What actually happened
In a letter dated September 1, 2026, and reported by AuntMinnie, the FDA cited O.N. Diagnostics over an uncleared update to VirtuOst VFA, its vertebral fracture assessment add-on for VirtuOst BCT — a CT-based "virtual stress test" that estimates bone strength and fracture risk using finite element analysis, built on bone-density research that originated at UC Berkeley.
The agency's complaint wasn't about the underlying science. It was about process. According to the letter, version 3.0.0 of VirtuOst VFA shipped with three changes the company had not cleared through a new 510(k): a new machine-learning algorithm for vertebral landmarking, a migration of the software from a desktop application to a web-based architecture, and a migration to a new technology platform. FDA said each of those changes, on its own, could significantly affect the device's safety or effectiveness — the statutory trigger under 21 CFR 807.81(a)(3) for requiring a new premarket submission before distribution.
The remedy FDA demanded was blunt: stop commercial distribution of version 3.0.0 until the issues in the letter are resolved. O.N. Diagnostics told AuntMinnie it is completing verification and validation of a revised version that removes the flagged machine-learning functionality and adds updated cybersecurity controls, alongside broader quality-system corrective actions, and that VirtuOst VFA remains available through the company's existing in-house service while remediation continues.
The rule this update tripped
VirtuOst VFA is regulated as software as a medical device (SaMD) — FDA-cleared software held to the same change-control rules as a physical device. FDA's own guidance on device modifications is explicit on this point: a new 510(k) is required whenever a legally marketed device is changed in a way that could "significantly affect the safety or effectiveness of the device," and manufacturers are expected to run a risk-based assessment of every modification against that standard — not wait for FDA to flag it after the fact. Retraining or replacing an algorithm and moving the software's underlying architecture are exactly the categories FDA's software-modification guidance calls out as presumptively significant, because both can change how the device performs on real patient data in ways a desktop-to-desktop version bump would not.
That is the generic lesson, independent of bone density or vertebral fractures: in regulated imaging software, a model update is not just a code deploy. It's a regulatory event, and the vendor — not the customer — is supposed to catch it before the update ships.
Why this is relevant to radiology AI buyers who've never heard of VirtuOst
VirtuOst isn't a competitor to AI reporting software, and this isn't a story about one company's technology being unsound. It's a story about a vendor's change-management discipline failing in a way that is structurally identical across any AI-enabled imaging product: a company updated the model and the platform underneath a cleared device, and shipped it to customers before FDA agreed the changes were safe. The same gap — pushing a retrained model, a new inference pipeline, or a new hosting architecture without the paperwork that change requires — is available to any vendor selling AI that touches a radiology report.
The table below maps the specific violations in the warning letter to the general question an imaging center should be asking any AI vendor during procurement or renewal.
| What FDA cited in the warning letter | The question it raises for any AI vendor |
|---|---|
| New ML algorithm for vertebral landmarking shipped without a new 510(k). | When your model is retrained or replaced, does the vendor file a new submission, or does it fall under an FDA-authorized predetermined change control plan? |
| Migration from desktop application to web-based architecture. | Has the vendor changed how or where the software runs (on-prem to cloud, new inference engine) since it was cleared — and was that change itself cleared? |
| Migration to a new technology platform. | Can the vendor show the 510(k) number currently in force matches the version actually running in your workflow today? |
| FDA ordered a stop to commercial distribution of the uncleared version. | If a vendor's clearance were challenged, does your contract and workflow let you keep operating safely, or does the report pipeline stop with it? |
None of this requires assuming bad intent. FDA's letter doesn't allege the updated algorithm performed worse — only that the company didn't get the required clearance before distributing it. That's precisely why change control matters as a due-diligence category of its own, separate from whether a product's published accuracy numbers look good: a vendor can have a strong clinical validation study behind its original clearance and still ship an unapproved update two versions later.
Where this fits with FDA's broader direction on AI updates
This case lands at the same moment FDA has told industry that managing the AI device lifecycle and predetermined change control plans are top guidance priorities for FY2027 — the mechanism that lets a vendor pre-clear how it will update a model without filing a new 510(k) every time. The O.N. Diagnostics letter is a real-world illustration of what happens when a vendor updates a cleared algorithm without that kind of pre-authorized plan, or without filing a new submission at all. For an imaging center, the practical takeaway isn't about this one bone-density tool — it's a reason to add vendor change-control history and current clearance status to the same evaluation process already covering accuracy, security, and postmarket error reporting. A foundation-model approach to CT reporting is reviewed by xAID's in-house radiologist on every study and delivered ready-to-sign, but the regulatory lesson here is vendor-agnostic: ask any AI vendor, including xAID, to show the clearance tied to the version actually running in production.
Frequently asked questions
What did the FDA warning letter to O.N. Diagnostics say?
In a letter dated September 1, 2026, the FDA told O.N. Diagnostics that it had distributed version 3.0.0 of VirtuOst VFA, its CT-based vertebral fracture assessment software, without clearing the update first. The agency said three changes — a new machine-learning algorithm for vertebral landmarking, a migration from a desktop application to a web-based architecture, and a technology platform migration — could each significantly affect the device's safety or effectiveness and therefore required new 510(k) submissions. FDA asked the company to stop commercial distribution of version 3.0.0 until the issues are addressed.
What is software as a medical device (SaMD) change control?
Under FDA regulation 21 CFR 807.81(a)(3), a cleared device's manufacturer must submit a new 510(k) before distributing a modification that could significantly affect the device's safety or effectiveness, or that changes its intended use. SaMD change control is the internal process a vendor uses to evaluate every software update — including new or retrained machine-learning models and architecture migrations — against that standard before shipping it, rather than treating updates as routine IT releases.
Why does a bone-density device warning letter matter to radiology AI buyers?
The product isn't a radiology-reporting AI, but the violation is generic to any AI-enabled imaging software: a vendor deployed a materially changed algorithm and platform without the regulatory clearance the change required. The same failure mode — pushing a model update without going back to FDA — applies to any AI vendor that reports CT, MRI, or other imaging studies. It is a concrete argument for asking vendors how they document and clear changes to their AI models, not a theoretical compliance question.
How can imaging centers vet an AI vendor's change-control practices?
Ask for the device's current 510(k) number and confirm the version in production matches what was cleared; ask how the vendor handles model updates (do they trigger a new submission or fall under an FDA-authorized predetermined change control plan); and ask for the vendor's quality system and change-log history, including any prior FDA correspondence. A vendor that cannot answer these clearly is asking a buyer to accept undocumented regulatory risk.
Source: AuntMinnie, "O.N. Diagnostics hit with FDA warning letter" (October 7, 2026); FDA, "Is a New 510(k) Required for a Modification to the Device?"; O.N. Diagnostics, VirtuOst BCT product overview. Figures are rounded as reported.