From Pixels to Partners: AI, LLMs and the Cloud Are Taking Radiology To a Higher Plane

For many of today’s radiologists, the reading room is beginning to look and feel more and more like the cockpit of a jetliner. There they sit, facing multiple screens crowded with multiple EHR tabs and multiple current and priors. Meanwhile, image volumes continue to increase along with exam complexity.

But now a not-so-new co-pilot is coming on board. Artificial intelligence (AI), fueled by large language models (LLMs) and the cloud, is reaching beyond image-reading assistance with an improved ability to prioritize cases, draft more detailed reports, and even take over many of the routine tasks that belabor radiologists on a daily basis.

And while there are headwinds to overcome, advocates argue that with seamless workflow design, explainable AI, responsible data practices, and cloud-scale infrastructure, radiology’s flight path points toward greater efficiency, stronger clinician support, and ultimately better patient care./

From Pixels to Purpose

In radiology’s early AI era, most innovation centered on “pixel intelligence”—algorithms that drew boxes around abnormalities such as subtle breast lesions that could go missed by the naked eye on mammograms. But AI is poised to take on more responsibility, says Tim Dawson, CTO and senior vice president of engineering at Optum Enterprise Imaging.

“We’ve shifted beyond focusing only on pixel intelligence. That was the focus for the first 5 to 10 years,” says Dawson. Today’s energy and excitement, he asserts, “is around large language models and where they can drive improvements.”

Rather than trying to read scans outright, emerging models can now improve orchestration of entire workflows: routing studies, prioritizing urgent cases, improving hanging protocols by automatically matching current and prior images, and pre-filling reports for radiologists. That shift reframes AI’s role from competitor to collaborator—reducing the grunt work so radiologists can focus on complex interpretation and clinical decision making.

Indeed, early AI detractors claimed AI would put radiologists out of their jobs; Dawson says those predictions missed the mark. It didn’t take long before it became clear that AI is best when it provides better tools. “I have an old t-shirt that says AI is not going to replace you,” he says. “Somebody using AI is going to replace you.”

PACS: The New Center of Operations

For decades, picture archiving and communication systems (PACS) have served as little more than storage and viewing tools. But today’s vision of PACS is that of an all-encompassing system that delivers the right comparison exams, an accurate patient history, and the right structured output at the right moment without radiologists having to dig for it.

“Physicians are really expecting it to be more of an end-to-end experience when making diagnoses and taking care of patients,” says Sonia Gupta, MD, chief medical officer at Optum Enterprise Imaging. “You want to be able to access the patient images, annotate them if needed, save them if it's a teaching case [and have] a fully unified view of all the patient data.

It’s got to be the foundation of the physician caring for the patient. That means not just [storing] and viewing images, but adding valuable context to inform patient care,” Dr. Gupta adds.

The Tech Backbone: The Cloud, LLMs and Agentic Workflows

Cloud infrastructure is making this vision feasible, says Jason Klotzer, customer engineering lead and a medical imaging expert at Google Cloud.

“We’re starting to transform and translate multiple data domains into something consumable by clinicians,” says Klotzer, who describes the elimination of constant toggling among multiple tabs for data in favor of an AI-enhanced PACS that integrates and seamlessly delivers images, reports, priors, and other data streams directly to the radiologist.

The result, he says, is less disruption, more continuity, and a reading environment that supports, rather than fragments, the radiologist’s attention. This may arguably be one of AI’s most valuable new capabilities. Radiologists cite repetitive clicks, manual hanging protocols, and redundant dictation among their daily frustrations, says Dawson, noting that the importance of seamlessness cannot be overstated.

“The best AI tools can be completely negated by a terrible workflow,” he says. “Radiologists don’t want another click or a new window—they want seamless integration. AI is all about eliminating toil. It’s the stuff you don’t want to do. That enables you to focus on helping patients, the reason you got into this business in the first place.”

Some examples of where AI can eliminate the menial work:

• Smart hanging protocols that automatically match current and prior images.
• Ambient AI transcription that listens while radiologists dictate their findings and then inserts them directly into reports.
• AI-assisted triage for screening programs in breast, lung, or prostate imaging, as well as in the emergency department.

Klotzer notes that increasingly sophisticated LLMs and foundation models are providing the behind-the-scenes scaffolding for these capabilities. Unlike earlier AI systems that required customized training for each task, today’s “agentic” models offer a flexible baseline that can be tuned to specific medical contexts, says Klotzer.

