Vision-Language Model-Powered Radiology Shows Gains in Detection and Efficiency in Research Study of Over 237,000 Exams

The models identified additional expert-confirmed findings for every condition tested, averaging approximately 40% enhanced detection, with error rates comparable to or lower than radiologists’ original reports and efficiency gains of 27% on CT and 31% on X-ray

NASHVILLE, Tenn.--(BUSINESS WIRE)--Mosaic Clinical Technologies, Inc.™ (Mosaic) today announced results from an ongoing, internal research study of how investigational vision-language models (VLMs)* performed on more than 237,000 CT and X-ray exams. Across all 50 conditions evaluated, VLMs developed by Cognita Imaging, Inc.™ (Cognita), Mosaic’s advanced AI lab, identified additional clinically important findings as confirmed by expert radiologists, enhancing detection by an average of approximately 40% and by as much as 110%. The models contribute to a draft report in MosaicOS, Mosaic’s cloud-native radiology operating system, for radiologists to review, edit and sign. Preliminary data also showed that clinically significant error rates in AI-drafted reports were comparable to or lower than those in the original radiologist reports for every model evaluated. In clinical research, radiologists using the models also saw average efficiency rise more than 27% for CT and 31% for X-ray.

Demand for medical imaging continues to grow much faster than the radiologist workforce, increasing pressure to deliver accurate, timely interpretations. Most radiology AI tools available today are built for a single task, such as flagging a possible intracranial hemorrhage or fracture, leaving radiologists to identify and describe all other findings unaided and generate the full report. Mosaic takes a more holistic approach: its investigational, proprietary VLMs analyze the entire exam, the prior exam and relevant clinical context to generate comprehensive natural-language observations that radiologists can confirm, edit or reject. Mosaic’s integrated reporting tools then transform the confirmed findings into a complete radiology report in real time.

Cognita’s VLMs support a broad range of X-ray (chest, abdomen and musculoskeletal) and CT exams (head, cervical spine, chest, abdomen and pelvis). The models are being used by radiologists in the Radiology Partners Clinical Services practice under a research protocol to support U.S. Food and Drug Administration (FDA) authorization. In March 2026, Cognita’s comprehensive chest X-ray model became the first generative AI radiology model to receive FDA Breakthrough Device designation, which is intended to speed review of promising technologies and gives Cognita prioritized interaction with the agency as the model moves through the review process.

To evaluate the models before deployment, the research team retrospectively analyzed 237,905 exams that radiologists had previously interpreted without AI assistance, identifying radiologist-VLM concordance and discordance across a total of 50 significant conditions. Panels of U.S. board-certified radiologists then reviewed a large random subset of cases in which the AI draft, the original radiologist report or both flagged a finding. Reviewers were blinded to the source of each finding. This approach allowed the research team to measure the additional, expert-confirmed findings the VLMs identified, known as the enhanced detection rate (EDR). Highlights included:

• CT quality results (71,707 exams): Substantial enhanced detection averaging 40.9% (range 5.2% to 109.6%) across 35 broadly-defined conditions on CT studies of the head, cervical spine, chest, abdomen and pelvis. For example, enhanced detection reached 73.4% for fractures on head CT, 68.1% for fractures on abdomen and pelvis CT, 50.4% for pancreatic masses and 40.6% for free air in the abdomen.
• Chest X-ray quality results (100,000 exams): Substantial enhanced detection averaging 43.4% (range 12.7% to 57.5%) across 10 conditions, including 57.5% for lung nodules or masses, 51.6% for malpositioned devices and 46.5% for pneumothorax.
• Musculoskeletal and abdomen X-ray quality results (66,198 exams): Substantial enhanced detection averaging 23.9% (range 6.8% to 45.8%) across five conditions, including 27.1% for bone infections and 45.8% for suspicious bone lesions, uncommon findings with high clinical impact.

