ThyroidXAgent: Auditable AI for Evidence-Grounded Ultrasound Diagnosis
ThyroidXAgent is an AI system designed for clinician interaction, which integrates specialized diagnostic tools for thyroid ultrasounds and maintains outputs as an auditable record for each case. Utilizing OpenThyroidDB—a comprehensive resource comprising around 0.3 million ultrasound images and 24,000 corresponding reports—the system underwent testing on 28,458 distinct cases, including 8,721 from 35 facilities in the private NHC-MISD-TUS cohort. It recorded a mean Dice score of 87.21% for nodule segmentation and an AUROC of 0.9466 for classifying benign and malignant cases. Additionally, it supports predictions for lymph-node metastasis and assessments of follicular variants, thereby streamlining thyroid diagnosis into a cohesive, auditable workflow that enhances clinical trust and review.
Key facts
- ThyroidXAgent is a clinician-interactive agentic AI system for thyroid ultrasound diagnosis.
- It coordinates specialized diagnostic tools and stores outputs as an auditable case-level evidence record.
- Developed using OpenThyroidDB, a multicentre, multitask resource with ~0.3 million ultrasound images and 24,000 paired reports.
- Evaluated on 28,458 non-overlapping test cases, including 8,721 from 35 centres in the private NHC-MISD-TUS cohort.
- Achieved mean Dice score of 87.21% for nodule segmentation.
- Achieved mean AUROC of 0.9466 for benign-malignant classification.
- Workflow supports lymph-node metastasis prediction and follicular variant assessment.
- Addresses the isolation of AI systems in thyroid diagnosis by integrating tasks into a single auditable workflow.
Entities
Institutions
- arXiv
- NHC-MISD-TUS