A new AI system has boosted radiology productivity by up to 40% in the first real-world deployment of generative artificial intelligence for medical imaging interpretation.
The tool, developed at Northwestern Medicine and tested across 11 hospitals, generated nearly 24,000 radiology reports over five months while maintaining clinical accuracy and identifying life-threatening conditions within seconds of imaging completion.
Real-World Performance Exceeds Expectations
Unlike previous AI tools that focus on single conditions, Northwestern’s system analyzes entire X-rays and CT scans, producing complete reports that radiologists can review and finalize. The average efficiency gain reached 15.5%, with some radiologists achieving improvements as high as 40%.
“This is, to my knowledge, the first use of AI that demonstrably improves productivity, especially in health care. Even in other fields, I haven’t seen anything close to a 40% boost,” said senior author Dr. Mozziyar Etemadi, an assistant professor of anesthesiology at Northwestern University Feinberg School of Medicine.
Built from Scratch, Not Big Tech
Rather than adapting existing models like ChatGPT, Northwestern engineers built their system from the ground up using clinical data from their own hospital network. This approach created a lightweight, specialized tool that requires far less computing power than commercial alternatives.
“There is no need for health systems to rely on tech giants,” said first author Dr. Jonathan Huang, a third-year medical student who holds a Ph.D. in biomedical engineering. “Our study shows that building custom AI models is well within reach of a typical health system, without reliance on expensive and opaque third-party tools like ChatGPT.”
Life-Saving Speed for Critical Cases
The system automatically flags life-threatening conditions like collapsed lungs before radiologists even view the images. During testing, it identified 72.7% of clinically significant pneumothorax cases with 99.9% specificity, delivering alerts within 24 seconds of image completion compared to the typical 24.5-minute delay for radiologist notifications.
“For me and my colleagues, it’s not an exaggeration to say that it doubled our efficiency. It’s such a tremendous advantage and force multiplier,” said co-author Dr. Samir Abboud, chief of emergency radiology at Northwestern Medicine.
Beyond Efficiency: Quality Maintained
Peer review of 800 studies showed no difference in clinical accuracy or textual quality between AI-assisted and traditional reports. The system maintained the same low rate of report corrections, with addenda needed in just 0.14% of AI-assisted reports compared to 0.13% of traditional reports.
What sets this system apart is its approach to report generation. Rather than simply flagging abnormalities, it produces 95% complete reports personalized to each patient and radiologist’s style, essentially functioning like a highly skilled trainee whose work requires final review.
Addressing Critical Shortage
The timing couldn’t be more crucial. By 2033, the U.S. expects a shortage of up to 42,000 radiologists as imaging volumes rise 5% annually while residency positions increase by just 2%. Northwestern’s tool offers a practical solution, helping radiologists clear backlogs and deliver results in hours instead of days.
During the study period, the AI system saved over 63 hours of documentation time, equivalent to reducing coverage needs from 79 to 67 radiologist shifts. This efficiency gain becomes particularly valuable in emergency situations where every minute counts.
Preventing Missed Diagnoses
The study revealed compelling examples of the system’s life-saving potential. In one case, a patient with a large pneumothorax was discharged from the emergency department after a preliminary reading missed the condition. The AI system had flagged it immediately, but since alerts weren’t yet live, the patient was only called back six hours later after an attending radiologist’s review.
Another patient undergoing pneumonia workup had a pneumothorax that went unnoticed until an oxygen desaturation event 11 hours after imaging. The AI had identified it within seconds of the scan completion.
Error Rates Tell the Story
The system’s sophistication shows in its editing patterns. For chest X-rays, the median word error rate between AI drafts and final reports was just 0.31, meaning roughly one in three words required modification. For non-chest studies, the rate was 0.63, reflecting the added complexity of musculoskeletal and other imaging.
Importantly, the system generated alerts for pneumothorax at a rate of just over one per day across the entire health system, demonstrating its ability to minimize alert fatigue while catching critical cases.
Future of Medical AI
“You still need a radiologist as the gold standard,” Abboud emphasized. “Medicine changes constantly — new drugs, new devices, new diagnoses — and we have to make sure the AI keeps up. Our role becomes ensuring every interpretation is right for the patient.”
The technology has two approved patents with others pending, and early-stage commercialization is underway. Etemadi’s team, which he describes as following a “Bell Labs-style” approach, has attracted talent from major tech and finance companies to work within the hospital system.
“We’re not just pushing health care AI forward — we’re advancing the fundamentals of AI at a fraction of the cost of the big AI labs. This is the start of the DeepSeek moment for health care AI,” Etemadi said.
The study, published in JAMA Network Open, represents the first prospective clinical evaluation of generative AI for radiology reporting, providing a roadmap for other health systems looking to improve efficiency without compromising patient care.
If our reporting has informed or inspired you, please consider making a donation. Every contribution, no matter the size, empowers us to continue delivering accurate, engaging, and trustworthy science and medical news. Independent journalism requires time, effort, and resources—your support ensures we can keep uncovering the stories that matter most to you.
Join us in making knowledge accessible and impactful. Thank you for standing with us!