Quantification of thyroid nodules in multiple ultrasonography systems

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Quantitative ultrasound (QUS) extracts tissue acoustic properties from pulse-echo signals.
• Assessing thyroid nodules requires accurate quantification of acoustic properties like attenuation (ATT) and speed of sound.
• Variability across ultrasound systems poses challenges for consistent quantitative analysis.

Purpose of the Study:

• To introduce QIT-net, a quantitative imaging technique for thyroid nodule assessment.
• To quantify acoustic attenuation (ATT) and speed of sound using multiple ultrasonography systems.
• To develop a robust method that handles variations in ultrasound device characteristics.

Main Methods:

• Employed a CNN-Transformer hybrid architecture to analyze radiofrequency (RF) data, capturing both local and global features.
• Utilized B-mode images as auxiliary input for improved performance in complex anatomical structures.
• Implemented a device-specific fine-tuning strategy and a parameter-sharing network architecture for cross-system training.

Main Results:

• The QIT-net demonstrated effective quantification of acoustic attenuation and speed of sound for thyroid nodules.
• The hybrid architecture successfully integrated local and global feature extraction from RF data.
• Auxiliary B-mode imaging and fine-tuning strategies enhanced robustness across different ultrasound devices.

Conclusions:

• QIT-net provides a reliable quantitative imaging technique for thyroid nodule assessment across diverse ultrasonography systems.
• The proposed deep learning approach effectively quantifies key acoustic properties.
• This method holds potential for improving diagnostic accuracy in thyroid nodule evaluation.