Reflections from RSNA 2025: AI is everywhere, but are we solving the right problems?

I just got back from RSNA 2025 in Chicago, and if there's one takeaway from walking those massive exhibit halls, it's this: everything is "AI." Seriously—every booth, every banner, every demo station had the letters "AI" somewhere prominent. The conference organizers say over 200 companies were in the dedicated AI Showcase alone.

But here's what struck me after dozens of conversations: most people couldn't actually tell you what makes one AI solution different from another. I'd ask vendors, "Okay, but what makes yours unique?" and the answers got pretty fuzzy pretty fast. Everyone's detecting nodules. Everyone's improving workflow. Everyone's using deep learning. The market has gotten crowded enough that genuine differentiation is hard to find, and I think that's telling us something about where we are as an industry.

To be fair, there were bright spots. Opportunistic screening—where AI extracts additional findings from routine scans, like coronary calcium scores from chest CTs—represents practical value. And the growing use of LLMs for catching speech recognition errors in reports is genuinely useful. But these workflow-oriented applications got far less booth space than the hundredth nodule detector.

United Imaging is making a splash

One company impossible to ignore was United Imaging. The Chinese manufacturer has become a real competitor to Siemens, GE, and Philips—and the numbers back it up. They now have over 400 installations across 70% of US states, and just announced they're tripling their Houston production footprint to meet demand. At RSNA they showcased their FDA-cleared 5T MRI, a new ultrasound product family, and their "Native AI" approach of building intelligence into hardware from day one.

For clinical trials, this growth has a flip side: more scanner variability. When your study sites include not just the big three but also United Imaging systems with different acquisition characteristics, harmonization gets harder. It's a good problem to have—more competition drives innovation—but it adds another layer of complexity to multi-site data management that most AI tools aren't designed to handle.

Clinical trials are still a mess

Anyone who works in imaging clinical trials knows this already, but RSNA was a good reminder: we still haven't figured out how to manage multi-site imaging data without pulling our hair out. You've got 12-15 centers across different countries, different scanners, inevitable protocol deviations, and someone has to harmonize all of that. AI can spot an anomaly in a brain scan, sure, but it's not going to fix the fact that your site in Boston is using a different acquisition protocol than your site in Barcelona.

The operational stuff—patient enrollment, protocol adherence, data transfer between sites—that's where studies still get bogged down. We don't talk about it as much because it's not sexy, but it's where the real bottlenecks are.

The international nature of radiology

Something I've always valued about radiology: it's genuinely international. You can't make breakthroughs in medical imaging without collaboration across borders. The international nature of RSNA - researchers from Germany, clinicians from Japan, and engineers from Brazil - shows exactly what makes multi-site coordination so critical and so hard. Progress happens when institutions share data and expertise, not when they work in isolation. The conversations I had reinforced how much the field depends on that global perspective.

The real problem AI hasn't solved

This year's RSNA theme was "Imaging the Individual"—and there's real promise there. Photon-counting CT with spectral analysis tailored to specific patients. AI models that adapt to individual anatomy. But here's the irony: the more we individualize imaging, the harder multi-site coordination becomes.

Anyone in clinical trials knows this pain. A dozen sites across four continents, different scanners, inevitable protocol deviations—and now growing equipment diversity as United Imaging gains ground alongside Siemens, GE, and Philips. Someone has to harmonize all of that before any analysis can happen. AI can spot an anomaly in a brain scan, but it won't fix the fact that your Boston site uses a different protocol than Barcelona, and one just installed a scanner with completely different acquisition characteristics.

We've gotten very good at analyzing individual images. What we haven't solved is coordination: getting images collected consistently across 20 sites, making sure data flows where it needs to go, keeping trials on schedule when centers use different EMRs and five scanner vendors. That's where studies actually get bogged down—not in the analysis, but in the operational chaos upstream.

Those problems don't get solved by better segmentation algorithms. They get solved by better systems thinking and better software for how medical imaging actually works.

Flying home, I'm optimistic about where the technology is headed, but I also think we need to redirect some of that AI energy toward the problems that are actually holding us back. Sometimes the hardest problems aren't the most exciting ones.

Paulo Rodrigues is CTO and Co-Founder of QMENTA, a medical imaging AI company providing cloud-based solutions for clinical trials and research.