Thyroid ultrasound image classification using a convolutional neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Ultrasound (US) is a standard diagnostic tool for thyroid nodules (TNs).
• Artificial intelligence (AI) integration in US for TN diagnosis is a growing research area.
• Accurate classification of benign versus malignant TNs is crucial for effective patient management.

Purpose of the Study:

• To develop an enhanced deep learning algorithm for classifying benign and malignant thyroid nodules using US images.
• To improve the diagnostic accuracy of AI in thyroid nodule classification.
• To compare the performance of the developed AI model against experienced radiologists.

Main Methods:

• A Visual Geometry Group (VGG)-16T model was developed, incorporating batch normalization and dropout layers into the VGG-16 architecture.
• The model was trained and validated using an internal dataset of 600 TNs from 592 patients.
• Performance was evaluated on an external dataset of 200 TNs from 187 patients using 10-fold cross-validation.

Main Results:

• The VGG-16T model achieved 87.43% sensitivity, 85.43% specificity, and 86.43% accuracy on the internal dataset.
• On the external dataset, the VGG-16T model achieved an Area Under the Curve (AUC) of 0.829.
• The AI model's AUC (0.829) surpassed that of radiologists with 15 years (0.705), 10 years (0.725), and 5 years (0.660) of experience.

Conclusions:

• The VGG-16T deep learning model demonstrates high accuracy, sensitivity, and specificity in distinguishing between benign and malignant thyroid nodules.
• The AI model's diagnostic performance exceeds that of experienced radiologists.
• This AI-powered tool can serve as a valuable assistant for radiologists in diagnosing thyroid cancer.