Northwestern Medicine Pioneers Generative AI Radiology Tool, Boosting Efficiency Without Sacrificing Accuracy

A groundbreaking study published in JAMA Network Open highlights how Northwestern Medicine is using a custom-built generative AI system to significantly improve radiologist productivity—by as much as 40% in some cases—while maintaining high standards of accuracy and quality.

Developed entirely in-house, the AI model can generate near-complete, patient-specific radiology reports from X-rays, tailored to each radiologist’s individual style. In real-world testing across 12,000 X-ray interpretations, the system delivered a 15.5% improvement in documentation efficiency without any reduction in diagnostic accuracy or report quality.

Researchers described the tool as “holistic,” capable of analyzing entire images for all potential clinical issues. The output—a report that is about 95% complete—is personalized for the patient and easily reviewed, edited, and finalized by the radiologist.

“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 Samir F. Abboud, MD, chief of emergency radiology at Northwestern Medicine.

Rather than adapting existing large language models such as ChatGPT, Northwestern’s engineers built their system from scratch using the organization’s own clinical data. This approach resulted in a “lightweight, nimble” AI model tailored to radiology needs. Embedded engineering specialists work alongside clinicians to refine the system, and the team has already secured two patents. The technology is in the early stages of commercialization.

Co-author Mozziyar Etemadi, MD, PhD, stressed the importance of accessibility. “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,” he said. “We believe that this democratization of access to AI is the key to drive adoption worldwide.”

The five-month prospective study, concluding in April 2024, involved nearly 24,000 radiographs. Half were interpreted with AI assistance, and half without. On average, AI-assisted interpretations took 160 seconds, compared to 189 seconds without the tool. A peer review of 800 cases found no difference in accuracy or text quality.

The model also demonstrated clinical vigilance, flagging unexpected cases of pneumothorax (collapsed lung) with 72.7% sensitivity and 99.9% specificity among almost 98,000 studies screened. Early, unpublished results suggest some radiologists experienced efficiency gains of up to 80%, and the tool has proven effective in analyzing CT scans as well.

Northwestern claims this is the first generative AI radiology tool to be integrated into live clinical workflows and to achieve high performance across all X-ray types, from “skulls to toes.” The system also monitors for urgent findings, cross-referencing patient records and immediately alerting radiologists when new critical conditions arise. Future development aims to detect delayed or missed diagnoses, such as lung cancer.

While the technology could help address the nationwide radiologist shortage, its creators stress that it is not designed to replace physicians. “You still need a radiologist as the gold standard,” Abboud said. “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.”