Dual-View Thyroid Ultrasound Classification via Dual Knowledge Distillation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Thyroid ultrasound diagnosis commonly uses transverse and longitudinal views.
• Existing deep learning models often process these views independently, limiting diagnostic robustness.
• There's a need for methods that effectively model cross-view semantic consistency.

Purpose of the Study:

• To propose an uncertainty-weighted mixture-of-experts (UMoE) framework for dual-view thyroid ultrasound classification.
• To enhance cross-view semantic consistency and diagnostic robustness in deep learning models.
• To improve the reliability of computer-aided diagnosis for thyroid nodules.

Main Methods:

• A shared Vision Transformer backbone with dual-view and single-view branches was utilized.
• Cross-view knowledge distillation aligned lesion representations using uncertainty-refined patch tokens.
• Cross-level knowledge distillation transferred holistic semantic knowledge from dual-view to single-view branches.

Main Results:

• The UMoE framework outperformed existing single-view and dual-view baselines in accuracy, F1-score, and AUC.
• Experiments were conducted on both in-house (DTN5K) and external public datasets.
• Ablation studies confirmed the effectiveness of the dual-view strategy and individual components.

Conclusions:

• The UMoE framework effectively improves cross-view consistency during training.
• It enhances the predictive ability of single-view branches, leading to more accurate thyroid ultrasound classification.
• This provides a clinically relevant framework for thyroid ultrasound analysis and computer-aided diagnosis.