AI in Radiology Right Now: What's Actually Happening in March 2026
A lot is moving in radiology AI right now — and not all in the same direction. Here's what's worth paying attention to this month.
Foundation Models Are Finally Hitting Clinics
At ECR 2026, ACR's Christoph Wald made the case that 2026 is the "year of the foundation model," with imaging practice on the precipice of deploying next-generation AI in real clinical settings.
That's not just conference hype. On March 4th, Nature published research on a model called Merlin that backs it up. Merlin is a 3D vision-language model trained on abdominal CT scans, radiology reports, and electronic health records. The results are notable: Merlin showed stronger off-the-shelf performance than other vision-language models across three hospital sites outside its original training center, outperforming second-best models by an average of 20%.
That last detail matters. Most radiology AI falls apart the moment it hits a hospital it wasn't trained on. Merlin's cross-site consistency is what makes it clinically interesting, not just academically interesting.
Physician AI Adoption Has Doubled — But the Tension Is Real
The AMA's 2026 Physician Survey found that more than 81% of physicians now use AI in their practices — more than double the 38% rate recorded in 2023. The average number of AI use cases per physician rose from 1.1 in 2023 to 2.3 in 2026, with medical research summarization and clinical documentation the most common applications.
But adoption and enthusiasm aren't the same thing. More than three-quarters of physicians believe AI improves their ability to care for patients, up from 65% in 2023 — yet 88% expressed concern about potential skill loss, particularly among those with 10 years or less in practice.
And on the question of patients using AI to interpret their own radiology results? Nearly half of physicians strongly oppose it, with 46% saying they would never or rarely want patients using AI to interpret radiology results without physician involvement.
The Malpractice Question Is Getting Serious
A study published March 10th in Nature Health tackled something most people in the field have been quietly wondering about: if AI flags a problem and a radiologist still misses it, who's liable?
Researchers presented 282 mock jurors with a hypothetical malpractice case where a patient suffered irreversible brain damage after a radiologist missed a brain bleed — even though AI correctly identified the CT as abnormal. In one condition, the radiologist interpreted the scan once after seeing AI feedback. In another, they read it first without AI, then again with it.
The difference in verdict was stark. Participants were more likely to side with the plaintiff in the single-read versus the double-read condition — 74.7% versus 52.9%.
In plain terms: how a radiologist uses AI affects whether a jury thinks they did their job. That's a workflow design problem with real legal consequences.
ECRI Named AI Diagnostics the #1 Patient Safety Threat of 2026
"Navigating the AI diagnostic dilemma" was named the No. 1 threat to patient safety in 2026, according to a report published March 9th by ECRI.
The report notes that AI isn't foolproof and can contribute to diagnostic mistakes — and that these models can perpetuate biases, lack transparency, and erode clinicians' critical thinking skills.
ECRI's position isn't anti-AI. They're explicit that "in order for AI to be used effectively in diagnosis, clinicians must view it as a tool designed to supplement and support clinical expertise — not replace it." The concern is speed of deployment outpacing training, governance, and accountability.
What Radiologists Actually Need From AI
Here's the tension that keeps surfacing across all of this: the AI tools getting built aren't always solving the right problems.
A blunt take from Diagnostic Imaging: "Radiologists don't need AI to detect things for them." Research shows radiologists only need 250 milliseconds to spot a finding on a chest X-ray. Where they actually get bogged down is cognitive and administrative load.
What the field needs are tools that synthesize findings, summarize prior exams, factor in clinician intent, and translate image data into actionable reports — tools that fit inside existing workflows rather than sitting beside them.
That's a harder product to build than another detection algorithm. But it's the one that would actually get used.
The Bigger Picture
NVIDIA's 2026 State of AI in Healthcare survey confirms the industry is moving from AI experimentation to execution, with medical imaging and drug discovery leading the way in ROI. AI adoption is up across every healthcare segment — with digital health leading at 78%, followed by medical technology at 74%.
But the smartest observation in the whole space right now might be this one: the ease of making a new model is at odds with the difficulty of encouraging widespread adoption.
Getting a model to work in a lab is one problem. Getting it to work in a department of tired radiologists who are managing liability risk, skill atrophy concerns, and a patient population that increasingly wants to run their own scans through ChatGPT — that's a different problem entirely.