The Agentic Future of Radiology
Dr. Ali Tejani (Chief Neuroradiology Fellow, University of California) & Dirk Smeets (CTO icometrix)
At the 2025 annual meeting of the Radiological Society of North America (RSNA), Dr. Ali Tejani (UCSF) delivered a thought-provoking perspective on where radiology is heading. His talk, and the conversations that followed, lingered with us at icometrix.
Radiology has never had more tools at its disposal. Detection algorithms for hemorrhage and nodules, quantification tools for lesions and atrophy, large language models drafting impressions, multimodal systems connecting image and text.
And yet, despite this unprecedented technological acceleration, many radiologists feel busier than ever.
The Missing Layer Is Not Another Algorithm
When we stepped back, the issue did not seem to be technological capability. The pieces already exist:
● DICOM structured reports
● FHIR interoperability
● Standardized reporting templates
● Common data elements
The infrastructure for intelligence is already there, the real question is orchestration.
If AI outputs remain isolated (screenshots, PDFs, secondary captures, …), they will always sit adjacent to clinical decision-making. But when AI results become structured, standardized, and embedded into the radiologist’s natural workflow, something changes.
The report stops being a static document. It becomes a clinical interface that enables downstream intelligent action.
We have seen this evolution before in neurology. In multiple sclerosis, structured MRI reporting moved from “nice to have” to essential for treatment decisions, a journey we’ve described in Progress(ion) in Multiple Sclerosis. Standardization reduces variability and enables precision in clinical management.
Radiology is now ready for the same step forward.
One Interface. One Decision Point.
Dr. Tejani emphasized something beautifully simple: radiologists do not need more dashboards, they need fewer decision points.
The future is not ten AI widgets, it is one unified reading experience.
Imagine AI findings appearing directly where the radiologist already works. Contextualized. Structured. Editable. Accepted or rejected with a single action. No hunting. No copy-paste. No translation into narrative prose.
Low interaction cost and high clinical value.
Why Structure Changes Everything
Templates from societies such as the American College of Radiology and American Society of Neuroradiology already define what must be reported in complex cases.
Take ARIA monitoring in Alzheimer’s disease. Severity grading drives immediate treatment decisions. When these requirements are encoded as structured, machine-readable fields, AI can reliably populate them. Generative models can check completeness. Missing elements can be flagged before sign-off.
The radiologist remains in control, but no longer alone.This shift from free text to common data elements is not administrative detail. It is the foundation for intelligent action.
And once data is structured, something powerful happens: reports can trigger workflows.
A significant ARIA finding does not just sit in a PDF. It can notify neurology, initiate follow-up imaging, and connect to treatment pathways.
The report becomes dynamic and guides real time management to avoid critical delays in care or complications.
Generative AI: Safety Net, Not Shortcut
Since 2022, large language models have matured rapidly. They can summarize, structure, suggest differentials, and detect inconsistencies.
Used carelessly, they create risk. Used responsibly, they could reduce it.
We are particularly encouraged by governance frameworks such as MI-CLEAR, developed in collaboration with RSNA and published in the Korean Journal of Radiology. Transparency, reproducibility, monitoring for drift. These are not academic exercises. They are prerequisites for robust and reproducible results.
In this context, generative AI becomes a safety layer. It ensures required fields are completed. It cross-checks guideline adherence. It reminds me without overwhelming.
It augments cognition instead of competing with it.
From Tools to Agents
The next frontier is what many now call “agentic AI.”
Not a single monolithic brain in the cloud but coordinated, specialized agents working quietly behind the scenes.
One agent references guidelines.Another structures data elements.
Another monitors consistency across timepoints.
Another prepares downstream triggers.
All under human oversight.
We believe the real opportunity is not building smarter individual algorithms. It is building systems that collaborate, with each other and with clinicians.
A Clinician-Driven Future
Perhaps the most important insight from our exchange is this: the agentic future of radiology will not be vendor-driven, it will be clinician-shaped.
Radiology can remain a collection of impressive but isolated tools or it can evolve into an intelligent, interoperable ecosystem.
From fragmentation to intelligence.From static reports to clinical pathways.From algorithms to agents.
The technology is ready.
The standards are ready.
The question is: are we ready to orchestrate them?

