Trust, Governance, and the Closed Loop: Leading the AI-Integrated Imaging Enterprise
By Professor & Dr. Anthony Noujaim Endowed Chair of Oncology. Director, Division of Oncologic Imaging and Radionuclide Therapy, Faculty of Medicine & Dentistry, University of Alberta
Parts 1 and 2 traced the arc from rule-based detection to foundation models, and from isolated tools to integrated imaging workflows. Part 3 asks the harder question: what does it take to lead an organization through that transition, and what does the landscape look like for those who get it right — and those who do not?
There is a version of this story that writes itself: AI arrives, radiology transforms, costs fall, outcomes improve, and the organizations that moved early reap the rewards. That version is not wrong — but it is incomplete in ways that matter to anyone responsible for making it happen rather than simply predicting it. The technology is real. The clinical evidence is accumulating. The commercial momentum, as the RadNet-Gleamer transaction made plain in early 2026, is now structural rather than speculative. What remains genuinely hard is not acquiring the tools. It is building the organizational conditions in which they deliver what the vendor deck promises.
The Governance Gap No One Advertises
FDA clearance is a threshold, not a guarantee. When applied to a community hospital scanner fleet, an older patient population, or an ethnic group that was underrepresented in the training data, a technology that was approved on a dataset derived from major academic medical centers in North America may exhibit subtle discrepancy. This is not a theoretical risk — it is a documented pattern in the published literature, and it is the reason that the most rigorous health systems treat AI deployment as a clinical program rather than a software installation. [1]
The underperformance is rarely dramatic. It does not announce itself. It accumulates silently in the subpopulations where the training data was thinnest — and those subpopulations, in radiology’s experience with algorithmic bias, are disproportionately the ones carrying the highest burden of undiagnosed disease. What that means in practice is investment in four things that do not appear in any vendor proposal: baseline performance measurement before go-live, prospective monitoring after it, a structured process for capturing radiologist disagreement with AI outputs, and governance accountability that sits at the clinical leadership level rather than in the IT department. None of these are expensive relative to the cost of the tools themselves. All of them are routinely skipped in the rush to deployment. The organizations that skip them discover, usually eighteen months later, that their AI tools are performing adequately on average and poorly in precisely the subpopulations where clinical need is highest. [2]
For the CFO, this is not an abstract governance concern. A radiology AI tool that generates a systematic false-negative pattern in a specific patient population is a liability that no clearance letter insulates against. The cost of getting that wrong — in patient harm, in litigation, in reputational damage to a referring physician network built over decades — dwarfs the cost of building the monitoring infrastructure that would have caught it.
The Data Asset Nobody Is Talking About
The most durable competitive advantage in the AI-integrated imaging enterprise is not the algorithm. It is the data. Specifically, it is the longitudinal, well-governed, clinically annotated imaging dataset that a high-volume imaging network accumulates over time and that no external vendor, however well-funded, can replicate by acquisition alone.
RadNet’s strategic logic in acquiring Gleamer is legible precisely here. With over eleven million scans performed annually across its U.S. network, RadNet is not simply buying AI capability — it is accelerating the construction of a proprietary training and validation environment that will compound in value with every scan performed. The AI tools improve because the data improves. The data improves because the volume grows. The volume grows because AI-enabled throughput and turnaround attract referring physicians who value speed and consistency. The loop closes on itself, and the gap between the organizations inside it and those outside it widens each year. [3]
For health system executives outside the outpatient radiology network model, the implication is equally clear: the imaging data your institution generates is a strategic asset, and the question of who owns it, who governs it, and who benefits from its use in AI training is a board-level conversation, not an IT procurement question. Health systems that sign AI vendor agreements without retaining data rights are, in a meaningful sense, subsidizing a competitor’s moat.
The stakes of that conversation have already become concrete in ways that should give every procurement committee pause. In 2024, Australia’s largest radiology provider, I-MED, was found to have transferred de-identified patient images — X-rays, CT scans, the imaging record of tens of thousands of patients — to a commercial AI company for model training, without explicit patient consent, prompting a formal investigation by the national privacy commissioner. The data were not stolen; they were shared under an agreement that most patients, and reportedly many clinicians, were unaware of. This is not an isolated incident — it is a structural feature of an industry in which imaging data has become extraordinarily valuable, and in which vendor agreements are increasingly written to capture it. Scanner manufacturers offering attractive incentives — lifetime software upgrades, all-inclusive configurations, performance guarantees — are not doing so without a commercial logic that extends beyond the hardware sale. Before any such agreement reaches a signature, the question of what happens to the imaging data generated on that equipment, and who retains the right to use it for AI training, deserves the same legal scrutiny as the indemnity clauses and the service-level commitments. The upgrade that comes at the cost of your patients’ imaging data is not, on any honest accounting, free.
The pixels became predictions. The predictions became products. The products are becoming infrastructure. What infrastructure becomes, in the hands of the organizations that govern it well, is an enduring advantage.
