Artificial intelligence-based multimodal multitask analysis of thyroid ultrasound image features predicts thyroid...

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Thyroid nodule ultrasound (US) imaging is crucial for diagnosis and clinical decision-making.
• Accurate characterization of US features aids in identifying malignant thyroid nodules.

Purpose of the Study:

• To evaluate an AI-assisted system (MDT-TC) for characterizing thyroid nodule US features.
• To assess the system's ability to assist radiologists in diagnosing thyroid cancer.

Main Methods:

• The MDT-TC system was trained and validated on B-mode US images from 7204 lesions.
• External validation was performed on three independent cohorts.
• Radiomic features, including echogenicity (ECH) and shape (SHA), were incorporated.

Main Results:

• The AI model achieved high accuracy in determining ECH (up to 87.56%) and SHA (69.21%) attributes.
• Area Under the Curve (AUC) values for MDT-TC were strong across internal and external validation cohorts (0.816–0.951).
• Radiologist diagnostic accuracy, particularly for junior doctors, significantly improved with MDT-TC assistance (p < 0.001).

Conclusions:

• The MDT-TC system effectively characterizes thyroid nodule US features.
• AI assistance significantly enhances diagnostic performance for thyroid cancer detection.
• MDT-TC shows potential to benefit clinical practice by improving diagnostic accuracy.