Modern medical ultrasound offers a relatively safe and noninvasive way to view tissues inside the body and serves as a guide for procedures, such as tissue biopsies. However, the algorithms used to form images from sound waves rely on a bevy of assumptions, such as the speed of sound in tissue, the density of tissues, and the number of times sound waves are reflected in the body. Because these assumptions aren’t constant for all body types, obtaining accurate images can be difficult for some individuals, such as those who are obese or who have unusually high ratios of fat to muscle.
Muyinatu Bell, an assistant professor of electrical and computer engineering, and her colleagues are developing a machine-learning approach to solve this problem. Their strategy involves training computers to see only structures of interest—say, a needle tip and kidney cyst for a drainage procedure—extracting out all “noisy” background material.
Their work thus far, training a system that uses simulated data and incorporates hundreds of speeds of sound, tissue densities, and needle locations, and then applies the training to real data, shows early promise in identifying correctly structures of interest. Recently, Bell received a Trailblazer Award from the National Institutes of Health for the progress she has made thus far.