Structured ACR TI-RADS features improve ChatGPT-5.4 thyroid nodule classification over image-only input

Adding ACR TI-RADS descriptors to ultrasound images increased ChatGPT-5.4’s accuracy for thyroid nodule classification to 84.2% vs 73.8% with images alone (P=0.001), but matched features-only input (84.2%).

• Accuracy significantly improved when clinician-recorded TI-RADS features were provided with images (84.2%) versus image-only (73.8%; P=0.001).
• Features-only input also achieved 84.2% accuracy; adding images to the features did not improve overall accuracy but shifted sensitivity (91.9% vs 96.3%) and specificity (68.2% vs 59.1%).
• The study included 202 nodules (66 benign, 136 malignant) from 153 patients with histopathological confirmation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.