Expert reviewers also rated the VLM drafts for errors. On chest X-rays, 91.2% of VLM drafts were error-free, compared with 83.0% of the original radiologist-only reports. On musculoskeletal and abdomen X-rays, 3.7% of VLM drafts contained a clinically significant error, compared with 6.5% of the original radiologist-only reports. On CT, the VLM drafts’ clinically significant error rates were comparable to or lower than those of the original radiologist-only reports.

Because promising results from retrospective studies do not always translate into real-world clinical practice, the team also measured how radiologists performed with the models in investigational clinical use. Among radiologists who had read at least 50 AI-drafted exams, average efficiency improved by more than 27% on CT exams and by 31% on X-rays, compared with their own performance before AI drafts were available.

Through Oct. 22, 2026, the investigational models had drafted nearly 8 million exams for radiologist review, including about 6.3 million X-rays and 1.7 million CT scans, and were drafting an average of more than 37,000 exams a day. Mosaic continuously monitors how radiologists and the models work together, including how often radiologists keep or change AI-drafted findings and how long reports take to sign, to identify and correct over-reliance on AI and other safety signals that require further review or intervention.

“Across both X-ray and CT, our VLMs identified additional clinically important findings for every condition we evaluated, while keeping radiologists at the center of the process. These preliminary results suggest our VLMs can support radiologists in improving both quality and efficiency. Early post-deployment monitoring has also shown encouraging quality results, and we continue to study human-AI performance as adoption expands and radiologists gain experience with the technology,” said Jason Poff, MD, Director, Clinical AI, Mosaic Clinical Technologies.

“The research that RP is doing using its Mosaic platform is measuring quality alongside efficiency. It’s the kind of evidence health systems need to evaluate the efficacy of these kinds of tools,” said Jon Foster, Executive Vice President and Chief Operating Officer of HCA Healthcare, Inc.

“The results of this research using our proprietary and fully sovereign vision-language models are nothing short of stunning in terms of the potential for the advancement of patient care, as well as in alleviating the impact on our healthcare system of the crippling radiologist shortage. In my opinion, radiology just had its ‘ChatGPT moment,’” said Rich Whitney, chairman and Chief Executive Officer of Radiology Partners, Inc.

Both research papers are available at Mosaic’s RSNA Event Hub. Mosaic, Cognita and Radiology Partners Clinical Services will be at RSNA 2026, Nov. 29-Dec. 3 in Chicago, at booth 5304 in the South Hall AI Showcase. Jason Poff, MD, Nina Kottler, MD, Chief Medical AI Officer and Louis Blankemeier, Ph.D., Chief Executive Officer of Cognita, are available for interviews before and during RSNA.

*CAUTION: Investigational device. Not for sale in the United States.

About Mosaic Clinical Technologies, Inc.

Mosaic Clinical Technologies, Inc., a wholly owned subsidiary of Radiology Partners, Inc., develops advanced technology and proprietary vision-language AI models for medical imaging, built in-house in collaboration with radiologists in the Radiology Partners Clinical Services practice. Its MosaicOS platform brings diagnostic tools, cutting-edge AI and radiology workflows together in a cloud-native operating system designed to support radiologists and the healthcare organizations they serve. Learn more at MosaicClinical.ai and connect with us on LinkedIn and YouTube.

About Cognita Imaging, Inc.

Cognita Imaging, Inc., Mosaic Clinical Technologies, Inc.’s advanced AI lab, develops proprietary vision-language foundation models and other generative AI capabilities for radiology. Its work is focused on helping radiologists deliver more accurate, efficient care and expanding access to high-quality healthcare. Cognita seeks to elevate the performance of all radiologists to the level of the best expert subspecialists and, longer term, to empower radiologists with digital imaging information not easily perceptible to the human eye. Learn more at cognita.ai.

About Radiology Partners, Inc.

Radiology Partners, Inc., through RP Clinical Services and its managed practices, together with Mosaic Clinical Technologies, Inc., provides radiology services and advanced technology to more than 3,400 hospitals and other healthcare facilities. Its network includes more than 4,000 radiologists who interpret more than 55 million imaging studies each year. Learn more at radpartners.com and connect with us on LinkedIn, X, Instagram and YouTube.