Bidirectional interaction directional variance attention model based on increased-transformer for thyroid nodule...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate classification of thyroid nodules is crucial for cancer diagnosis.
• Ultrasound image noise and subtle contour/texture differences challenge existing models.
• Low classification accuracy hinders effective diagnosis of benign versus malignant nodules.

Purpose of the Study:

• To develop an advanced deep learning model for precise thyroid nodule classification.
• To enhance the model's ability to overcome noise and subtle feature variations in ultrasound images.
• To improve diagnostic accuracy for differentiating benign and malignant thyroid nodules.

Main Methods:

• Proposed the Increased-Transformer (IFormer) for global feature modeling and noise reduction.
• Introduced the Bidirectional Interaction Directional Variance Attention (BIDVA) module for focused feature extraction.
• Developed a Multi-Dimensional Loss (MD Loss) function to improve class separation and handle imbalance.

Main Results:

• IFormer-DVNet achieved 76.55% accuracy on the TNCD dataset and 93.02% on a private dataset.
• The model demonstrated superior performance across all evaluation metrics compared to state-of-the-art networks.
• The Increased-Transformer effectively mitigated noise interference, and BIDVA enhanced focus on informative regions.

Conclusions:

• IFormer-DVNet offers a significant advancement in automated thyroid nodule classification.
• The proposed architecture effectively addresses challenges posed by ultrasound image noise and subtle diagnostic features.
• This model holds promise for improving the accuracy and efficiency of thyroid cancer diagnosis.