Ultrasound Signal Processing: From Models to Deep Learning.

Area of Science:

• Medical Imaging
• Signal Processing
• Artificial Intelligence

Background:

• Conventional ultrasound reconstruction algorithms rely on physical principles, but their assumptions limit image quality.
• Data-driven deep learning methods offer flexibility but require extensive training data and can be model-agnostic.
• Existing methods face challenges with robustness and environmental sensitivity in medical ultrasound.

Purpose of the Study:

• To provide an overview of model-based deep learning techniques in medical ultrasound.
• To discuss diverse ultrasound applications of these hybrid methods.
• To inspire further research and address opportunities in ultrasound signal processing.

Main Methods:

• Reviewing recent literature on model-based deep learning for ultrasound.
• Analyzing the combination of data-driven deep learning with domain knowledge.
• Exploring various ultrasound imaging applications utilizing these techniques.

Main Results:

• Model-based deep learning solutions demonstrate high robustness in ultrasound imaging.
• These hybrid methods require fewer parameters and less training data compared to conventional neural networks.
• The integration of domain knowledge enhances the performance and reliability of deep learning in ultrasound.

Conclusions:

• Model-based deep learning represents a promising paradigm for advancing medical ultrasound signal processing.
• These techniques offer improved performance and efficiency over traditional and purely data-driven methods.
• Future research in model-based deep learning holds significant potential for the field of medical ultrasound.