What Is a Radiology Interpretation Worth? Measuring Value in the AI Era
RADIOLOGY IN THE AGE OF AI & VLMS | ARTICLE 14 OF 14
If AI can generate a report, what is a radiologist’s signature worth? The answer depends on what the signature is certifying, and that question has never been more important to answer clearly.
The Old Model and Why It Is Under Pressure
For most of the past three decades, the economic logic of radiology was simple and stable: read more, earn more. Volume drove revenue. Speed was the premium skill. The RVU system rewarded throughput, and practices that optimized their workflows around throughput did well. That model is not gone. But AI is creating enough pressure on the throughput premise that every radiologist who has not yet thought seriously about where the value in their work lives should start now.
What the Signature Is Really Certifying
The wrong framing is replacement. A foundation model that screens a non-contrast CT simultaneously for breast and lung cancer findings, matched to mammography-level performance for breast lesions, does not make the radiologist irrelevant. It makes the radiologist’s judgment about what the AI found, how confident it is, whether its flags are contextually appropriate for this patient, and whether it missed something that a trained eye would catch the specific thing that earns the signature.¹ A brain MRI that simultaneously predicts IDH mutation status, estimates dementia risk, characterizes brain age, and marks potential stroke timing creates an enormous new volume of findings to be interpreted, prioritized, communicated, and acted upon.² The radiologist who interprets a study that dense is not doing less work. The work is different. The question is how to value it.
The Reimbursement Pressure Is Not Going Away
The RVU model does not currently have a good answer to that question. An AI-assisted read of a chest CT may take less clock time than an unassisted one, and if payers respond to that productivity gain the way they have historically responded to every productivity gain in radiology, they will use it as justification for further reimbursement compression. It is the pattern of the last decade, and there is no reason to expect it to change on its own. What can change is the argument radiologists make about what the signature is worth. The case is for ours to make.
Outcomes Data Is the Only Evidence That Matters
The evidence that matters is outcomes data. Not AUC. Not reader study performance on a curated validation set. The question that determines whether an AI-assisted interpretation added value is whether the patient did better. Was the right downstream action taken? Was the incidental finding followed up? Was the subtle opacity on the CT that the AI flagged and the radiologist confirmed actually the beginning of something that needed to be caught? These are Stage 4 questions in the epistemology this series has been building toward since Article 12, and no vendor and no existing monitoring infrastructure answers them systematically.
What the Market Is Already Pricing
The practices that are positioned to answer those questions will be worth something specific and demonstrable that others will not. This is not an abstract observation. RadNet has allocated over $340 million in acquisitions in early 2026, targeting imaging centers and AI companies simultaneously, with a stated goal of approximately $140 million in AI annual recurring revenue.³ The valuation logic embedded in that acquisition pace is not primarily about read volume. It is about what happens when imaging network scale and AI performance infrastructure sit on the same platform: the ability to generate outcome data at a level no single-site practice can match. The practices being acquired are not the ones with the highest throughput. They are the ones with the monitoring infrastructure to prove what their AI is doing.
Three Stakeholders, One Infrastructure
The three-stakeholder structure this series has returned to across Articles 6, 12, and 13 resolves cleanly into an economic argument here. For the individual radiologist, the value question is about trust: does this AI output warrant my endorsement for this patient in this clinical context, and can I document why? For practice leadership, the question is operational: is this tool performing at the level we contracted for, and can we quantify the value it is or is not delivering? For compliance, the question is audit readiness: if a liability event occurs, can we show what the AI produced, what the radiologist did with it, and whether the system was functioning within its validated parameters at that moment? The same monitoring infrastructure answers all three questions at different levels of aggregation. That infrastructure is not a vendor’s responsibility. Vendors have every incentive to define value in terms of what their benchmarks measure, which is agreement with a reference standard, not downstream patient outcomes. Closing that gap is clinical work, and it belongs to physicians.
