What Did RSNA 2016 Get Right and Wrong About Where Radiology Was Heading?

Ten years is a long time in technology. In radiology, it can feel like several lifetimes.

RSNA 2016 was a moment when AI and deep learning moved from academic research into mainstream conversations about day-to-day workflow and patient care. The exhibit floor was full of possibility and chatter about one (in)famous prediction (more on that below). 

For those familiar with the alternate definition of R.S.N.A., there’s something fitting about looking back before everyone heads to Chicago. With 10 years of hindsight, it’s interesting to revisit which predictions proved prescient, which missed the mark and what both can tell us about where the industry may head next.

Past: What Radiology Was Talking About at RSNA 2016

Deep learning was one of the dominant narratives, but there were very different ideas about what it meant for radiology.

In “When Machines Think: Radiology’s Next Frontier,” Keith J. Dreyer, DO, PhD, a radiology leader from Harvard Medical School and Massachusetts General Hospital, presented a symbiotic vision: AI could complement human expertise and help move radiology toward more personalized, data-driven medicine.

Then there was the prediction that generated considerably more attention. Earlier that November, Geoffrey Hinton said in a panel discussion the industry should “stop training radiologists,” predicting that deep learning would outperform radiologists within five to 10 years.

The exhibit hall was also enthusiastic about IBM Watson Health’s “Eyes of Watson,” while emerging deep learning companies such as Enlitic demonstrated new approaches to clinical image interpretation, but others fell victim to the (the first?) demo versus reality debate.

Simon Harris of Signify Research noted that December, “Vendors need to complete and promote clinical studies to validate their claims, otherwise marketing soundbites may impede the acceptance of deep learning in radiology. More customer education is required so that the conversations at next year’s RSNA move on from ‘what’s deep learning?’ to ‘tell me how can deep learning help me do my job better.’”

AI wasn’t the only major conversation. Gadolinium retention was driving scrutiny of linear versus macrocyclic contrast agents. PI-RADS Version 2,  LI-RADS and BI-RADS reflected a broader push toward standardization. The impending MACRA and MIPS go-live lead to discussions on leveraging data to maximize reimbursement.  

Present: What Radiology Is Living 

One prediction was spot on: Radiology became the proving ground for clinical AI. Today, the majority of FDA-authorized AI-enabled medical devices are in radiology. The replacement narrative, however, went very differently.

Instead of obsolescence, radiology is confronting rising imaging demand and persistent workforce constraints. Those challenges — not to mention burnout and declining reimbursement — have changed the problems AI is being asked to solve.

AI can now identify potentially time-sensitive findings, such as large vessel occlusions, intracranial hemorrhage and pulmonary embolism, to help prioritize studies and accelerate care coordination. It’s also moving beyond detection into reporting, documentation, clinical context and other work surrounding image interpretation.

That’s a very different future from the autonomous machine radiologist imagined in some of the 2016 predictions. Instead, AI’s role is increasingly about extending the radiologist’s capacity.

Other debates evolved, too. Concerns about gadolinium retention accelerated movement toward more stable macrocyclic agents. Dense-breast notification progressed from a state-by-state movement to a federal requirement.

The Misses: What the Industry Got Wrong

Perhaps the biggest miscalculation was underestimating how difficult it would be to turn an accurate algorithm into an effective clinical workflow. 

The first generation of radiology AI was dominated by point solutions — individual models built for individual findings. I could spend hours linking to all the announcements from various companies touting their individual solutions, but you probably remember this era.  

Ten years later, we’ve learned that at enterprise scale, managing disparate solutions without a unifying infrastructure quickly becomes very complicated. Every additional solution could mean another integration, workflow, validation process and system to manage.

IBM Watson Health offers another useful lesson. Its ambition was to bring together enormous amounts of medical knowledge and patient information to help clinicians make better decisions. But real-world healthcare data proved fragmented, messy and difficult to operationalize. IBM ultimately sold the Watson Health assets in 2022.

Interestingly, the underlying idea behind that technology never disappeared. With large language models, foundation models and multimodal AI, the industry is once again exploring systems capable of working across different forms of clinical information.

The last decade also demonstrated that regulatory authorization isn’t the finish line. Patient populations, scanners, protocols and clinical environments differ and data will drift overtime. Governing and monitoring how it performs in the real world is increasingly important.

Possible: What Comes Next

If the RSNA 2026 program is any indication, the next decade may broaden the definition of what radiology AI actually does.

Sessions on foundation models are exploring systems that move beyond single tasks toward multimodal, multidisease and more general-purpose capabilities. Research into opportunistic screening is asking what else can be learned from images patients are already getting. And work combining imaging with clinical data and even pathology suggests that the next generation of AI may increasingly reason across information that’s traditionally lived in separate places.

That begins to look surprisingly similar to the symbiotic future Dr. Dreyer described a decade ago: human expertise amplified by computational intelligence and broader clinical context.

By 2036, the biggest change may be that radiologists spend much less time thinking about whether they’re “using AI.” The most effective AI may simply become part of the infrastructure — connecting information, anticipating what is needed and quietly removing work that doesn't require a radiologist's expertise.

Heading to RSNA 2026? Let’s Connect.

Join us in Chicago to hear from the people shaping what’s next , see what’s already possible and explore how AI can help put radiologists — and radiology — at the center of care. Schedule a meeting.