Optimizing Thyroid Nodule Evaluation: AI Integration Into the Thyroid Imaging Reporting and Data System Through...

Area of Science:

• Endocrinology
• Radiology
• Artificial Intelligence

Background:

• Thyroid nodules are common endocrine abnormalities requiring risk stratification.
• Ultrasound is the primary diagnostic tool, but ACR-TIRADS faces interobserver variability and time constraints.
• Artificial intelligence (AI) can automate processes and improve diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate an AI model for thyroid nodule risk stratification based on ACR-TIRADS.
• To assess the model's performance in classifying nodules as suspicious or non-suspicious.

Main Methods:

• Retrospective analysis of 139 thyroid ultrasound images (out of 1000 initially collected).
• Annotation of images based on ACR-TIRADS features (composition, echogenicity, shape, margins, echogenic foci).
• Training a vision-language AI model (LLaVA-Med) using domain-specific pretraining and fine-tuning.

Main Results:

• The AI model achieved 67% accuracy, 71% sensitivity, 53% specificity, and 84.6% precision.
• An F1 score of 77% was obtained, with performance favoring sensitivity.
• The model demonstrated potential in reducing the likelihood of missed malignant nodules.

Conclusions:

• The vision-language AI model shows potential as a supportive screening tool for thyroid nodule risk stratification.
• Higher sensitivity and explainable outputs are key advantages, especially in resource-limited settings.
• Further validation and refinement are necessary for clinical implementation.