Ask a health system what it is doing about AI in imaging, and you will usually hear about models. Which vendor, which finding, which pilot. Ask Dr. Marc Kohli, and the answer starts somewhere much less exciting.
Kohli is Medical Director of Imaging Informatics at UCSF Health. He co-chairs the RSNA and ACR Common Data Elements effort, chaired the SIIM board, and still reads clinically every week. On episode three of Flight Path he made a case that is easy to nod along with and hard to actually fund: what decides whether your AI works is the plumbing underneath it.
Not the algorithm. The routing, the standards, the quality of what you are able to hand a model in the first place.
UCSF's posture is deliberately unhurried. "We've been fairly conservative, wanting to make sure that the AI tools we purchase are solving an actual clinical problem," Kohli said. "So we're starting with the clinical need rather than starting with the desire for AI."
That sentence is worth sitting with, because most AI programs run the other way around. The tool arrives first and the problem gets reverse-engineered to fit it. Starting from the clinical need is slower and it produces fewer press releases. It also means you can say what a deployment is for.
When Kohli talks about what has actually made AI practical at UCSF, he does not name a model. He names a router.
"We put in a DICOM router almost ten years ago. That DICOM router has made it much easier for us to send images to the right system for processing and inference. Being able to do that in a rules-based way is incredibly valuable. We basically use it as our AI orchestrator." He credits the insight to his colleague John Mongan.
A decade-old piece of routing infrastructure, bought for reasons that had nothing to do with machine learning, turns out to be the thing that lets a health system point studies at the right engine at the right moment. Kohli's practical advice follows from that: look for a product that lets you write flexible routing rules, and push the people selling you anything to support real interoperability standards. FHIR first. DICOM already gives imaging a head start that most of healthcare does not have.
The other half of the plumbing is what you write down.
"If I write something in free text and you read it, you may not understand the exact same thing that was in my brain," Kohli said. "I think of communication as a potentially lossy compression process. I've got a concept in my head, I spit out something else, and then you have to decompress that again and hopefully arrive at the same mental model."
That is the argument for Common Data Elements, and it is a better argument than "structured data is good." RadLex gave the field codes for concepts, so a computer reading "left upper lobe" knows what it is looking at. Common Data Elements add question and answer pairs on top. Is pneumonia present, yes or no. If yes, where. The point is not to strip language out of radiology. It is to remove ambiguity where ambiguity is not doing any work.
Kohli is careful about the limit. Radiologists are often not delivering an answer at all. They are delivering a differential, a range of possibilities, a degree of uncertainty. Prostate imaging has PI-RADS and a precise shared language for it. A less specific study needs room for a broad differential. "We kind of need both to meet the needs of radiology practice."
Any push toward structure that cannot hold both of those at once will get quietly abandoned by the people who have to use it.
Kohli was in the room early enough to remember the skepticism. HL7 Version 2 was everywhere. Why would anyone move?
"I just want to celebrate that FHIR is now kind of a household word in the informatics community. A lot of companies are building on top of FHIR." Part of why it took hold is that its Diagnostic Report Framework carries a free-text report and structured data in the same resource, with a specified way to do it. Part of it is plainer than that: FHIR uses JSON, so a developer who has never touched healthcare can pick it up without learning the habits of HL7 V2 first.
Here is where the plumbing stops being a technical topic.
"Where we do create these data sets, they have bias issues associated with them because those data sets are not representative of the patient population across the United States or broadly," Kohli said. "That limits the ability of algorithms to generalize."
Asked who loses, he was specific: "The other population at high risk of losing out is the people in critical access hospitals and rural populations, especially in the United States. If we don't work on creating larger data sets that are more representative, you're still going to have things benefiting people on the coasts and kind of missing the middle of the country."
A model performs best on patients who look like the ones it was trained on. If the training data comes from a handful of well-resourced systems, the tool arrives at a rural hospital already worse at its job, and nobody in that building can tell by looking at it. The gap does not announce itself.
Kohli sees the opening in the same place. Health systems away from the coasts hold data that these models need and often lack the capacity to train on it themselves. Connecting that data to organizations that can, whether academic or commercial, is how the pools get bigger and more representative. Which puts a lot of weight on something unremarkable: whether a health system can find what is in its own history, and de-identify it well enough to share.
Kohli's governance work at UCSF turned out to be mostly teaching. Not fairness metrics and model cards, at least not at first, but the more basic question of how AI differs from every kind of software people already knew.
And then there is the shortcut everyone wants to take. "A lot of people come to me and say, well, I want to do AI. And then the challenge is that they don't want to do the unsexy thing of learning the fundamentals of informatics. But I really don't think you can do the AI without understanding the fundamentals of informatics. If you don't know what DICOM is, you're kind of going to be sunk when it comes to AI orchestration."
None of those questions are about AI. That is the point.
Flight Path Ep. 3: Marc Kohli on data standards and AI readiness. Full conversation, chapters, and show notes.
Mach7 builds vendor-neutral enterprise imaging: the Vendor Neutral Archive for one governed record of every study, the Communication Workflow Engine for rule-based DICOM routing and normalization, the eUnity enterprise viewer, AI enablement, and Flamingo, the next Mach7 product set.