X-rays, asking five radiologists to interpret images both with and without the help of AI. Introducing AI-generated reports helped cut reading times from 34.2 seconds down to 19.8. Agreement and quality scores also improved (both P < .001) while sensitivities increased for certain clinical concerns. 

“The results of our study demonstrated that preliminary reports created by a multimodal generative AI model could aid radiologists in chest radiograph interpretation in terms of reading time, report quality, and accuracy,” radiologist Eun Kyoung (Amy) Hong, MD, PhD, and co-authors wrote March 11. 

Researchers utilized the publicly available PadChest set of X-ray images, obtained between 2009 to 2017, for their research. The model deployed (AIRead) comes from Seoul artificial intelligence firm Soombit.ai. Radiologists participated in the study in late 2023, conducting two separate sessions. In the first, they evaluated chest X-rays and created reports without AI, while the second had them use AI-generated documents as preliminary drafts. Two separate thoracic radiologists with over 20 years of experience separately evaluated the accuracy and quality of reports from both sessions using a five-point scale. 

Report agreement scores—the degree of concordance between the radiologist’s opinion and the report—were 5 both with and without the help of AI. Report quality also stayed the same with a median score of 4.5 with AI and the same without. Radiologists’ sensitivity for detecting various abnormalities increased “significantly,” the authors noted, with gains for mediastinal silhouettes (increasing from 84.3% to 90.8%) and pleural lesions (77.7% to 87.4%). 

“Additionally, we noticed individual variability in the effect of introducing AI-generated reports in improving radiologists’ performance,” the authors noted. “For example, reading time increased for one reader, who tended to be skeptical of the AI-generated reports throughout the study, whereas other readers tended to be more accepting and demonstrated high preference for AI-generated reports as preliminary readings.”

In a corresponding editorial, experts noted that multimodal GenAI presents a “transformative opportunity” to increase the efficiency and accuracy of radiologist reporting. However, ongoing refinement of these models will be “critical” in safeguarding quality and patient safety. 

“As AI reporting tools continue to be developed, close attention must be paid to developing standardized methods for evaluating AI-generated reports and educating radiologists about the limitations of AI reporting systems,” Paul S. Babyn, MD, and Scott J. Adams, MD, PhD, with the University of Saskatchewan, wrote Tuesday. “Ultimately, if developed and integrated thoughtfully, GenAI tools—working in synergy with radiologist expertise—have the potential to enhance workflow efficiency and elevate the overall standard of patient care.”