Can GPT-5.0 Interpret Thyroid Ultrasound Images? A Comparative TI-RADS Analysis with an Expert Radiologist

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Radiology

Background:

• Large language models (LLMs) are being explored for medical image interpretation, including thyroid ultrasounds (USs).
• Assessing thyroid nodules requires evaluating subtle echogenic and morphological details, a challenge for AI.
• The American College of Radiology (ACR) Thyroid Imaging Reporting and Data System (TI-RADS) provides a standardized framework for evaluating thyroid nodules.

Purpose of the Study:

• To evaluate GPT-5.0's performance in interpreting thyroid US images based on TI-RADS criteria.
• To compare GPT-5.0's assessments with those of an expert radiologist.
• To determine the diagnostic accuracy of GPT-5.0 for thyroid nodule malignancy risk stratification.

Main Methods:

• Retrospective analysis of 100 thyroid US cases with cytology-confirmed diagnoses (Bethesda II, V-VI).
• Independent assessment of TI-RADS features (composition, echogenicity, shape, margin, echogenic foci) by GPT-5.0 and a radiologist.
• Analysis of inter-rater agreement using Cohen's kappa and diagnostic performance metrics (sensitivity, specificity) with TR4-TR5 as positive for malignancy.

Main Results:

• GPT-5.0 showed substantial agreement with the radiologist for composition, shape, and margin, but only moderate for echogenicity and poor for echogenic foci.
• GPT-5.0 exhibited a tendency to up-classify nodules, assigning more TR4-TR5 categories, indicating a risk-averse bias.
• The radiologist achieved superior diagnostic performance (sensitivity 89%, specificity 85%) compared to GPT-5.0 (sensitivity 67%, specificity 49%), with GPT-5.0 frequently misclassifying benign nodules as suspicious.

Conclusions:

• GPT-5.0 can identify some TI-RADS features but struggles with microcalcifications and overestimates malignancy risk.
• The current performance limits GPT-5.0's standalone clinical utility for thyroid nodule risk stratification.
• Domain-specific training and adaptation may enhance LLMs for adjunctive roles in thyroid nodule assessment.