Deep Learning Technology for Classification of Thyroid Nodules Using Multi-View Ultrasound Images: Potential Benefits...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid nodule classification relies on accurate imaging and cytopathology.
• Deep learning offers potential for automated analysis of medical images.
• Evaluating deep learning for thyroid nodule classification is crucial.

Purpose of the Study:

• To assess the effectiveness of deep learning algorithms for classifying thyroid nodules on ultrasound images.
• To compare the performance of different convolutional neural network (CNN) architectures.

Main Methods:

• Retrospective analysis of 1,048 thyroid nodules from 943 patients.
• Utilized CNNs (ResNet, DenseNet, EfficientNet) and Siamese networks for multi-view analysis.
• Defined thyroid cancer based on Bethesda categories V (suspicious) and VI (malignant).

Main Results:

• Deep learning models achieved high accuracy in thyroid nodule classification.
• Longitudinal images demonstrated superior prediction ability compared to transverse images.
• Multi-view analysis using paired images significantly improved model performance, with accuracies up to 0.83.

Conclusions:

• CNN algorithms show significant accuracy in classifying thyroid nodules from ultrasound images.
• Deep learning holds promise as a tool for thyroid cancer diagnosis.
• Clinical performance may be influenced by image quality and acquisition variability.