Robust Deep Learning for Pulse-Echo Speed of Sound Imaging via Time-Shift Maps

Area of Science:

• Medical Imaging
• Ultrasound Technology
• Artificial Intelligence in Medicine

Background:

• Accurate spatial distribution of longitudinal speed of sound (SoS) is crucial for ultrasound image quality and diagnostic value.
• SoS imaging enables aberration correction and offers a novel contrast mechanism for disease diagnosis.
• Pulse-echo SoS imaging presents significant challenges, often limiting accuracy compared to physics-based methods.

Purpose of the Study:

• To develop a robust deep learning (DL) approach for accurate pulse-echo SoS imaging.
• To learn the nonlinear mapping between measured time shifts and SoS without reliance on specific forward models.
• To enhance SoS imaging performance through various optimization strategies.

Main Methods:

• A DL model was developed to learn the nonlinear mapping from time shifts to SoS.
• Time-shift maps were computed using a common mid-angle configuration, normalized beamformed data, and accounting for depth-dependent frequency.
• The structural similarity index measure (SSIM) was integrated into the loss function for global structure learning.
• A two-stage training strategy involved ray-tracing synthesis for pretraining and full-wave simulations for fine-tuning.

Main Results:

• The DL model demonstrated robustness and generalizability across diverse conditions.
• A simulation-trained model successfully reconstructed SoS maps from experimental phantom data.
• The DL approach improved reconstruction accuracy and contrast-to-noise ratio compared to physics-based inversion methods.
• The developed method showed high accuracy and robustness in phantom experiments.

Conclusions:

• The developed DL approach offers a powerful and accurate method for pulse-echo SoS imaging.
• This technique overcomes limitations of traditional physics-based approaches, improving image reconstruction.
• The findings highlight the potential of DL to advance ultrasound imaging for enhanced diagnostic capabilities.