Learnability of Thyroid Nodule Assessment on Ultrasonography: Using a Big Data Set

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Thyroid nodule assessment via ultrasonography (US) is crucial for differentiating benign and malignant cases.
• Training novice readers requires effective methods and tools to achieve diagnostic proficiency.

Purpose of the Study:

• To assess the learnability of thyroid nodule interpretation using a large dataset of US images.
• To evaluate the diagnostic utility of AI-CAD for readers with diverse experience levels in thyroid imaging.

Main Methods:

• Six college freshmen underwent training using a large dataset (13,560 images) of thyroid nodules.
• Diagnostic performance was assessed pre- and post-training on a separate test set (282 images).
• Readers' performance was compared with experienced radiologists and further evaluated with AI-CAD assistance.

Main Results:

• Most students showed improved diagnostic performance after the training program.
• While experienced readers initially outperformed students, five students achieved comparable results to one resident.
• AI-CAD significantly enhanced sensitivity, accuracy, and AUROC for several students and improved performance for all experienced readers.

Conclusions:

• Self-learning with big data US image sets can supplement traditional training for thyroid nodule assessment.
• AI-CAD offers a valuable tool to augment the diagnostic capabilities of readers across different experience levels in thyroid imaging.