Mosaic Clinical Technologies Reports AI Vision-Language Models Improve Radiology Detection by Up to 110% Across More Than 237,000 Imaging Exams
Mosaic Clinical Technologies, Inc. has announced promising preliminary research findings demonstrating the potential of its investigational artificial intelligence (AI) vision-language models (VLMs) to improve the detection of clinically important findings in medical imaging. An internal research study evaluating more than 237,000 computed tomography (CT) and X-ray examinations found that the company’s models identified additional clinically significant findings across all 50 conditions assessed.
The models enhanced detection by an average of approximately 40%, with improvements reaching as high as 110%. The research also indicated that AI-generated draft reports achieved clinically significant error rates comparable to or lower than those in original radiologist reports, while radiologists using the technology experienced efficiency improvements of more than 27% for CT examinations and 31% for X-rays.
Mosaic Clinical Technologies, Inc. (Mosaic) announced the findings from its ongoing internal research program evaluating the performance of investigational vision-language models developed by Cognita Imaging, Inc. (Cognita), its advanced artificial intelligence research laboratory. The study examined how these proprietary AI models performed across a large collection of medical imaging examinations, with the objective of assessing their ability to support radiologists in identifying clinically meaningful abnormalities and preparing comprehensive diagnostic reports.
The research involved 237,905 previously interpreted CT and X-ray examinations. Across all 50 conditions evaluated, Cognita’s VLMs identified additional clinically important findings that were subsequently confirmed by expert radiologists. The results demonstrated an average enhanced detection improvement of approximately 40%, with certain conditions showing improvements of up to 110%.
In addition to detection performance, the study evaluated the accuracy and quality of AI-generated draft reports. Preliminary findings showed that clinically significant error rates in the AI-drafted reports were comparable to or lower than those observed in the original reports prepared by radiologists without AI assistance.
The company also assessed the technology in investigational clinical settings, where radiologists using the models experienced average efficiency improvements exceeding 27% for CT examinations and 31% for X-ray examinations. These findings suggest that comprehensive AI-based reporting systems could potentially help radiology departments manage increasing imaging workloads while supporting diagnostic quality.
Addressing Growing Demand for Medical Imaging
The demand for diagnostic medical imaging continues to increase across healthcare systems, placing additional pressure on radiologists and diagnostic imaging departments. The growing volume of CT scans and X-ray examinations, combined with a shortage of radiology professionals, has created challenges in delivering timely and accurate interpretations.
Radiologists must review complex imaging studies, identify abnormalities, compare current examinations with previous scans, consider relevant patient information and prepare detailed reports. These responsibilities require substantial time and specialized expertise, particularly when imaging volumes are high.
Many existing radiology AI solutions are designed to perform individual diagnostic tasks. For example, certain tools focus on identifying specific conditions such as intracranial hemorrhage, fractures or other targeted abnormalities. Although these technologies can help identify particular findings, radiologists must still independently evaluate the remainder of the examination and prepare the complete diagnostic report.
Mosaic is pursuing a broader approach through its proprietary vision-language models. Rather than concentrating on a single abnormality, the investigational models are designed to analyze the entire imaging examination and generate comprehensive natural-language observations.
The technology evaluates the current examination alongside relevant prior imaging studies and available clinical context. It then produces an AI-generated draft containing potential findings for radiologists to review.
Radiologists remain responsible for evaluating the proposed observations and determining whether they should be accepted, modified or rejected. Once the findings are confirmed, Mosaic’s integrated reporting capabilities transform them into a complete radiology report in real time.
This approach is intended to support the entire reporting workflow rather than simply identifying isolated abnormalities. By combining image analysis, clinical context and automated report generation, Mosaic aims to help radiologists improve productivity while maintaining professional oversight of diagnostic decisions.
Cognita’s Vision-Language Models Support Multiple Imaging Categories
Cognita’s investigational VLMs have been developed to support a broad range of medical imaging examinations, including several categories of X-rays and CT scans.
For X-ray imaging, the models support chest, abdomen and musculoskeletal examinations. Their CT capabilities cover examinations of the head, cervical spine, chest, abdomen and pelvis.
The models are designed to recognize multiple types of clinically relevant findings across these imaging categories and generate corresponding natural-language observations.
Radiologists working within the Radiology Partners Clinical Services practice are using the models under an investigational research protocol. The research program is intended to generate evidence supporting the company’s efforts to obtain authorization from the U.S. Food and Drug Administration (FDA).
An important regulatory milestone was achieved in March 2026, when Cognita’s comprehensive chest X-ray model became the first generative AI radiology model to receive FDA Breakthrough Device designation.
The FDA Breakthrough Devices Program is designed to accelerate the development and review of medical devices and technologies that have the potential to provide more effective diagnosis or treatment for serious or life-threatening conditions.
The designation provides eligible developers with prioritized interaction with the FDA during the regulatory review process. For Cognita, this milestone represents progress in advancing its generative AI technology toward potential clinical use.
