Impact of Clinician-Recorded ACR TI-RADS Features on ChatGPT-5.4-Based Thyroid Nodule Classification.
TL;DR
Clinician-recorded structured ACR TI-RADS features improved ChatGPT-5.4 thyroid nodule classification accuracy when provided alongside ultrasound images, however, the combined strategy did not achieve higher overall accuracy than features-only input.
Abstract
Background
Multimodal large language models (LLMs) have shown potential in medical image analysis, but their performance in the direct interpretation of thyroid ultrasound images remains limited. Whether clinician-recorded structured sonographic features can improve LLM-based thyroid nodule classification is unclear.
Objective
This study aimed to evaluate the effect of clinician-recorded structured ACR TI-RADS features on ChatGPT-5.4 thyroid nodule classification by comparing image-only, features-only, and combined image-and-feature input strategies.
Methods
In this prospective cross-sectional study, 202 thyroid nodules from 153 patients with histopathological confirmation were included. ChatGPT-5.4 Thinking was evaluated under three input conditions: image-only analysis based on transverse and longitudinal grayscale ultrasound images, features-only analysis based on the five clinician-recorded ACR TI-RADS descriptors, and TI-RADS-informed analysis combining the same images with the structured descriptors. Agreement with pathology was assessed using Cohen's kappa, and diagnostic performance was evaluated using sensitivity, specificity, and accuracy. Pairwise comparisons were performed using generalized estimating equations with Holm correction for multiple testing.
Results
Of the 202 nodules, 66 were benign and 136 were malignant. Accuracy was significantly higher with TI-RADS-informed than with image-only input (84.2% vs 73.8%; Holm-adjusted P = 0.001). Features-only input also achieved an accuracy of 84.2%, which was significantly higher than that of image-only input (Holm-adjusted P = 0.005). Compared with the features-only strategy, the TI-RADS-informed strategy had lower sensitivity (91.9% vs 96.3%; Holm-adjusted P = 0.088) but higher specificity (68.2% vs 59.1%; Holm-adjusted P = 0.026). Cohen's κ values were 0.415, 0.606, and 0.625 for image-only, features-only, and TI-RADS-informed analyses, respectively.
Conclusion
Clinician-recorded structured ACR TI-RADS features improved ChatGPT-5.4 thyroid nodule classification accuracy when provided alongside ultrasound images. However, the combined strategy did not achieve higher overall accuracy than features-only input.