Large language models (LLMs) are reshaping radiology, but their integration into the reading room is far from straightforward. 2026 ARRS Annual Meeting Categorical Course Director Yee Seng Ng, MD, outlines the most significant barriers to adoption and why the solutions are more complicated than they appear.
Security Slows Adoption
Most widely available LLMs live on proprietary cloud platforms. Sending protected health information (PHI) outside a hospital network creates immediate compliance issues, and current guidelines from major organizations explicitly prohibit using public LLMs for protected patient information. Even if models claim not to store or reuse data, rads cannot verify how patient information is handled, refined, or monetized.
Private Solutions, But Not for All
Institutions can build private LLM instances behind their firewall, but this requires substantial infrastructure, IT support, and vendor partnerships—resources generally limited to large academic centers. Local installations using open-source models (via interfaces such as webUI) avoid cloud exposure but introduce new challenges: maintenance, computing requirements, and accessibility across the department.
Where Is NLP Already Helping?
Even though LLMs aren’t widely used at the point of care, rads are already benefitting from improved natural language processing (NLP) embedded in commercial reporting tools. Examples include:
- Automated impression generation from narrative text
- Converting freeform dictation into structured reports
- Organizing sentences under correct headings
These features accelerate reporting and reduce cognitive load without exposing PHI externally.
Error Prevention Still Matters
Simple NLP tools remain some of the most valuable. PowerScribe’s laterality and gender checks prevent avoidable mistakes that can undermine confidence in a report. Tools that flag mismatched anatomy—such as referencing a prostate in a female patient—provide immediate, low-friction safety nets that rads consistently appreciate.
Bottom Line
Security and workflow realities remain the biggest obstacles to adopting LLMs for radiology reporting. Until private, institution-controlled LLMs become practical and widely available, rads will continue to rely on integrated NLP tools that improve.
Leave a Reply