What RSNA 2025 taught me

RSNA, the Radiological Society of North America's annual meeting in Chicago, is the biggest radiology gathering of the year. Every vendor with an AI product is there, and so is everyone who wants to buy, fund or argue about one. The common expectation going in, mine included, is that the exhibition floor will be dominated by autonomous AI: software that reads the study and writes the report. I went for the first time in 2025, and what I found was rather different. This is what I wrote that week, lightly cleaned up, followed by what I think it means for centres here in India.

I went in expecting the hype

I've been building AI in radiology for the past 3 years. This week I attended RSNA 2025 for the first time. And I was shocked, but in a very pleasant way. We are in the midst of a hype "Agentic AI" wave, courtesy Silicon Valley. I went in expecting to hear booths promising to replace clinicians with autonomous software.

But... that is not what I found. What I saw and heard was actually more grounded, encouraging and realistic.

Flashy demos were few and far between. The impact of AI was happening in the mundane daily workflows. The companies getting the most crowd (and traction) were those that were building the infra to help radiologists scale, not the ones that were trying TO BE radiologists!

The narrow, unglamorous workflows got the real attention

Even narrow, niche but important workflows like triaging emergency pathologies or mapping severity to prioritise queues were well thought out and received deeper discussions amongst administrators, IT teams and radiologists.

First, a few words on why that matters. A worklist (the queue of studies waiting to be reported) in a busy department is usually sorted by priority and then by time of arrival. A head CT with an unsuspected bleed, ordered as routine, can sit next to a routine follow-up scan, and whichever came first gets read first. Triage AI looks at the study as it arrives and moves the urgent one up.

Nobody's job is replaced; the radiologist still reads both, and the patient who needs it gets read sooner. Mapping severity to the queue works the same way, only with a scale instead of a flag.

It is not the sort of thing that makes headlines. But it is the sort of thing an administrator, an IT head and a radiologist can all sit around a table and agree on, because each of them gets something from it, and that is what I saw happening.

But there is this irony... and a testament to the kind of crazy times we live in. These are the kinds of companies that make real-world impact, and yet are rarely ever covered by tech journalists.

Honest integration wins mindshare

While it is still early days in solving the interoperability problem, I found vendors that stayed honest in their integration approach got a leg up in the battle of mindshare. The best tools plugged into existing PACS as simple widgets. Most were unobtrusive.

I keep coming back to that line. A radiologist reads study after study in a shift, with muscle memory built over years. Any tool that asks them to open a new window, learn a new layout or log in somewhere else is asking for something they do not have, which is spare attention. The advantage of a widget inside the existing PACS is exactly that it asks for none; its disadvantage is that it is still a guest in someone else's system, limited to whatever that system lets it see. The tools that were winning had accepted that trade-off and worked within it.

Foundation models and the messiness of hospitals

As is the vogue, I did meet up with people who were building foundational models. While the technology they are building sounds impressive - in healthcare, technology is often the easy part. The jury is still out on whether they can navigate the operational messiness of the real world, aka hospital providers.

By operational messiness I mean the things a model never sees in a benchmark: studies that arrive with the wrong body part in the header, priors (earlier scans of the same patient) stuck in another system, a radiologist who reports across three centres, a referring doctor waiting on a phone call. Getting a model to perform on a curated dataset is one problem. Getting it to show up at the right moment inside that daily mess is another, and it is mostly an infrastructure problem rather than a modelling one.

What I brought home

I return with the satisfaction and hope that AI will indeed find a home in radiology - but in the messy workflows, and as an unobtrusive smart technician that allows radiologists to both scale and excel at their job.

What this means for a centre

If you run a centre in India and you are being pitched AI, RSNA's lesson translates fairly directly. Before asking how accurate a tool is, ask how much it changes your radiologists' day. A few questions follow:

1. Does the result show up inside the viewer and worklist they already use, or somewhere else?
2. Does it need another login?
3. Does it help with something that happens every day, like getting urgent studies read first, or something that happens once a month?
4. What does it need from your other systems (the RIS, the PACS, the priors) to work, and who will maintain that connection?

So where does that leave a centre being pitched AI? The tools that ask for the least change are the ones your team will actually use, and a tool nobody opens has no accuracy at all. It is worth weighing that before any benchmark figure.

FAQ