Development and clinical validation of an artificial intelligence based model for thyroid nodule malignancy risk...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid nodules are common, and early diagnosis of thyroid cancer is crucial.
• Ultrasound is a key diagnostic tool, but interpretation can be subjective and experience-dependent.
• Developing objective diagnostic aids can improve thyroid cancer detection rates.

Purpose of the Study:

• To develop and validate a computer-aided diagnostic framework for thyroid nodule analysis.
• To utilize the Cancer Imaging Reporting and Data System (C-TIRADS) guidelines for standardized risk stratification.
• To assess the performance of an AI model in assisting physicians with thyroid ultrasound interpretation.

Main Methods:

• A C-TIRADS-guided framework was developed, including nodule detection, feature classification, and risk scoring modules.
• The AI model was trained on thyroid nodule datasets and validated on an independent cohort of 303 nodules.
• Key ultrasound features (composition, echogenicity, margin, shape, echogenic foci) were classified.

Main Results:

• The AI model achieved an accuracy of 0.862 in classifying thyroid nodules.
• Physician accuracy improved from 0.705 without AI to 0.845 with AI assistance.
• The system demonstrated potential as a decision-support tool for C-TIRADS classification.

Conclusions:

• The proposed AI framework can localize nodules, classify ultrasound features, and stratify malignancy risk.
• The tool shows promise in assisting physicians with thyroid ultrasound interpretation.
• Further multicenter studies are needed for routine clinical application.