“[Agents] can orchestrate a lot of things that you would normally have to build hardcoded rules around,” he notes, citing routine tasks, context switching, summarization and, in particular, patient triage.

“It’s not just, ‘hey, does this patient have to be seen very quickly’ but it’s ‘is this person in the right facility to get the care that they need? Do we have people on staff to cater to what they need?’” he says. “Triage is absolutely a fantastic agentic use case right now because the agentic workflow can orchestrate many different variables simultaneously to make sure that you're getting the right outcome within that moment.”

Trust and the Human Connection

For all its promise, though, AI adoption hinges on one significant factor: trust. In Dawson’s view, trust of the technology improves with experience and with transparency—ie, the concept of “explainable AI.” Radiologists must be able to see not just the what but also the why behind AI outputs. Seeing the why is easier with pixel-based overlays—where clinicians can see for themselves whether an algorithm circled the right abnormality. But for more abstract tasks this will be more challenging, and the best systems must be able to “show their work,” as Dawson puts it.

Regulators are grappling this issue. A 2024 review reported that as of October 2022, the US Food and Drug Administration (FDA) had cleared or authorized 521 AI-enabled medical devices, 75% of which were in radiology. However, the same review found that only 46% of FDA-approved AI devices reported detailed performance metrics, and just 3.6% included race or ethnicity data—raising concerns about bias and generalizability.¹

The Economics of Adoption

Beyond trust, AI must also be able to demonstrate its value in economic terms. Burnout-driven attrition is expensive, and AI tools that streamline workflow are expected to offer returns far beyond productivity gains. In fact, evidence is mounting that AI-enabled tools can offer such value.

One recent study found that AI-assisted reporting cut average report creation time by nearly 50 percent without compromising accuracy.² Another analysis of an AI stroke management platform projected a 451% return on investment (ROI) over five years, increasing to 791% when accounting for savings in radiologist time.³ For health systems facing razor-thin margins, those savings matter.

For Dr. Gupta, the point is not to replace radiologists, but to protect them: “The cost of replacing that position can range anywhere from $250,000 to $1 million for the health system because of all the credentialing costs, the patient appointments that have to be rescheduled now because this physician is no longer there, and a whole flurry of administrative costs,” she observes.

The Invisible Future

Asked to imagine radiology five years from now, Dawson envisions AI disappearing into the fabric of practice. “We stop talking about AI,” he says. “It just becomes invisible —tools [that] anticipate the needs of the user, prioritize the most important cases, and put the right information at our fingertips. That maximizes productivity, effectiveness, and accuracy, which ultimately improves patient care.”

Dr. Gupta echoes the sentiment: “My hope is that … we're not even thinking about it, it’s just in the workflow. It’s working in the background, and the radiologist is able to focus more on the diagnostic piece, on the patient care,” she says.

Partners in Patient Care

AI, LLMs, and cloud computing are reshaping radiology from the inside out —far from replacing radiologists, they are instead taking them to a higher plane where they can serve as integrators and consultants across the entire patient care continuum.

The winners will not be those who resist change, but those who master AI as a quiet but powerful partner. For radiologists, the choice is clear: embrace the tools, demand transparency, and stay at the center of care. As Dawson reiterates, the future is not about radiology vs AI — it’s about radiologists working with AI.

“This is a transformational period in healthcare,” agrees Klotzer. “The technology is finally mature enough to make a huge impact in patient care.”

Schedule a Live RSNA Demo Today!

References

1. Muralidharan V, Adewale BA, Huang CJ, et al. A scoping review of reporting gaps in FDA-approved AI medical devices. NPJ Digit Med. 2024;7(1):273. Published 2024 Oct 3. doi:10.1038/s41746-024-01270-x
2. Bharadwaj P, Nicola L, Breau-Brunel M, et al. Unlocking the value: quantifying the return on investment of hospital artificial intelligence. J Am Coll Radiol. 2024;21(10):1677-1685. doi:10.1016/j.jacr.2024.02.034
3. Rajmohamed RF, Chapala S, Shazahan MA, Wali P, Botchu R. Evaluating the accuracy and efficiency of AI-generated radiology reports based on positive findings-a qualitative assessment of AI in radiology. Acad Radiol. Published online September 26, 2025. doi:10.1016/j.acra.2025.09.012