The FDA faces mounting pressure to establish clear rules for artificial intelligence tools that analyze medical images, as regulatory uncertainty threatens to slow adoption of promising diagnostic technologies. Radiology Partners and its technology division, Mosaic Clinical Technologies, submitted a formal petition to the U.S. Food and Drug Administration on August 12, asking the agency to clarify how existing medical device regulations should apply to commercially available AI vision-language models. These AI systems perform medical image analysis to support diagnosis and treatment, but lack standardized oversight.

Why Are Regulators Struggling to Keep Up With Medical AI?

Vision-language models represent a new category of AI that can process and interpret medical images in ways previous tools could not. Unlike traditional AI systems trained for a single narrow task, these foundation models are trained on massive datasets to solve multiple downstream tasks. They can be fine-tuned on institution-specific data after purchase, creating a gray area in regulatory responsibility.

The core problem is ambiguity. "Differing interpretations have emerged regarding the application of existing FDA medical device requirements, creating uncertainty for developers, healthcare organizations, clinicians and patients," wrote Mike Peresie, president of Mosaic Clinical Technologies, in the petition to the FDA. This uncertainty has real consequences. Radiologists and hospital systems currently bear substantial responsibility for determining whether these models are appropriate for their intended use, even though the models may not have undergone premarket safety and effectiveness review.

What Specific Risks Do Unregulated AI Models Pose?

Radiology Partners identified several safety concerns that underscore why regulatory clarity matters. Without standardized oversight, AI models risk being trained on biased or unrepresentative datasets, lacking transparency about how they reach conclusions, and missing basic cybersecurity protections. These gaps create tangible patient harm potential. Additionally, some developers disclaim responsibility for their models' reliability or compliance with legal requirements, providing few warranties about performance or fitness for diagnostic use.

The absence of established standards for validating AI model performance and quality means downstream users like radiologists cannot easily assess whether a tool is safe before deploying it in clinical settings. This puts clinicians in an impossible position: they must make safety determinations about complex AI systems without clear regulatory guidance or baseline performance data.

How Can Healthcare Organizations Navigate AI Deployment Safely?

While awaiting FDA guidance, healthcare organizations and developers can take several steps to reduce risk and ensure responsible AI adoption:

• Validation Requirements: Establish internal protocols to validate AI model performance on your institution's specific patient populations and imaging equipment before clinical deployment, rather than relying solely on vendor claims.
• Transparency Documentation: Require vendors to provide detailed documentation about model training data sources, potential biases, explainability features, and cybersecurity controls before integration into clinical workflows.
• Performance Monitoring: Implement ongoing monitoring systems to track AI model accuracy and safety in real-world clinical use, with clear protocols for flagging performance degradation or unexpected errors.
• Clinician Oversight: Maintain radiologist review of all AI-assisted diagnoses rather than treating AI outputs as autonomous decisions, ensuring human expertise remains central to patient care.
• Compliance Consistency: Document all AI deployment decisions and maintain records demonstrating that your institution applied consistent standards across all AI tools, protecting both patients and the organization legally.

Radiology Partners is asking the FDA to take specific regulatory actions. The petition seeks clarification on whether a model marketed with expectations of subsequent fine-tuning on institution-specific data should itself be classified as a medical device requiring premarket clearance or approval. The group also wants the FDA to ensure consistent application of medical device requirements across all commercially available vision-language models.

"Clarity on standards will ultimately accelerate both the development and adoption of safe and trusted clinical diagnostic technologies," stated Mike Peresie, president of Mosaic Clinical Technologies.

Mike Peresie, President of Mosaic Clinical Technologies

The timing of this petition reflects a broader shift in healthcare AI adoption. As these foundation models become increasingly commercialized and deployed in clinical settings, the regulatory gaps have grown more urgent. Peresie emphasized that clear expectations around validation, performance monitoring, transparency, and quality management "would support patient safety, promote responsible innovation and provide greater consistency across the healthcare ecosystem".

Peresie

The FDA's response could reshape how AI tools are developed and deployed across radiology departments nationwide. Clear regulatory standards would benefit multiple stakeholders: patients gain assurance that AI tools meet safety baselines; healthcare providers reduce legal and clinical risk; developers understand what validation is required before commercialization; and radiologists can confidently integrate AI into their diagnostic workflows. Without such clarity, the uncertainty may paradoxically slow the adoption of AI tools that could improve diagnostic accuracy and patient outcomes.