Hierarchical Temporal Attention Network for Thyroid Nodule Recognition Using Dynamic CEUS Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Contrast-enhanced ultrasound (CEUS) is crucial for thyroid nodule diagnosis, offering real-time vascular visualization.
• Current deep learning methods often overlook the clinical diagnostic dependency, where benign/malignant differentiation precedes pathological typing.
• A hierarchical approach is clinically relevant, mirroring the diagnostic workflow.

Purpose of the Study:

• To propose a novel hierarchical temporal attention network (HiTAN) for thyroid nodule diagnosis using dynamic CEUS imaging.
• To unify dynamic enhancement feature learning and hierarchical nodule classification within a single deep framework.
• To model the native diagnostic dependency in nodule classification.

Main Methods:

• A hierarchical temporal attention network (HiTAN) was developed, decomposing nodule diagnosis into a two-stage classification task.
• Gated Recurrent Units (GRUs) were employed to model the diagnostic dependency between stages.
• A local-to-global temporal aggregation (LGTA) operator with an attention mechanism was designed for comprehensive temporal fusion.

Main Results:

• The HiTAN method demonstrated the efficacy of learning and fusing dynamic enhancement patterns.
• Experimental results validated the effectiveness of the hierarchical diagnosis mechanism.
• The proposed approach successfully integrated dynamic feature learning and classification.

Conclusions:

• The developed HiTAN method offers an effective approach for thyroid nodule diagnosis using dynamic CEUS.
• The hierarchical temporal attention mechanism improves diagnostic accuracy by modeling clinical workflow.
• This study highlights the potential of deep learning in advancing CEUS-based medical image analysis.