Building the Closed-Loop Framework
The closed-loop outcomes framework this series has introduced as a forward concept is not a theoretical ideal. The first practical implementations are already identifiable: MSK MRI reports correlated with subsequent surgical decisions, lung nodule findings correlated with biopsy confirmation or progression on follow-up CT, chest AI flags correlated with clinical action rates at the referring team level. There is an important caveat that deserves to be stated directly: downstream clinical events are a signal, not a gold standard. Surgeons operate on ambiguous findings. Biopsy rates reflect clinical culture as much as finding acuity. Interpreting outcome data requires the same contextual judgment that interpreting imaging does. But over time and at scale, this is the only framework that can move the value argument beyond the performance metrics that vendors control.
The Radiologist Who Can Answer the Hard Questions
The radiologist who can walk into a meeting with practice leadership and say, “Here is what our AI is producing, here is how those outputs are correlating with downstream clinical events, and here is where the performance is drifting from what we contracted for” is not describing a compliance function. That radiologist is describing exactly the Doctor’s Doctor role this series identified in Article 13: the clinician who translates AI performance into clinical accountability, who makes the AI usable to the people responsible for acting on it. That role is not being filled by vendors. It is not being filled by IT. It requires a physician who understands both what the model is doing and what the clinical consequences of its errors look like.
The Trust Band as an Economic Argument
The three-part competency frame introduced earlier in this series as a thought experiment is the operational answer to the value question. A ‘Fellow Level’ trust band on an AI tool means something you can defend: this model’s outputs, in this practice’s case mix, over this time window, have performed at a level that justifies the reading workflow we have built around it. A ‘Senior Resident Level’ band means something different: this model assists but requires active verification, and the documentation trail should reflect that. A ‘Junior Resident Level’ designation means the tool is new to this environment, or performing outside its validated range, and the oversight posture needs to match. None of this framing is typically available from a vendor’s installation report. It comes from building the measurement infrastructure to know what the tool is doing in your specific clinical environment.
The Durable Competitive Asset
The economic case for building that infrastructure is straightforward. A practice that can document AI-assisted judgment with a performance record that spans the four stages of the epistemology: agreement, correctness, decision-level impact, and outcomes, is the practice whose signature means something defensible and specific. In a market where payers are already looking for reasons to compress per-study reimbursement and plaintiff attorneys are already developing the precedents for AI-assisted malpractice claims, that documentation infrastructure is not overhead. It is the practice’s most durable competitive asset.
This is the final article in a series but a few more topics have surfaced and I will do at least two more. One is leveraging the checklist mindset and the other is why I feel this is the best time in history to be a radiologist.
For those who know me, I am taking this series and your feedback and creating a short book, which will be available soon on Amazon.
Most clinical AI systems are evaluated before deployment and assumed to perform the same in production. In reality, performance shifts across sites, scanners, populations, and workflows, and those shifts are rarely measured systematically.
If you are building or deploying AI in radiology, including VLM-based reporting or multi-model orchestration systems, you need a way to monitor real-world behavior continuously. This includes tracking disagreement with clinicians, identifying drift, and understanding failure modes over time.
Veriloop provides a vendor-agnostic observability layer for clinical AI. We sit downstream of your model and workflow, measuring performance where it matters: in production, across real cases, with real users.
This is not model evaluation. This is system monitoring.
Contact:
References
¹ Qian X et al.. “A foundation model for breast and lung cancer screening using non-contrast computed tomography.” Nature Health, February 5, 2026. https://www.nature.com/articles/s44360-026-00055-8
² Kann BH, Tak D et al.. “A generalizable foundation model for analysis of human brain MRI.” Nature Neuroscience, February 5, 2026. https://www.nature.com/articles/s41593-026-02202-6
³ RadNet M&A Strategy. “RadNet Has Allocated Over $340M in Acquisitions Already in 2026.” Radiology Business, March 2026.https://radiologybusiness.com/topics/healthcare-management/mergers-and-acquisitions/radnet-has-allocated-over-340m-acquisitions-already-2026-leaders-discuss-why-and-whats-next