Ultrasound deep beamforming using a multiconstrained hybrid generative adversarial network.

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Signal processing

Background:

• Conventional delay-and-sum (DAS) beamforming offers speed but lacks spatial resolution in ultrasound imaging.
• Adaptive beamforming methods improve image quality but introduce high computational complexity, limiting clinical use.
• There is a need for advanced beamformers that overcome spatiotemporal resolution limitations.

Purpose of the Study:

• To introduce a novel deep-learning-based algorithm, the multiconstrained hybrid generative adversarial network (MC-HGAN) beamformer.
• To achieve rapid, high-quality ultrasound imaging by establishing a direct mapping from radio frequency signals to reconstructed images.
• To develop a generalizable beamformer adaptable to various ultrasonic emission modes for clinical applications.

Main Methods:

• Developed a hybrid generative adversarial network (GAN) model with two branches for extracting radio frequency-based and image-based features.
• Integrated features using a fusion module within the hybrid GAN architecture.
• Implemented a multiconstrained training strategy using intermediate network outputs for comprehensive guidance.

Main Results:

• The MC-HGAN beamformer demonstrated superior image quality compared to existing deep learning-based methods.
• Evaluated performance using similarity-based and ultrasound-specific metrics across diverse datasets (line-scan and plane wave modes).
• Showcased high robustness across different clinical datasets and potential for real-time imaging.

Conclusions:

• The MC-HGAN beamformer effectively overcomes spatiotemporal resolution bottlenecks in ultrasound imaging.
• This deep learning approach offers a promising solution for high-quality, computationally efficient ultrasound imaging.
• The algorithm's adaptability and robustness suggest significant potential for widespread clinical adoption.