Deep learning-based classification of thyroid nodules using uncertainty-aware multi-modal ultrasound imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate thyroid nodule differentiation is vital for early thyroid cancer diagnosis.
• Current deep learning approaches primarily use grayscale ultrasound, limiting diagnostic potential.
• Integrating multiple ultrasound modalities can offer more comprehensive nodule characterization.

Purpose of the Study:

• To enhance thyroid nodule classification performance by combining B-mode, color Doppler (CD), and shear wave elastography (SWE) data.
• To develop and validate a customized deep learning architecture for multi-modal ultrasound analysis.
• To assess the diagnostic superiority of multi-modal imaging over uni-modal and bi-modal approaches.

Main Methods:

• A prospective study included 506 thyroid nodules from 422 subjects.
• A novel deep learning network integrated MobileNetV2 with depth-wise separable convolutions, attention-based pooling, and self-attention.
• An uncertainty-aware fusion strategy was employed for robust multi-modal data integration.

Main Results:

• The multi-modal approach achieved high accuracy (0.95), sensitivity (0.98), specificity (0.92), and F1 score (0.95).
• Area under the ROC curve (AUC) reached 0.97, outperforming uni-modal (AUC 0.73-0.90) and bi-modal (AUC 0.90-0.97) data.
• The proposed network demonstrated comparable or superior performance to state-of-the-art models with a smaller architecture.

Conclusions:

• Multi-modal ultrasound imaging, combined with a customized deep learning network, significantly improves thyroid nodule classification.
• This integrated approach enhances diagnostic performance, offering a more effective tool for thyroid cancer detection.
• The developed network provides an efficient and accurate method for analyzing complex ultrasound data.