Radiology Partners and its technology division, Mosaic Clinical Technologies, are asking the U.S. Food and Drug Administration (FDA) for greater clarity on how vision-language models (VLMs) used in medical imaging should be regulated. Mosaic submitted a citizen petition on August 12, arguing that these AI systems represent an important advance in diagnostic medicine but that uncertainty over how existing medical-device requirements apply is creating problems for developers, healthcare organizations, clinicians, and patients.
A key question is whether a commercially distributed imaging model should itself be considered a medical device when it is intended to be fine-tuned later using institution-specific data. This matters because VLMs are foundation models trained on large datasets and capable of performing multiple downstream tasks. Unlike conventional imaging AI designed for a narrowly defined purpose, these newer systems can potentially be adapted to different clinical environments, making it less obvious when FDA clearance or approval should be required.
The petition comes as regulators and the industry are already debating how to evaluate increasingly sophisticated medical AI. There is currently no established standard for validating the performance and quality of imaging foundation models, raising questions about how developers should demonstrate safety and effectiveness before deployment and how hospitals should validate models on their own patient populations. The FDA has separately released a discussion paper seeking feedback on regulating generative-AI medical devices, including risk assessment, premarket evaluation, and post-market monitoring.
The broader issue is finding a balance between innovation and patient safety. Radiology Partners is not simply asking for less regulation; it is seeking clearer and more consistent rules so developers and healthcare providers know what regulatory obligations apply to these rapidly evolving systems. The debate is particularly important because AI models can change after deployment and may perform differently across institutions and patient populations. Clear requirements for validation, monitoring, updates, and accountability could therefore help the FDA encourage adoption of useful imaging AI without sacrificing clinical safety.
