The Best Time in History to Be a Radiologist
Radiologists are not being replaced. We are becoming radiologists with better tools, deployed one trusted workflow at a time.
There is a quote that has been circulating in radiology circles for several years now. You have probably heard it: AI will not replace radiologists, but radiologists who use AI will replace those who don’t. The more time I spend thinking about how this transition is playing out across different practice environments, the more I think this framing is wrong.
Replacement implies a binary. A before and after. A moment when the radiologist who adapted wins and the one who didn’t loses. That is not what I am seeing, and it is not what the evidence supports.
What I am seeing is something slower, messier, and considerably more interesting.
Start with the scope of what radiology actually does. Charles Kahn and colleagues have spent years building and refining a formal ontology of radiology’s differential diagnosis space, a structured documentation of the thousands of conditions, anatomic relationships, imaging patterns, and causal chains that constitute what a radiologist looks for across every modality and subspecialty.¹ That documented finding space is enormous. Current AI addresses a small fraction of it with sufficient confidence to be clinically useful.
The earliest deployments cluster where value is clearest and risk is most manageable: extremity radiographs in outpatient orthopedic settings, lung cancer screening CTs at academic centers, neuro-oncology follow-up MRIs where the prior is known and the question is well-defined. These are not the same clinical environment. They are not the same patient population, the same vendor, or the same modality. They were not adopted at the same time or in the same sequence.
That heterogeneity is not a problem to be solved. It is the actual shape of how this transition works.
Over the past several articles, I have written about what can go wrong: the ways standard benchmarks fail to capture real-world performance,² how post-deployment drift goes unmeasured because the infrastructure to measure it does not exist,³ the liability exposure that builds quietly in practices without audit trail documentation,⁴ the correlated failure modes that compound when AI misses something a radiologist would also have been primed to miss.⁵
That material is not pessimistic. It is a prerequisite for the optimism that follows. You cannot build the infrastructure to supervise these tools if you do not understand what they are doing, where they are fragile, and what questions to ask. The infrastructure is the point.
Consider what the progression looks like from inside a practice. A community teleradiology group and an academic subspecialty center are not adopting AI on the same timeline or for the same use cases. A VA system and a private outpatient chain have different formularies, different patient demographics, and different operational constraints.
Progress moves in surges rather than a straight line: rapid expansion into new tasks, then consolidation as tools hit harder territory, then renewed capability as the systems mature.
Some of those expansions will work well.
Some will need to be walked back.
There will be cycles. That is not a failure of AI. That is how complex tools get adopted into complex systems.
What changes across all of these environments is not whether the radiologist is present. It is what the radiologist is doing. At ECR 2026, Harvey and Snoeckx both framed the radiologist’s evolving function around exactly this point: not the generation of a report, but accountability for the judgment applied to what the AI generated.⁶ That shift has structural consequences.
The radiologist whose practice has a monitoring protocol, a documented performance record for its deployed tools, and an audit trail that holds up under scrutiny is practicing differently from one who does not.
The difference is not visible in throughput numbers. It is visible when something goes wrong, and in how the practice is valued when it comes time to negotiate, partner, or sell.
Residency is part of this too. The next generation of radiologists will train with these tools in the workflow. That is a feature, not a concern. Teaching a resident to use AI well, to understand where a particular model has been validated and where it has not, to know when to trust the output and when to override it, is training a radiologist for the environment they will practice in for the next thirty years. Those skills are harder to teach without infrastructure, which is another reason to build the infrastructure now rather than waiting for a curriculum update to mandate it.
Radiology has absorbed categorically different tools before and come out with greater capability each time. The transition from film to PACS eliminated the physical constraints that limited how many patients a radiologist could serve and how quickly. Voice recognition was resisted and then became something no one could imagine working without. Automated impression generation arrived quietly and is now ordinary. Each of those transitions felt, in the moment, like a threat to something essential. Each turned out to be an expansion of what was possible.
The current transition is larger than any of those. The tools are more capable, the finding space they are beginning to address is vast enough that the progression curve will play out for decades, and the radiologist’s comparative advantage, judgment applied to AI output with full accountability for that judgment, is more valuable now than at any prior point in the history of the specialty. That is the conclusion that follows from the research documented across this body of work.
My book expanding on all of these is now available.
A Radiologist’s Introduction to Foundation Models: The Best Time in History to Be a Radiologist takes everything in this series deeper: the architecture of these models, the failure modes nobody is measuring, the liability gap most practices have not recognized yet, and the practical protocol for building a monitoring layer from the ground up.
It is on Amazon now: https://a.co/d/0fjYa3Nz
Radiologists are not being replaced. We are becoming radiologists with better tools, deployed one trusted workflow at a time.
That is why this is the best time in history to be one.
References
1. Kahn CE Jr, et al. Biomedical Ontologies to Guide AI Development in Radiology. PMC8669056. https://pmc.ncbi.nlm.nih.gov/articles/PMC8669056/
2. Kim et al. Safety-Aware ROC Framework. npj Digital Medicine 2026. https://www.nature.com/articles/s41746-026-02450-7
3. Royal College of Radiologists expert panel. Post-deployment monitoring and safety reporting of AI medical imaging devices in clinical practice. London: RCR, March 2026. https://www.rcr.ac.uk/our-services/all-our-publications/clinical-radiology-publications/post-deployment-monitoring-and-safety-reporting-of-ai-medical-imaging-devices-in-clinical-practice [Note: reflects UK NHS regulatory context; monitoring principles apply directly to US deployment.]
4. Bernstein, Sheppard, Bruno, Lay, Baird (Brown / Penn State / Seton Hall). Nature Health, March 10, 2026. https://doi.org/10.1038/s44360-026-00085-2
5. Asadi, O’Sullivan et al. MIRAGE: The Illusion of Visual Understanding. arXiv:2603.21687v2, March 2026. [Preprint.] https://arxiv.org/abs/2603.21687
6. ECR 2026 CXR Session (Rylands-Monk / Harvey / Snoeckx). AuntMinnie Europe, March 11, 2026. https://www.auntminnieeurope.com/resources/conferences/ecr/2026/article/15819005/ecr-does-ai-toll-the-beginning-of-the-end-for-chest-xray-reporting
Clinical AI does not fail at deployment. It fails quietly after deployment.
If you are building or deploying radiology AI, especially vision-language models or automated reporting systems, post-deployment monitoring is no longer optional. Regulators, including the FDA, increasingly expect continuous assessment of real-world performance, including drift, site variability, and human-AI disagreement.
Veriloop is a vendor-agnostic clinical AI observability layer designed to measure how models behave in real workflows. We track agreement, error patterns, and trust over time, helping teams understand where AI is reliable, where it breaks, and how much workload it can safely absorb.
We do not replace regulatory approval or pre-deployment validation. We provide the missing layer after deployment, where safety, performance, and trust actually evolve.
If you are building, deploying, or evaluating clinical AI systems and need real-world monitoring, reach out at ty@orainformatics.com