Regulating radiology AI medical devices that evolve in their lifecycle

TL;DR

The paper addresses how radiology AI devices that evolve after deployment face distribution drift and regulatory hurdles that impede timely updates. It surveys the US FDA PCCP mechanism and the EU AI Act, arguing that lifecycle regulatory controls can enable safe, monitored updates without full re-approval. It proposes building an infrastructure that includes structured reporting, continual-learning methods, uncertainty estimation, and radiologist-in-the-loop quality assurance to support post-market learning. The authors highlight remaining gaps in evaluation metrics, data governance, and cross-population safety, urging standards and third-party assessments to ensure trustworthy deployment.

Abstract

Over time, the distribution of medical image data drifts due to factors such as shifts in patient demographics, acquisition devices, and disease manifestations. While human radiologists can adjust their expertise to accommodate such variations, deep learning models cannot. In fact, such models are highly susceptible to even slight variations in image characteristics. Consequently, manufacturers must conduct regular updates to ensure that they remain safe and effective. Performing such updates in the United States and European Union required, until recently, obtaining re-approval. Given the time and financial burdens associated with these processes, updates were infrequent, and obsolete systems remained in operation for too long. During 2024, several regulatory developments promised to streamline the safe rollout of model updates: The European Artificial Intelligence Act came into effect last August, and the Food and Drug Administration (FDA) issued final marketing submission recommendations for a Predetermined Change Control Plan (PCCP) in December. We provide an overview of these developments and outline the key building blocks necessary for successfully deploying dynamic systems. At the heart of these regulations - and as prerequisites for manufacturers to conduct model updates without re-approval - are clear descriptions of data collection and re-training processes, coupled with robust real-world quality monitoring mechanisms.

Figures (4)

• Figure 1: From bottom to top: Requirements set by recent regulatory efforts in EU and USA, building blocks for effectively designing dynamic systems, and potential benefits in the form of risk reduction, performance increase, and sustainability.
• Figure 2: Approval process for a deterministic locked AI system (top) vs. a lifelong learning system (bottom). In the second case, a description of the planned modifications and protocols outlining how they will be implemented and monitored are submitted as additional documentation during initial clearance. Modifications following these specifications, such as re-training the model with additional training data, do not require re-approval.
• Figure 3: Continuously learning systems adapt to new environments, obtaining better performance on local data distributions. By leveraging more data overall, global performance may also increase on yet-unseen or older domains. However, it is also possible that the system forgets how to cope with earlier data or that failures and/or biases are introduced into the model.
• Figure 4: Advantages and disadvantages of locked vs. continual learning systems with respect to resource utilization and model performance.