The company continues to evaluate its models through research and investigational clinical applications, with the objective of demonstrating their performance, reliability and potential contribution to radiology workflows.
Large-Scale Research Evaluated More Than 237,000 Examinations
To assess the performance of its AI models before broader deployment, Mosaic’s research team conducted a retrospective analysis involving 237,905 CT and X-ray examinations that had originally been interpreted by radiologists without AI assistance.
The evaluation examined areas of agreement and disagreement between the original radiologist reports and the findings generated by the VLMs. A total of 50 clinically significant conditions were included across the imaging categories.
Following the initial analysis, panels of U.S. board-certified radiologists reviewed a substantial random subset of cases in which the original radiologist report, the AI-generated draft or both identified a potential finding.
Importantly, the expert reviewers were blinded to the source of each finding. This meant that they did not know whether an observation had originated from the original radiologist or the AI model.
The independent review process was designed to determine whether the identified findings were clinically valid and to establish how many additional abnormalities could be confirmed through expert evaluation.
Mosaic used this methodology to calculate the enhanced detection rate (EDR), a measure of the additional clinically important findings identified by the VLMs and confirmed by expert radiologists.
The results showed additional detection improvements across all 50 conditions assessed, although the magnitude of improvement varied depending on the imaging category and specific condition.
CT Imaging Demonstrates Average Enhanced Detection of 40.9%
The CT component of the research included 71,707 examinations covering the head, cervical spine, chest, abdomen and pelvis.
Across 35 broadly defined clinical conditions, the VLMs achieved an average enhanced detection rate of 40.9%. Individual condition results ranged from 5.2% to 109.6%, demonstrating improvements across multiple categories of diagnostic findings.
Several clinically important abnormalities showed particularly encouraging detection results.
For example, the models achieved an enhanced detection rate of 73.4% for fractures identified through head CT examinations. For fractures detected in abdomen and pelvis CT examinations, the enhanced detection rate reached 68.1%.
The models also demonstrated a 50.4% enhanced detection rate for pancreatic masses, an important category of findings that may require further clinical investigation and management.
For free air in the abdomen, the enhanced detection rate was 40.6%. Identifying this finding can be clinically important because it may indicate an underlying condition requiring urgent evaluation.
These findings indicate that the models could potentially help radiologists identify additional abnormalities across different anatomical regions rather than focusing on one specific diagnostic category.
The research team emphasized that expert radiologist confirmation was central to the evaluation, helping distinguish clinically meaningful findings from observations that might not be relevant to patient care.
Chest X-Ray Models Show 43.4% Average Detection Improvement
The chest X-ray component represented the largest individual imaging category in the study, covering 100,000 examinations.
Across 10 evaluated conditions, the VLMs achieved an average enhanced detection rate of 43.4%, with individual results ranging from 12.7% to 57.5%.
The highest improvement was observed in the detection of lung nodules or masses, where the enhanced detection rate reached 57.5%.
The models also demonstrated a 51.6% enhanced detection rate for malpositioned medical devices and a 46.5% rate for pneumothorax.
These findings are relevant to routine chest imaging, which is widely used across hospitals and healthcare facilities for evaluating respiratory conditions, monitoring disease progression and assessing the position of medical devices.
Chest X-ray examinations frequently require radiologists to identify multiple findings within a single image. The ability to analyze the entire examination and generate structured observations could help support a more comprehensive review process.
Mosaic’s approach is designed to integrate these observations into a draft report that radiologists can evaluate and finalize, rather than requiring them to work with separate AI-generated alerts.
Musculoskeletal and Abdomen X-Ray Results
The research also evaluated 66,198 musculoskeletal and abdomen X-ray examinations.
Across five assessed conditions, the VLMs achieved an average enhanced detection rate of 23.9%, with individual results ranging from 6.8% to 45.8%.
Among the findings, the models achieved a 27.1% enhanced detection rate for bone infections and 45.8% for suspicious bone lesions.
Although these findings may be less frequently encountered than some common imaging abnormalities, they can have considerable clinical importance.
Bone infections and suspicious bone lesions may require additional diagnostic testing, specialist assessment or further clinical investigation. Identifying such abnormalities during the initial imaging interpretation can support timely follow-up.
The results suggest that comprehensive AI models may have value in identifying both common imaging abnormalities and less frequently encountered findings that could otherwise be overlooked.
AI-Generated Reports Demonstrate Encouraging Quality Results
In addition to evaluating enhanced detection, Mosaic’s research team examined the quality and accuracy of the AI-generated draft reports.
Expert radiologists reviewed the reports to identify errors and assess whether the observations were clinically appropriate.
For chest X-ray examinations, 91.2% of the VLM-generated drafts were rated as error-free. This compared with 83.0% of the original radiologist-only reports.
The findings suggest that the AI-generated drafts performed favorably in this evaluation, although the results remain preliminary and should be interpreted within the context of the study methodology.
