Regulatory Pathways for AI in Nuclear Medicine: FDA 510(k) and Beyond

The regulatory landscape for AI-based medical imaging software has changed more in the past five years than in the preceding two decades. FDA's Digital Health Center of Excellence, established in 2020, has accelerated clearance review timelines and provided clearer guidance on clinical evidence standards. The 2021 AI/ML-Based Software as a Medical Device (SaMD) action plan introduced the concept of predetermined change control plans - allowing cleared AI devices to update algorithms within defined parameters without requiring a new 510(k) submission. For nuclear medicine AI developers, understanding these frameworks is prerequisite to building a viable regulatory strategy.

What Requires FDA Clearance vs. What Does Not

The threshold for FDA regulation of nuclear medicine software is function: does the software make, or assist in making, a clinical determination that informs patient treatment decisions? Software that meets this threshold is a Software as a Medical Device (SaMD) and requires either FDA clearance (510(k) or De Novo) or PMA approval before commercial distribution in the United States.

The distinction matters practically. Image quality enhancement software that improves signal-to-noise ratio in PET images without making any clinical claims - not claiming to improve lesion detection, not calculating SUV values for clinical interpretation - may be positioned as general-purpose image processing software outside the SaMD definition. Once the software claims to assist lesion detection, calculate quantitative metrics used in clinical decision-making, or generate preliminary diagnostic reports, it falls squarely within the SaMD framework and requires regulatory clearance.

Nuclear medicine dosimetry calculation software that is used to determine patient treatment activities is a medical device under 21 CFR 880.6310 (software for clinical decision support). Automated dosimetry tools used in radiopharmaceutical therapy planning are subject to FDA oversight when they influence administered activity decisions. This is a regulatory position that has sometimes surprised hospital physicists who assumed internally-developed calculation tools were exempt - they are not, once distributed beyond the developing institution.

The 510(k) Pathway: Substantial Equivalence in Practice

510(k) clearance requires demonstrating substantial equivalence to a legally marketed predicate device. The predicate does not need to be AI-based - it needs to have the same intended use and the same or equivalent technological characteristics. For nuclear medicine image enhancement software, appropriate predicates include previously cleared image processing software for nuclear medicine (several exist, cleared under product code IYO and LLZ), volumetric segmentation tools, and computer-aided detection software.

The critical challenge in nuclear medicine AI 510(k) submissions is clinical evidence. FDA expects analytical performance data (validation against labeled datasets demonstrating algorithm accuracy) and clinical performance data (demonstrating that the algorithm performs as intended in a patient population representative of the intended use). For lesion detection software, this means reader studies comparing physician diagnostic performance with and without AI assistance, on an appropriate test set covering the range of indications in the intended use description.

Sample size requirements for reader studies depend on the design. A non-inferiority study comparing AI-assisted performance against unassisted performance on a well-powered dataset (typically 200-400 studies per indication, with representation of positive and negative cases and disease severity strata) is the standard design. The endpoint must be pre-specified as either sensitivity/specificity at a fixed operating point, or AUC from a full ROC analysis. Showing statistically equivalent or superior diagnostic performance to unassisted reading, in a blinded study design, is the evidentiary standard FDA uses to evaluate substantial equivalence.

De Novo When No Predicate Exists

When a nuclear medicine AI software performs a function without a cleared predicate - either because the function is genuinely novel or because available predicates differ in technological characteristics in ways that create unacceptable safety concerns - the De Novo pathway provides a route to clearance for Class II devices. De Novo submissions include clinical evidence requirements similar to 510(k) but additionally require proposed special controls that FDA can use to establish a regulatory pathway for subsequent similar devices.

For nuclear medicine, automated dosimetry for radiopharmaceutical therapy is an area where the predicate landscape is thin. Several cleared dosimetry calculation tools exist, but AI-based dosimetry incorporating automated organ segmentation from SPECT/CT represents a technological extension that may not map cleanly to existing predicates. The De Novo pathway, while slower and more resource-intensive than 510(k), is appropriate when the predicate strategy is genuinely unclear - attempting a 510(k) with weak predicate support risks rejection on substantial equivalence grounds, a more costly outcome than choosing De Novo from the outset.

The Predetermined Change Control Plan

The most significant recent regulatory development for nuclear medicine AI is FDA's guidance on predetermined change control plans (PCCP). A PCCP is a document submitted with the initial 510(k) that describes the types of algorithm modifications the developer anticipates making post-clearance and the validation methodology that will be used to ensure those modifications do not reduce device performance below the cleared specification.

In practice, a PCCP for nuclear medicine AI might specify that the detection model may be retrained with additional training data from new scanner types, provided that validation on a held-out test set demonstrates maintained sensitivity and specificity within specified confidence intervals. Modifications that fall within the PCCP scope can be implemented without new 510(k) submission - reducing the regulatory burden of iterative model improvement, which is the normal development pattern for AI tools.

The PCCP must be clinically and statistically rigorous. FDA reviews the PCCP as part of the initial 510(k) and may request modifications before clearance. A well-crafted PCCP that defines modification categories, validation criteria, and testing datasets clearly can significantly reduce the post-market regulatory overhead for algorithm updates that would otherwise trigger a 510(k) supplement or new submission.

IEC 62304 and Software Lifecycle Documentation

FDA expects AI medical device software to be developed under a recognized software development lifecycle standard. IEC 62304 is the international standard for medical device software lifecycle processes, and FDA considers it a recognized standard that creates a rebuttable presumption of compliance with quality system requirements. For nuclear medicine AI developers, implementing IEC 62304 processes - software architecture documentation, unit and integration testing, software problem resolution, and configuration management - is not optional if FDA clearance is the goal.

The documentation burden of IEC 62304 compliance is substantial for a startup team more accustomed to rapid iteration than to documented test protocols and formal change control. The regulatory strategy question is therefore whether to establish IEC 62304-compliant processes from the beginning of software development - which slows initial development speed - or to adopt those processes after a product is more mature - which requires retrospective documentation that FDA reviewers view with skepticism.

The practical recommendation from regulatory consultants who have worked through multiple nuclear medicine AI clearances is to establish IEC 62304-compliant development processes from the first clinical version of the software, before any human subjects data is collected for validation. Retrofitting process documentation to software that was developed without it is more expensive and time-consuming than building the processes correctly from the start.

Post-Market Surveillance Under PCCP

AI devices cleared with PCCPs carry post-market surveillance obligations that conventional medical device software does not. The developer must monitor algorithm performance in deployed use - not just test data - and report significant performance deviations to FDA. This requires building surveillance infrastructure: real-world performance data collection from deployed sites, statistical monitoring for performance drift relative to cleared specifications, and an escalation pathway for reporting anomalies.

As we note in our article on nuclear medicine AI workflow integration, the practical logistics of post-market surveillance require deployed sites to participate in data sharing under appropriate data use agreements. This is a component of the commercial contract that is sometimes overlooked at deployment - creating a situation where the regulatory obligation exists but the data needed to fulfill it is not available. Addressing surveillance data collection as a deployment requirement, not an afterthought, is the appropriate operational approach.