A novel ultrasound image diagnostic method for thyroid nodules

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Thyroid nodules are increasingly common, necessitating accurate diagnosis for effective treatment.
• Current ultrasonography diagnosis is subjective, leading to high rates of missed and misdiagnosis.
• Deep learning offers potential to enhance diagnostic accuracy in thyroid nodule classification.

Purpose of the Study:

• To develop and evaluate a novel deep learning diagnostic model for improved classification of benign and malignant thyroid nodules.
• To enhance the accuracy and reliability of thyroid nodule diagnosis using ultrasonography through an AI-driven approach.

Main Methods:

• A localization-classification deep learning strategy was employed, incorporating a multiscale localization network and a two-way classification network.
• An improved attention mechanism was utilized to boost feature extraction performance.
• Fusion of deep features, shallow features, and nodule aspect ratio for final classification.

Main Results:

• The multiscale localization network achieved an accuracy of 93.74%.
• The classification network demonstrated an accuracy of 86.34%, specificity of 81.29%, and sensitivity of 90.48%.
• The proposed model outperformed the TNSCUI 2020 competition's champion model by 1.52% in accuracy.

Conclusions:

• The proposed deep learning model effectively diagnoses benign and malignant thyroid nodules.
• This AI-based approach offers a promising solution to improve the accuracy and reduce subjectivity in clinical diagnosis of thyroid nodules.
• The model's superior performance suggests its potential for clinical application in thyroid nodule management.