For musculoskeletal and abdomen X-ray examinations, 3.7% of the VLM-generated drafts contained a clinically significant error, compared with 6.5% of the original reports prepared without AI assistance.
For CT examinations, the clinically significant error rates associated with AI-generated drafts were comparable to or lower than those observed in the original radiologist-only reports.
These findings are important because improving detection alone is not sufficient for clinical AI applications. AI systems must also generate reliable observations and avoid introducing clinically significant errors into diagnostic documentation.
Mosaic’s evaluation therefore considered both the additional findings identified by the models and the quality of the resulting draft reports.
The company continues to position radiologists as the final decision-makers, with AI-generated observations subject to professional review, editing and approval.
Clinical Use Shows Improvements in Radiologist Efficiency
Recognizing that retrospective research findings do not necessarily translate directly into real-world clinical performance, Mosaic also evaluated the models during investigational clinical use.
The company assessed how radiologists performed when using AI-generated draft reports in their regular imaging interpretation workflows.
The analysis included radiologists who had interpreted at least 50 examinations supported by AI-generated drafts.
Compared with their individual performance before AI drafts became available, participating radiologists experienced average efficiency improvements of more than 27% for CT examinations and 31% for X-ray examinations.
These results provide additional evidence about the potential operational benefits of integrating comprehensive AI reporting technology into radiology workflows.
By generating draft observations and supporting report preparation, the models may help reduce the time radiologists spend documenting findings and preparing final reports.
However, the company continues to study human-AI collaboration to understand how performance changes as radiologists gain experience with the technology and adoption expands.
Nearly Eight Million Exams Drafted Through October 2026
Mosaic also reported substantial investigational use of its technology.
Through October 22, 2026, the models had generated draft reports for nearly 8 million examinations for radiologist review.
This included approximately:
- 6.3 million X-ray examinations.
- 1.7 million CT examinations.
- More than 37,000 examinations drafted per day on average.
The scale of this activity provides an opportunity for Mosaic to monitor how its AI systems perform across large numbers of examinations and different clinical workflows.
The company continuously evaluates interactions between radiologists and the AI models, including how frequently radiologists retain, modify or reject AI-generated findings.
It also monitors the time required for reports to be reviewed and signed.
These measurements are intended to help identify potential safety concerns, including excessive reliance on AI-generated observations, changes in reporting behavior and other signals that may require further investigation or intervention.
Maintaining appropriate human oversight is a central element of Mosaic’s approach. The company aims to ensure that AI supports radiologists rather than replacing their clinical judgment.
Company Leadership Highlights Potential Impact on Patient Care
Jason Poff, MD, Director of Clinical AI at Mosaic Clinical Technologies, said the VLMs identified additional clinically important findings across every condition evaluated in both CT and X-ray examinations.
He emphasized that the technology is designed to keep radiologists at the center of the reporting process while potentially improving diagnostic quality and operational efficiency.
Poff also noted that early post-deployment monitoring has produced encouraging quality results. The company will continue evaluating human-AI performance as more radiologists adopt the technology and become familiar with its capabilities.
Jon Foster, Executive Vice President and Chief Operating Officer of HCA Healthcare, Inc., highlighted the importance of evaluating clinical quality alongside operational efficiency.
He said research conducted through the Radiology Partners platform provides the type of evidence healthcare systems need when assessing the effectiveness and potential value of AI-based clinical tools.
Rich Whitney, Chairman and Chief Executive Officer of Radiology Partners, Inc., described the findings as a major development for radiology and highlighted the potential of proprietary vision-language models to advance patient care.
He also emphasized the potential role of these technologies in helping address the impact of the radiologist workforce shortage on healthcare systems.
Research Presentations Planned at RSNA 2026
Mosaic confirmed that both research papers detailing its CT and X-ray report-drafting findings are available through the company’s RSNA Event Hub.
Mosaic Clinical Technologies, Cognita Imaging and Radiology Partners Clinical Services will participate in the 2026 Radiological Society of North America (RSNA) Annual Meeting, scheduled for November 29 through December 3, 2026, in Chicago.
The organizations will be located at Booth 5304 in the South Hall AI Showcase, where attendees can learn more about their research and investigational AI technologies.
Jason Poff, MD; Nina Kottler, MD, Chief Medical AI Officer; and Louis Blankemeier, Ph.D., Chief Executive Officer of Cognita, will be available for interviews before and during the event.
The research findings highlight the potential of comprehensive vision-language models to support radiology through improved detection, automated draft reporting and increased workflow efficiency. As Mosaic and Cognita continue their clinical research and regulatory efforts, further evaluation will be important to establish the technology’s safety, reliability and performance across broader healthcare settings.
With nearly 8 million examinations already processed for radiologist review through October 2026, the companies are continuing to investigate how AI-assisted reporting can contribute to diagnostic quality, radiologist productivity and the evolving demands of medical imaging services.
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.