ChatGPT-4 Vision: a promising tool for diagnosing thyroid nodules

Area of Science:

• Artificial Intelligence in Medical Imaging
• Diagnostic Ultrasound
• Thyroid Nodule Analysis

Background:

• Thyroid nodules are common, requiring accurate risk stratification for malignancy.
• Current ultrasound analysis relies on experienced sonographers and established guidelines like C-TIRADS.
• AI tools offer potential for standardized and consistent image interpretation.

Purpose of the Study:

• To evaluate the diagnostic efficacy and consistency of ChatGPT-4 Vision in analyzing ultrasonic images of thyroid nodules.
• To compare the performance of ChatGPT-4 Vision against sonographers with varying experience levels.

Main Methods:

• Prospective study involving 124 patients with pathologically confirmed thyroid nodules.
• Ultrasound images were analyzed by ChatGPT-4 Vision and two sonographers using C-TIRADS guidelines.
• Performance was compared against fine needle aspiration (FNA) biopsy or surgical pathology results.

Main Results:

• ChatGPT-4 Vision achieved a sensitivity of 86.2% and specificity of 60.0% (AUC 0.731), comparable to a resident physician but lower than an attending physician.
• AI demonstrated good consistency with resident classification (Kappa=0.729) but lower consistency with pathological diagnosis compared to the attending physician (Kappa=0.457 vs. 0.816).

Conclusions:

• ChatGPT-4 Vision shows clinical utility in thyroid nodule risk stratification, with performance comparable to resident physicians.
• The AI's standardization capabilities align with precision medicine and may support future integration with advanced ultrasound techniques.