Thyroid nodule and lymph node metastasis assessment from ultrasound images using deep learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Preoperative differentiation of thyroid nodules is crucial for treatment planning.
• Ultrasound (US) examinations can have diagnostic inconsistencies due to variability.
• Computer-assisted diagnosis (CAD) using AI can enhance US diagnostic accuracy.

Purpose of the Study:

• To develop and validate an AI-powered ultrasound diagnostic model.
• To precisely differentiate benign thyroid nodules (BTN), non-metastatic malignant thyroid nodules (NMTN), and metastatic malignant thyroid nodules (MMTN).

Main Methods:

• A multi-center dataset of 209 patients was used, with 195 selected for analysis.
• A convolutional neural network (CNN) AI model was trained on 3,537 preprocessed ultrasound images.
• Model performance was evaluated using ROC curve analysis and a confusion matrix, comparing it with radiologists' diagnoses.

Main Results:

• The AI model achieved high AUC values: 0.97 for BTN, 0.99 for NMTN, and 0.96 for MMTN.
• AI model's diagnostic accuracy (95%) surpassed that of junior (73%) and senior (84%) radiologists.
• The AI model significantly outperformed radiologists in differentiating NMTN.

Conclusions:

• The AI-based ultrasound model demonstrates excellent performance in classifying thyroid nodules.
• This AI tool shows significant value in supporting clinical decision-making for thyroid nodule management.
• AI-powered US offers a promising approach to overcome diagnostic variability in thyroid nodule assessment.