Sunday, November 29 | 9:20 a.m.-9:30 a.m. | S1-SSHN01-3 | Room E352
In this scientific presentation, researchers will share findings from a large, real-world deployment of AI software for thyroid ultrasound.
AI has shown promise for improving efficiency and accuracy and for reducing inter-reader variability in thyroid ultrasound. However, its clinical impact ultimately depends on whether radiologists trust and adopt its outputs in routine practice, noted presenter Milad Mohammadzadeh, PhD, of AI software developer See-Mode Technologies and colleagues. To gauge this, the group prospectively assessed how often radiologists kept the TI-RADS descriptors generated by an AI system.
The software (See-Mode) was deployed at 297 clinics across the U.S. and automates nodule detection, characterization, and report generation. The researchers used a validated large language model pipeline to compare radiologists’ final signed reports from September 2025 through April 2026 against the AI-generated structured reports.
Across more than 95,000 studies covering over 187,000 nodules, radiologists left all AI-generated descriptors unchanged, and therefore agreed with the AI’s TI-RADS level, in nearly 89% of nodules. Agreement was even higher for individual descriptors, ranging from about 95% to 98%, according to the researchers.
“This high human-AI concordance is important for building trust in AI outputs, which is critical to successful long-term integration of AI into routine clinical practice,” the researchers wrote in the abstract.
Learn more at this Sunday morning talk.
