Stratifying High-Risk Thyroid Nodules Using a Novel Deep Learning System.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Current ultrasound-based thyroid nodule classification is subjective and time-consuming.
• Artificial intelligence (AI) shows promise in enhancing the accuracy of malignancy prediction for thyroid nodules.
• The Swin Transformer (Swin-T) is a state-of-the-art AI model for image classification.

Purpose of the Study:

• To evaluate the efficacy of the Swin Transformer model in classifying thyroid nodules using ultrasound images.
• To compare the performance of Swin-T against the ResNeSt50 model for thyroid nodule classification.

Main Methods:

• Prospective collection of ultrasound images from 139 malignant and 235 benign thyroid nodules (Jan 2016-June 2021).
• Classification of thyroid nodules using Swin-T and ResNeSt50 AI models.
• Performance evaluation using sensitivity, specificity, receiver operating characteristic (ROC) analysis, and McNemar test.

Main Results:

• Swin-T achieved higher average sensitivity (82.46%) and specificity (84.29%) compared to ResNeSt50 (72.51% and 77.14%).
• Swin-T demonstrated a superior area under the curve (AUC=0.91) versus ResNeSt50 (AUC=0.82).
• The McNemar test confirmed significantly better performance for Swin-T.

Conclusions:

• The Swin Transformer model offers a more accurate and reliable method for classifying thyroid nodules from ultrasound images.
• Swin-T can serve as a valuable tool to support shared decision-making between physicians and patients regarding thyroid nodule management.
• This AI approach is particularly beneficial for nodules with high-risk sonographic features.