Advances in high-resolution ultrasound imaging with deep learning

Researchers at the Beckman Institute for Advanced Science and Technology have developed a new technology that makes ultrasound localization microscopy, an emerging diagnostic tool for high-resolution microvascular imaging, more practical in clinical settings. Their approach uses deep learning to advance ULM’s post-processing pipeline.

Their technique, called localization with context-hour ultrasound localization microscopy, or LOCA-ULM, appears in the journal Communicate with nature.

“I’m interested in making ULM faster and better so that more people can use this technology. I think deep learning computational imaging-based tools will continue to play an important role in limiting the spatial and temporal resolution of ULM.” “said first author Yirang Shin, a graduate student in the Department of Electrical and Computer Engineering at the University of Illinois at Urbana-Champaign.

Ultrasound localization microscopy works by injecting microbubbles into blood vessels to act as contrast agents. Microbubbles have been approved by the FDA for clinical use. Ultrasound waves can penetrate deep into the tissues of the body and pinpoint the location of these microbubbles (each only a few microns in size) as they travel through the blood. Researchers use microbubbles to track blood flow velocity and create spatial images of blood vessels on a microscopic scale.

The current imaging speed of ULM limits its practical application as a diagnostic tool in the medical field and as a basic scientific research tool. Improving imaging speed requires a higher concentration of microbubbles in blood, which makes post-processing more difficult, Shin said.

The researchers’ new method demonstrates improved imaging performance and processing speed, greater sensitivity of functional ULMs, and overall superior in vivo imaging. It exhibits improved computational and microbubble localization performance and is applicable to different microbubble concentrations.

“It really beats traditional microbubble localization methods; That’s what works,” said Pengfei Song, a Beckman Fellow, YT Low Faculty Fellow and assistant professor in Illinois’ Department of Electrical and Computer Engineering.

To create microbubbles localization faster, more accurate and more efficient, researchers developed a simulation model based on generative adversarial network (GAN). This simulation creates a realistic microbubbles deep context-aware neural network DECODE training signal.

Carl R. of Pengfei Song Illinois. Voss is an assistant professor of bioengineering, biomedical and translational sciences at the Institute for Genomic Biology.

Other co-authors include Matthew R. Lorison, Yek Wang, Shi Chen, Qiu Yu, Zhijie Dong, and Mark A. Anastasio.