The Closed-Loop Imaging Pathway: What It Looks Like at Scale
The integrated imaging enterprise of 2030 does not look like a radiology department with AI tools bolted on. It looks like a clinical workflow in which AI is the connective tissue at every stage, and in which the stages themselves are no longer experienced as separate. [4]
A referring oncologist documents a clinical question. An AI-assisted decision-support layer confirms the imaging modality and protocol appropriate to that question, flags any prior relevant studies, and routes the order directly to the optimal scanner slot given patient location, urgency, and equipment availability. The scanner acquires the study at the lowest dose consistent with diagnostic quality, using deep learning reconstruction to produce a clean image in minutes rather than the processing intervals of the prior generation. The AI triage layer pre-populates structured report elements, cross-references current findings against the patient’s longitudinal imaging record, reads the study as soon as it is finished, and flags time-critical findings for urgent radiologist evaluation. The radiologist reviews, edits, and authenticates — spending the majority of their time on the interpretive and communicative work that requires their training, rather than on the administrative and templating work that does not. The signed report reaches the referring physician’s screen, structured for rapid clinical assimilation, before the patient has left the building. [5]
That is not a speculative scenario. Elements of it are operational today at leading imaging networks. The question for most organizations is not whether this pathway will exist — it will — but whether they will be inside it or outside it when it becomes the standard of care.
The ROI That Compounds
The financial case for the AI-integrated imaging enterprise is not a single number. It is a stack of returns that compound against each other over time, and understanding the stack is what separates a sophisticated capital allocation decision from a technology purchase.
At the departmental level, the immediate returns are operational: more studies completed per radiologist per shift, faster turnaround from acquisition to signed report, fewer repeat scans driven by protocol error or image quality failure, and lower consumable costs per study from deep learning reconstruction. These are measurable within the first operating year and, at meaningful scale, they are material.
At the network level, the medium-term returns are competitive: referring physician retention driven by turnaround time and report quality; market share capture from community providers who cannot match the throughput or the AI-assisted diagnostic consistency of a networked imaging enterprise; and the reduction of the radiologist recruitment premium that currently burdens any organization trying to grow imaging volume faster than the specialty’s training pipeline allows.
At the enterprise level, the long-term return is the data asset itself. A well-governed longitudinal imaging dataset, linked to clinical outcomes, treatment records, and genomic data where available, is the foundation for the next generation of AI development — and for the research partnerships, licensing arrangements, and regulatory submissions that will define the most valuable players in health AI over the next decade. The organizations building that asset today, at the cost of governance infrastructure and data discipline, are making an investment whose returns will not appear on next quarter’s dashboard but will determine who leads this market in 2032.
Three Decisions That Will Define the Upcoming Five Years
Every healthcare executive with imaging in their portfolio faces a version of the same three decisions, and the window for making them well is narrowing.
Own the workflow layer, not just the tools. The RadNet-Gleamer model is the clearest available signal of where durable advantage lies. Owning individual AI applications is a procurement strategy. Owning the integrated workflow layer that connects referral, protocol, acquisition, reconstruction, triage, and reporting — and that generates proprietary data at every step — is a business strategy. The distinction matters enormously when the tools themselves are increasingly commoditized.
Treat data governance as infrastructure, not compliance. The health systems that will lead in AI are not necessarily those with the most sophisticated algorithms today. They are those with the cleanest, best-governed, most longitudinally complete imaging data. That is an organizational capability built over years, not a problem solved by a vendor contract. The time to build it is before you need it to validate the next generation of tools, not after.
Invest in the human layer with the same seriousness as the technology layer. The radiologist who brings tumor board experience, clinical relationships, and interpretive depth to complex oncology cases is not a transitional resource to be managed toward obsolescence. They are the quality anchor of the AI-integrated enterprise, the accountability layer that no regulatory framework will remove, and the clinical intelligence that keeps referring physicians coming back. Workforce strategy in the AI era is not about managing fewer radiologists — it is about deploying better-supported radiologists on the work that justifies their training, and building the organizational culture in which AI is understood as the tool that makes that possible.
The View From 2030
The consolidation that the RadNet-Gleamer transaction signals will continue. The number of standalone AI radiology vendors will shrink as networks acquire capability and scale crowds out point solutions. The regulatory environment will tighten around post-market surveillance and algorithmic transparency, rewarding organizations that built monitoring infrastructure early. The health systems that effectively managed the imaging data asset will have advantages over those that did not, and the asset will be increasingly acknowledged as a balance sheet item.
The technology is not the hard part. It never was. The hard part is building the organizational clarity, the governance discipline, and the long-term investment perspective that turns a collection of AI tools into an integrated clinical and commercial enterprise. The organizations that do that work — unglamorous, incremental, and largely invisible to the market until it suddenly is not — are the ones that will define what radiology looks like for the next generation of patients, clinicians, and investors.
The pixels became predictions. The predictions became products. The products are becoming infrastructure. What infrastructure becomes, in the hands of the organizations that govern it well, is an enduring advantage.
References
- Benjamens S, Dhunnoo P, Mesko B. The state of artificial intelligence-based FDA-approved algorithms in medicine. NPJ Digit Med. 2020;3(1):130.
- Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447–453.
- Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44–56.
- Kahn CE Jr. From images to actions: opportunities for artificial intelligence in radiology. Radiology. 2017;285(3):719–720.
- RadNet Inc. RadNet acquires Gleamer, making DeepHealth the largest provider of radiology clinical AI solutions worldwide [press release]. Los Angeles: RadNet; 2026 Mar 2.
This concludes the three-part series: AI in Oncologic Imaging. Parts 1–3 are available in full at CXOTech MedicalTech Magazine digital archive.