Auditable Agentic AI Improves Thyroid Ultrasound Diagnosis and Reporting

Thyroid ultrasound diagnosis is a complex process involving lesion localization, measurement, risk stratification, and reporting, tasks often handled in isolation by existing AI systems. This research introduces ThyroidXAgent, an innovative clinician-interactive agentic AI system designed to coordinate these specialized diagnostic tools. It generates an auditable, case-level evidence record, providing comprehensive support for clinical review. Developed using OpenThyroidDB, a vast multicenter resource of ultrasound images and reports, ThyroidXAgent was rigorously evaluated on over 28,000 test cases. The system achieved high accuracy, with a mean Dice score of 87.21% for nodule segmentation and a mean AUROC of 0.9466 for benign-malignant classification across diverse datasets. It also supported lymph-node metastasis and thyroid carcinoma classification with strong AUROCs. ThyroidXAgent's evidence-grounded report generation outperformed multimodal language-model baselines, and a new clinical semantic metric, ThyClinScore, showed strong correlation with expert judgment. Clinically, the system improved physician classification accuracy, increased report diagnostic consistency from 70.3% to 86.2%, and significantly reduced segmentation and reporting times by 35.9% and 27.4% respectively, demonstrating the value of auditable, clinician-correctable agentic AI in medical diagnosis.