AI in Radiology: From Isolated Algorithms to Scalable Clinical Impact

Artificial intelligence in radiology is often discussed in broad, aspirational terms, but far less attention is paid to what happens after algorithms are cleared, purchased, and deployed. In a recent discussion hosted by Applied Radiology, experts examined how AI is being implemented at scale and what it takes to translate technical capability into meaningful clinical impact.

During the conversation, Avi Sharma, MD, host of Applied Radiology’s AI Podcast was joined by co-host Lawrence Tanenbaum, MD, and Greg Sorenson, MD, Chief Science Officer at RadNet, and, to explore how AI moves from isolated tools to enterprise-level infrastructure. The discussion focused less on individual algorithms and more on workflow, adoption, and sustainability in real-world imaging environments.

Dr. Sharma opened by noting that AI is frequently treated as a catchall term rather than a defined clinical tool. He emphasized the need to move beyond buzzwords and examine how AI is actually used in practice. That perspective framed much of the discussion, particularly as Dr. Sorenson described RadNet’s experience deploying AI across a large outpatient imaging network.

According to Dr. Sorenson, technology alone does not drive outcomes. He explained that early success required rethinking workflows rather than simply layering AI onto existing processes. Without controlling how AI outputs are reviewed, acted upon, and integrated into daily operations, the technology would not have delivered measurable benefit at scale.

One of the most frequently cited examples involved screening mammography. Dr. Sorenson described how RadNet paired AI with a redesigned workflow that introduced a targeted second-review process for a small subset of AI-flagged cases. While the majority of these cases were ultimately concordant, the subset in which radiologists disagreed proved clinically meaningful. That approach, supported by data from a large U.S. study involving more than 500,000 women, demonstrated a notable increase in cancer detection when AI was used as part of a structured safeguard rather than a standalone reader.

Radiologist adoption emerged as a central theme throughout the discussion. Dr. Sorenson acknowledged that skepticism was common during early deployment, particularly among experienced readers concerned about false positives or workflow disruption. Over time, however, exposure to real cases in which AI identified missed findings helped shift perception. He noted that some of the strongest initial critics ultimately became advocates once the clinical value became tangible.

Dr. Tanenbaum expanded on the human dimension of adoption, describing how AI support can reduce the cognitive burden associated with final interpretation. He emphasized that confidence at the point of sign-off has practical value, particularly in high-volume environments where uncertainty and fatigue can accumulate. Reducing the lingering question of whether something was missed can meaningfully change the reading experience.

The conversation also addressed the economic realities of AI adoption. Dr. Sorenson described how reimbursement has not always kept pace with innovation, requiring providers to explore alternative models. In some cases, RadNet experimented with patient self-pay or employer-sponsored programs to support early deployment. A more established pathway has emerged in thyroid ultrasound, where AI-driven automation improved reporting efficiency and supported billing under a Category III CPT code. Faster sign-off and improved consistency, he noted, created operational and clinical benefits that extended beyond reimbursement alone.

Looking beyond individual use cases, Dr. Tanenbaum urged imaging leaders to think more broadly about infrastructure. He pointed to RadNet’s remote operations model as an example of how AI can support distributed expertise, enabling radiologists and technologists to collaborate across sites. Rather than focusing solely on algorithm performance, he argued that scalable impact depends on how AI connects clinical, operational, and enterprise workflows.

Dr. Sorenson reinforced that view, noting that the long-term value of AI lies in integration rather than accumulation. Deploying dozens of disconnected tools, he suggested, creates complexity without solving underlying problems. Success depends on aligning AI with clinical priorities, operational realities, and measurable outcomes.

As the discussion concluded, Dr. Sharma reflected on a recurring theme that surfaced across use cases. AI delivers the most value when it functions as part of a human-centered workflow rather than as a replacement for clinical judgment. Throughout the conversation, that principle remained consistent. Meaningful AI adoption in radiology depends not just on algorithms, but on thoughtful integration, clinician trust, and systems designed for real-world practice.