Computer vision systems used in clinical settings -- radiology AI that assists with image interpretation, pathology algorithms that analyse histology slides, surgical guidance systems that interpret intraoperative video -- are Software as a Medical Device and require FDA clearance before deployment in US clinical settings. By 2024, the FDA had cleared over 950 AI/ML-based medical devices, the majority of which are radiology and imaging applications. The cleared devices represent the leading edge of clinical AI deployment, and the pathway they followed is well-documented enough to plan against.
510(k) Clearance for Computer Vision SaMD
The 510(k) pathway is available when a predicate device exists -- an already-cleared device with the same intended use and similar technological characteristics. For computer vision applications in radiology, the predicate landscape is now well-populated: most modality-specific AI applications have predicates that enable 510(k) submissions. The 510(k) requires demonstrating substantial equivalence to the predicate -- that the new device is at least as safe and effective as the predicate for the intended use.
For AI/ML-based SaMD, substantial equivalence demonstration requires providing performance data on a test dataset that is independent of the training data, comparison of performance against the predicate or against physician performance as a benchmark, and technical documentation of the algorithm architecture, training methodology, and validation protocol. The FDA's guidance on clinical performance assessment for AI/ML-based SaMD provides specific guidance on dataset requirements.
De Novo for Novel Computer Vision Applications
When no predicate exists -- for a novel clinical application or a novel imaging modality -- the De Novo pathway provides a route to classification and clearance without a predicate. De Novo results in the FDA establishing a new device type with special controls that subsequent similar devices can use as a predicate. The De Novo pathway is significantly more resource-intensive than 510(k) and requires demonstrating both safety and effectiveness, not just substantial equivalence.
Locked vs Adaptive Algorithms and the PCCP
The most consequential distinction for a computer vision SaMD development team is whether the algorithm is locked or adaptive. A locked algorithm does not change after deployment without a new regulatory submission. An adaptive algorithm changes -- through retraining, fine-tuning, or architecture modification -- during deployment. Most production AI systems are adaptive. The Predetermined Change Control Plan is the mechanism the FDA accepts for adaptive algorithms: it specifies the algorithm change protocol, the performance monitoring plan, and the methodology for determining when a change is within the approved scope versus when it requires a new submission.
A cleared PCCP gives the developer a framework for updating the algorithm without returning to FDA for each update, provided the updates stay within the PCCP boundaries. Most AI teams discover the PCCP constraint after their model has already improved in production -- at which point they must either seek a new clearance for the changed algorithm or roll back to the cleared version.
Clinical Validation Dataset Requirements
The validation dataset for a computer vision SaMD submission must satisfy several requirements distinct from the requirements of a general ML benchmark. The dataset must be representative of the intended use population -- the full range of patients, scanner types, acquisition protocols, and image quality levels the device will encounter in clinical practice. A dataset consisting entirely of high-quality images from a single academic medical centre is not representative of the diverse clinical environment. FDA reviewers scrutinise representativeness and will request additional data if the submitted dataset does not adequately cover the intended use population.
Post-Market Surveillance Obligations
FDA clearance of an AI SaMD is not the end of the regulatory engagement. Post-market surveillance obligations require monitoring the device's performance in clinical use, reporting adverse events through MedWatch, and submitting periodic performance reports. For adaptive algorithms with a PCCP, the post-market performance monitoring plan specified in the PCCP is an ongoing obligation. Model performance in production on diverse patient populations may differ from validation performance, and those differences must be monitored and reported.
The engineering behind this article is available as a service.
We have done this work — not advised on it, not reviewed documentation about it. If the problem in this article is your problem, the first call is with a senior engineer who has solved it.