Rad AI Study Finds Specialty Models Beat General LLMs in Radiology Workflows

New updates have been reported about Rad AI.

Rad AI has published peer-reviewed research in npj Digital Medicine showing that its domain-specific radiology AI more closely matches radiologist expectations than general-purpose large language models when generating radiology impressions, the core section of reports that drives treatment decisions. Conducted with Moffitt Cancer Center and based on 200 oncologic CT reports, the study is one of the first multi-stakeholder evaluations focused on how AI-generated impressions perform in real clinical workflows.

The research compared radiologist-written impressions with outputs from a radiology-specific model fine-tuned on institutional data and from a generic LLM, finding the domain-specific system closely aligned with human radiologists on completeness, correctness and conciseness. General-purpose models were consistently ranked lower by radiologists, with usability gaps of roughly 28% to nearly 50%, and Rad AI’s tuned model also produced high-quality impressions faster while maintaining the concise, high-signal format clinicians prefer.

Risk to patients remained low for all systems, with harm scores of around 1.0 to 1.2 on a 3-point scale where 1 indicates minimal risk, reinforcing that the main differentiator is workflow fit rather than safety alone. Rad AI’s chief medical information officer, Andrew Del Gaizo, said the data underscore that customization and alignment with actual radiology practice are critical for clinician confidence, not just raw accuracy.

The study found radiologists favor succinct summaries, while oncologists often value more detailed narratives, highlighting a need for adaptable AI that tailors outputs to specialty-specific preferences. Moffitt’s Trevor Rose noted that impression quality remains partly subjective even within a single specialty, implying health systems should prioritize configurable tools instead of seeking one standard output format.

For Rad AI, the findings validate its strategy of building purpose-built radiology solutions such as Rad AI Reporting, Impressions and Continuity, rather than relying on off-the-shelf general models. This strengthens Rad AI’s competitive positioning as hospitals and imaging groups increasingly differentiate between generic LLM offerings and clinically tuned systems that can scale efficiency, support growing imaging volumes and maintain high standards of care.

The publication adds scientific backing to Rad AI’s commercial traction, following recent recognition on Deloitte’s Fast 500, CNBC’s Disruptor 50 and Fast Company’s Most Innovative Companies lists, signaling strong growth and market acceptance in healthcare AI. As health systems advance their AI procurement, the study is likely to influence purchasing decisions toward domain-specific platforms like Rad AI’s, with implications for contract wins, pricing power and long-term integration into radiology workflows.