Simulation-to-real generalization for deep-learning-based refraction-corrected ultrasound tomography image...

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence

Background:

• Ultrasound computed tomography (UCT) image reconstruction is computationally intensive with iterative methods.
• Deep learning offers faster direct reconstruction but struggles with the simulation-to-real data gap.
• Lack of real labeled data hinders neural network training for UCT.

Purpose of the Study:

• To enhance the simulation-to-real generalization of deep learning models for UCT.
• To develop strategies for improving neural network performance on real UCT measurement data.
• To enable faster and more accurate UCT image reconstruction using deep learning.

Main Methods:

• Developed a Fourier-transform-integrated neural network.
• Implemented measurement-domain data augmentation techniques.
• Utilized a self-supervised learning-based patch-wise preprocessing neural network.
• Evaluated methods on simulation and real measurement datasets from two prototype machines.

Main Results:

• Deep learning methods improved neural network robustness against noise.
• Enhanced generalizability to real measurement data was demonstrated.
• Superior imaging quality compared to traditional iterative methods was achieved.
• Real-time 2D-image reconstruction capability was enabled.

Conclusions:

• The developed strategies effectively bridge the simulation-to-real gap in UCT deep learning.
• Deep learning models can outperform conventional iterative reconstruction algorithms in UCT.
• This research facilitates the practical application of deep learning for UCT image reconstruction using simulation data.