Fully Complex-Valued Gated Recurrent Neural Network for Ultrasound Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Ultrasound imaging is a vital diagnostic tool, but traditional methods like delay-and-sum (DAS) beamformers suffer from low resolution and contrast.
• Adaptive beamformers (ABFs) enhance image quality but often require significant computational resources, hindering real-time application.
• Deep learning offers potential for rapid ultrasound signal processing and image construction.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for improving ultrasound image quality.
• To leverage complex-valued signals and recurrent neural networks for enhanced ultrasound reconstruction.
• To address the computational cost and real-time performance limitations of existing methods.

Main Methods:

• A fully complex-valued gated recurrent neural network (GRU) was proposed for ultrasound image reconstruction.
• The model incorporated complex-number calculations, considering the temporal attributes of ultrasound signals.
• Complex batch normalization, analytic signals, and complex weights were analyzed for model optimization.

Main Results:

• The proposed complex-valued GRU model demonstrated enhanced performance in reconstructing high-quality ultrasound images.
• Analytic signals combined with complex weights significantly improved model performance.
• The model achieved superior results compared to seven state-of-the-art methods in experimental evaluations.

Conclusions:

• The fully complex-valued GRU model represents a significant advancement in ultrasound imaging technology.
• This approach effectively improves image quality while maintaining computational efficiency for real-time applications.
• The findings suggest a promising direction for future developments in AI-driven medical ultrasound.