MedVersa: A Generalist Foundation Model for Diverse Medical Imaging Tasks.
Authors
Affiliations (5)
Affiliations (5)
- Department of Biomedical Informatics, Harvard Medical School, Boston, MA, USA.
- School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University, Beijing, China.
- Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry, India.
- Department of Radiodiagnosis and Interventional Radiology, All India Institute of Medical Sciences, New Delhi, India.
- Scripps Research Translational Institute, Scripps Research, La Jolla, CA, USA.
Abstract
Current medical artificial intelligence (AI) systems are often limited to narrow, task-specific applications, which constrains their impact on routine clinical care. Generalist models that can learn from multimodal inputs and produce flexible outputs across imaging workflows may better support scalable, real-world deployment. We developed MedVersa, a generalist foundation model trained on tens of millions of compiled medical instances from diverse imaging scenarios. MedVersa is designed to accept heterogeneous inputs, including images and clinical text, and generate outputs such as classifications, segmentations, and narrative reports. We evaluated MedVersa on nine medical imaging tasks against leading task-specialized systems, and we conducted blinded radiologist assessments of AI-generated and human-written reports, as well as prospective user studies measuring report-writing time and discrepancy rates. Across tasks, MedVersa achieved performance that was competitive with or exceeded that of specialist systems. In blinded evaluations of chest radiograph reports, radiologists judged AI-generated and human-written reports to be clinically equivalent in 64% of all cases, with equivalence rising to 91% for normal studies. In user studies, the use of MedVersa was associated with notable reductions in report writing time and clinically relevant discrepancies compared with standard reporting workflows. MedVersa demonstrates that a single, flexible, multimodal foundation model can match or surpass task-specialized medical imaging systems while supporting radiologist efficiency and consistency. These findings highlight the potential of generalist multimodal AI systems to advance AI-enabled medical imaging workflows and augment clinical expertise.