Diagnostic performance of multimodal ultrasound-based deep learning models in differentiating benign and malignant thyroid nodules
By
Huajie Ding
Lei Na
Meiling Hao
Wanlou Chen
Zhen Zhang
June 29, 2026
Clinical Scorecard: Evaluation of Deep Learning Models Utilizing Multimodal Ultrasound for Distinguishing Between Benign and Malignant Thyroid Nodules
At a Glance
Category Detail
Condition Thyroid Nodules Key Mechanisms Multimodal ultrasound including superb microvascular imaging (SMI) and shear-wave elastography (SWE) enhances diagnostic performance. Target Population Patients with surgically or pathologically confirmed thyroid nodules. Care Setting Clinical practice involving thyroid nodule assessment.
Key Highlights
ResNet50 model achieved the highest diagnostic performance (AUC: 0.931). Study involved 735 patients and 15,373 multimodal US images. Comparison of four DL models: ResNet50, DenseNet121, VGG16, and GoogLeNet. ResNet50 model's accuracy (0.871) was better than junior radiologists (0.810) and comparable to intermediate radiologists (0.886). Grad-CAM visualization indicated focus on clinically relevant thyroid nodule regions. Guideline-Based Recommendations
Diagnosis
Utilize multimodal ultrasound techniques for thyroid nodule assessment. Management
Consider deep learning models for differentiating benign and malignant thyroid nodules. Monitoring & Follow-up
Regularly evaluate the performance of DL models against radiologist assessments. Risks
Potential for misclassification of nodules due to excessive focus on local features. Patient & Prescribing Data
735 patients with thyroid nodules undergoing surgery or biopsy.
Integration of multimodal US data may improve diagnostic accuracy and reduce unnecessary procedures.
Clinical Best Practices
Incorporate advanced imaging techniques like SMI and SWE in routine assessments. Utilize deep learning models to assist radiologists in diagnosis. Related Resources & Content
