Competitive intelligence from clearances and approvals, not slide decks

Predicates cited by recent AI/ML-enabled radiology 510(k)s and FDA's treatment of technological differences

Methodology

I searched the FDA 510(k)/De Novo database for AI/ML- and deep-learning-enabled radiology and medical-imaging devices cleared from January 2024 onward, using lexically diverse keyword sets (deep learning image analysis, convolutional neural networks, computer-assisted triage/notification, MRI/PET reconstruction and denoising, CADe/CADx detection, and AI quantification). That discovery pass returned 58 distinct cleared devices. From those I drilled a diverse, recent subset of 21 clearances spanning the four dominant AI/ML radiology archetypes: computer-aided triage/notification (product code QAS), computer-aided detection/CADx (QFM, MYN, QBS), image reconstruction/enhancement (LNH, LLZ), and quantitative post-processing (QIH). For each I read the 510(k) summary to extract the cited predicate(s) and reference device(s), the technological differences FDA identified, and the rationale and testing FDA relied on to conclude those differences did not raise new or different questions of safety and effectiveness.

What predicates are being cited

The clearest pattern: recent AI/ML radiology devices overwhelmingly cite another AI/ML device of the same functional class as the primary predicate, frequently an earlier version of the same product line, plus one or more AI reference devices to bridge indication or feature gaps. FDA's substantial-equivalence logic for this class is now well-worn: same principles of operation, same DICOM inputs, same output (worklist prioritization / notification or an enhanced image), same parallel-to-standard-of-care workflow, with differences confined to the specific finding, algorithm architecture, or performance level.

Computer-aided triage and notification (QAS / QFM)

Image reconstruction and enhancement (LNH / LLZ)

Quantitative post-processing (QIH)

How FDA treated the technological differences

Five recurring treatments emerge across the subset.

1. "The difference is only the finding or the algorithm, not the principle of operation." For triage/notification software, FDA repeatedly held that a different target finding (LVO vs. ICH, aortic dissection vs. pneumothorax) or a different/improved AI model is a technological difference that does not raise new questions of safety and effectiveness, because the subject and predicate share intended use, DICOM inputs, worklist-prioritization output, and operate in parallel to standard of care without altering the original image or removing/de-prioritizing cases. This is the explicit logic in syngo.CT LVO Detection (only the clinical finding differs from the hemorrhage predicate) ¹⁴⁰¹⁴¹, BriefCase CARE Multi-Triage (fine-tuned modules from a "locked foundation model," same triage function) ¹³⁶, and both Annalise Enterprise clearances, where FDA localized the differences to "set of findings and algorithm" and "AI models/performance of algorithm" ^109110115.

2. Performance shortfalls handled through benefit-risk framing rather than rejection. In Annalise Enterprise (K253818) FDA acknowledged NCCT is a less advanced modality than CTA, so infarct-detection performance could be expected to be poorer than CTA-based approaches, and still cleared it, citing same principles of operation, subgroup analyses across vascular territories, and a benefit-risk discussion showing the device performs as intended ¹¹⁴¹¹⁵.

3. Expanded indication supported by a targeted standalone study. Rayvolve broadened the population from adults to pediatric patients aged 2+; FDA accepted overlapping AUC confidence intervals between the pediatric and adult (predicate) datasets as evidence of statistically non-inferior performance ¹³³¹³⁴. Brainomix 360 Triage Stroke added a subarachnoid-hemorrhage indication and a re-architected ICH deep-learning model, and FDA held that offering three configurable workflows in one device and improving the algorithm did not raise different questions, supported by clinical validation across ICH subtypes ^127128129.

4. Reconstruction/enhancement devices: image quality is "the primary question of safety and effectiveness," proven on the bench. For the same-family upgrades (AIR Recon DL adding DL Phase Correction, Sonic DL adding a 3D CNN and extending to 7.0T), FDA relied on nonclinical phantom and in-vivo bench testing (PSNR, RMSE, SSIM, ADC accuracy, low-contrast detectability), supplemented for Sonic DL by Likert-score reader studies, concluding no new hazards and adequate diagnostic image quality ^117119120144.

5. Quantification devices: "deterministic derivatives" and no-new-diagnostic-output arguments. AI4CMR v2.0 added 2D flow outputs that FDA accepted as deterministic mathematical derivatives of already-validated core metrics, so clinical validation was not required (multi-vendor agreement testing and Dice-coefficient segmentation validation sufficed) ¹³⁵. NeuroQuant added T2*/SWI as a data source but kept the same volumetric-quantification intended use, so V&V plus clinical validation of the reports/overlays supported SE ¹¹².

Evidence used to close the gap

• Standalone algorithm performance (retrospective, blinded, multi-reader ground truth) is the workhorse for triage/CADe devices. Examples: BriefCase mCTA VO study, sensitivity 81.4% / specificity 85.5% against a 2+1 majority-vote reader panel, with time-to-notification of 1.09 min vs. 2.23 min for the predicate ¹⁵¹¹⁵²; CogNet AI-MT+ DBT study, AUROC 0.9548 on 806 women ¹³⁹.
• Time-to-notification benchmarking against the predicate is now a standard triage endpoint (Annalise 81.6 s ¹¹⁴; BriefCase 1.09 min ¹⁵¹).
• Nonclinical bench + limited reader studies dominate reconstruction/enhancement clearances (AIR Recon DL, Sonic DL) ¹²⁰¹⁴⁴.
• Software V&V under IEC 62304 / ISO 14971 / IEC 62366-1, with clinical testing waived where the predicate itself was not clinically validated (AI-Rad Companion Brain MR explicitly states no clinical tests because predicate K232305 was not clinically validated) ¹⁴²¹⁴³.
• Predetermined Change Control Plans (PCCPs) appear in the review record for at least one device (a2z-Unified-Triage) ¹⁵⁹, consistent with FDA's current approach to AI/ML lifecycle management.

Bottom line

Recent AI/ML radiology 510(k)s cite a same-class (often same-family) AI predicate plus reference devices, localize the differences to finding, algorithm, or feature, and close the gap with standalone performance and bench testing. FDA's reasoning is now formulaic: shared intended use, inputs, output, and workflow mean the difference is "technological only" and raises no new questions of safety and effectiveness.