Deep Learning Aids Development of Super-Resolution Ultrasound

Researchers at the Beckman Institute for Advanced Science and Technology used deep learning to develop a new framework for super-resolution ultrasound.

Traditional super-resolution ultrasound techniques use microbubbles: tiny spheres of gas encased in a protein or lipid shell. Microbubbles are considered to be a contrast agent, which means they can be injected into a blood vessel to increase the clarity of an ultrasound image.

Conventional ultrasound has been commonplace for over 50 years. The development of super-resolution technology in the last decade has introduced new challenges. Super-resolution ultrasound provides a much clearer picture than the traditional method. Although useful for research and diagnostics, its processing speeds are much slower.

"Ultrasound is expected to be a real-time imaging modality.”

This challenge prompted Song to team up with fellow Beckman researcher Dr. Daniel Llano, an associate professor of molecular and integrative physiology and a neurologist at Carle Foundation Hospital. Together, the researchers tested a new approach to super-resolution ultrasound technology.

Their paper appears in IEEE Transactions on Medical Imaging.

“As engineers, we develop tools that we think will be useful for researchers, but sometimes we miss the mark,” Song said. “This is a case where the user of the technology, like Professor Llano, tells us how we have to improve the technology: make it faster.”

Traditional super-resolution ultrasound techniques produce crisp, vibrant images, but the process is lengthy because it requires a very low concentration of microbubbles. For researchers like Llano, every minute counts.

In response to Llano’s feedback, the Song group returned to the drawing board and decided to revamp the super-resolution technology, forgoing microbubble localization and tracking entirely. Instead of evaluating data frame by frame, the researchers used a holistic approach and evaluated the information spatiotemporally — over space and time. Using an artificial intelligence network, the technology was able to determine the speed of blood flow and convert a blurred image to a clearer one with a high resolution.

Because conventional super-resolution ultrasound is so slow, the end product is likened to a still image. But with the researchers’ new method, blood flow can be visualized in real time.

“To the best of our knowledge, this is the first paper that achieved direct calculation of the blood flow velocity, both speed and direction, using raw ultrasound data without any explicit microbubble localization or tracking,” Song said.

As a result, processing speeds have been reduced from minutes to seconds, and post-processing can be done in real time. The researchers hope that speeding up the higher-resolution technology will make it a useful option for clinicians.

“We’ve done human imaging before with conventional imaging, but it’s challenging,” Song said. “We think that this technique has the potential for super resolution to be finally used in a clinical setting.”

Collaboration between the two research groups was made possible through the shared environment at Beckman.

“Professor Llano’s home department is [molecular and integrative physiology], so without Beckman this collaboration would not have been possible, because we need to have a common lab space,” Song said. “It’s really the common physical space that made this happen.”

Reference: Chen X, Lowerison MR, Dong Z, Sekaran NVC, Llano DA, Song P. Localization free super-resolution microbubble velocimetry using a long short-term memory neural network. IEEE Trans Med Imaging. 2023:1-1. doi: 10.1109/TMI.2023.3251197

This article has been republished from the following materials. Note: material may have been edited for length and content. For further information, please contact the cited source.