Radiology's AI Turn Is Erasing the Line Between Vendor and Practice

Radiology has long been the proving ground for medical AI. It has structured data, high imaging volumes, and a chronic workforce shortage that makes automation attractive on economic grounds alone. What is changing now, according to reporting from STAT, is the depth of that integration. Some radiology practices are no longer simply buying AI tools off a vendor's shelf. They are building, tuning, and iterating on models internally, in a way that starts to look less like clinical deployment and more like software development.

That distinction matters more than it sounds. When a hospital licenses a diagnostic AI product from an FDA-cleared vendor, the regulatory and liability lines are relatively clear. The vendor owns the model, the practice owns the clinical judgment, and the two responsibilities sit side by side. When a practice starts customizing algorithms, retraining them on its own patient population, or building proprietary tools to triage studies, those lines blur. The practice becomes, in effect, a co-developer of the technology it uses to make diagnoses.

Why the Blurring Matters

Radiology groups have real incentives to move in this direction. Off-the-shelf AI tools are trained on datasets that may not reflect a given practice's patient mix, equipment, or case volume. A model tuned on general population chest CTs may underperform on a rural hospital's older scanner fleet or a specialty center's unusual case distribution. Practices that can adjust models to their own environment stand to gain real accuracy improvements. That is the pitch, and it is not an unreasonable one.

But the operational upside comes with governance costs. FDA clearance pathways were built around the premise of a static, or near-static, product. A cleared algorithm behaves the same way in Ohio as it does in Oregon. Once a practice starts modifying that algorithm post-clearance, even informally, it raises questions about which version of the tool is actually operating on a given patient, who validated that version, and who is accountable if it errs. Radiology groups that once outsourced these questions to vendors are now, in some cases, inheriting them.

This is not a hypothetical concern for an industry already navigating scrutiny elsewhere in health tech. STAT's broader coverage this cycle includes a separate flashpoint: Medicare's push to tighten remote patient monitoring billing rules, which drew formal opposition from UnitedHealth Group, CVS Health, and Kaiser Permanente. The proposal would require RPM services, currently often delivered by third-party vendors working under a billing provider's supervision, to be performed directly by the billing provider's own employees. The specifics differ from radiology's AI question, but the underlying tension is the same. Regulators are trying to figure out where vendor responsibility ends and provider accountability begins, at exactly the moment technology is making that boundary harder to draw.

Key Risks

The clearest risk sits with liability. If a radiology practice modifies a diagnostic algorithm and a missed finding leads to a malpractice claim, the question of whether the original FDA clearance still applies becomes central to the case. Plaintiffs' attorneys will have an obvious target in a hybrid model where the practice itself contributed to the tool's behavior.

There is also a reimbursement dimension. Payers reimburse based on defined clinical services, not on the provenance of the software behind them. A practice that has effectively become a technology developer may find itself in an awkward position when justifying costs or outcomes tied to internally modified tools, particularly if those tools have never gone through a formal clearance update.

Workforce economics cut both ways here. AI-native practices argue they can offset radiologist shortages by extending capacity, letting a smaller clinical staff handle more volume with algorithmic support. That is a genuine efficiency argument. But it also concentrates risk. A practice with deep internal AI development capability is making itself dependent on that capability performing correctly, at scale, across a growing caseload, with less external validation than a purely vendor-supplied model would carry.

None of this suggests the trend will slow. Health systems facing radiologist vacancies and rising imaging volumes have clear incentive to pursue any tool that adds throughput. The direction of travel favors deeper AI integration, not less.

The Opportunity

For vendors, the shift creates a bifurcated market. Some radiology groups will remain straightforward customers, buying cleared products and using them as intended. Others will demand more flexible, customizable platforms that let internal teams adjust model behavior, essentially asking vendors to sell development infrastructure rather than finished software. Companies that can serve both segments, without compromising the regulatory clarity that cleared products depend on, have an edge.

There is also an advisory opportunity building around governance. As practices take on more technology-development responsibility, they will need frameworks for validating internally modified models, documenting changes, and managing the liability exposure that comes with it. That is a service gap, and one that neither traditional AI vendors nor traditional malpractice insurers are fully equipped to fill yet.

Regulators, for their part, will likely need to revisit how clearance frameworks handle post-market modification, particularly as more clinical practices start behaving like software shops. The FDA has already wrestled with adaptive algorithms that learn continuously. AI-native radiology practices doing manual customization on top of cleared tools present a related but distinct version of the same problem, and one that current guidance does not cleanly address.

What Investors Should Watch

Investors with exposure to diagnostic AI vendors should watch how quickly practices move from buyers to co-developers, since that shift changes the vendor's role from software supplier to infrastructure and governance partner, a higher-margin but higher-liability position. Watch, too, for signs of regulatory response: any FDA movement on post-clearance modification rules would directly affect how much customization practices can safely pursue. Finally, the Medicare RPM dispute is worth tracking as a proxy fight. How CMS ultimately resolves the vendor-versus-provider accountability question in remote monitoring will likely set precedent language that radiology, and other AI-heavy specialties, will eventually have to contend with as well.