Using machine learning to address medical imaging problems
Jun Zhang enjoys explaining his research to non-scientists. In a nutshell, it’s related to artificial intelligence (AI), something most people have heard about—even kids.
“If I’m talking to kids who like video games, I’ll say that AI can help them program a player with super-human game performance, and they find that pretty exciting,” says Zhang, who has been a UWM professor of Electrical Engineering since 1988.
Apart from video games, Zhang’s research has potential applications that range from medical technology to financial investing. A more precise description of what he does is machine learning. The two terms are often used interchangeably, but AI is a much broader field that includes machine learning, which Zhang defines as “a statistical approach for solving three types of problems.”
First, machine learning be used to classify data (the input) into one of several fixed categories (the output), as in face recognition software; second, it can learn input-output relations from the data to make predictions; and third, it can solve complex decision problems traditionally tackled by optimal control theory.
In Zhang’s applied projects, which are supported by Catalyst Grants from GE Healthcare, machine learning techniques are used to address two medical imaging problems.
The goal of the first project is to reduce the scatter in X-ray images that is created by physical structures in the human body or other random factors. Although a grid on the X-ray machine is able to block some of this scatter, it’s expensive and can be difficult to use in some situations. Zhang’s research team develops software that can generate a sharper image at a lower cost.
“We produce images under perfect laboratory conditions that do or don’t have scatter,” he explains. “Then we use machine learning techniques to map, or convert, noisy and blurry images into clearer images.”
The second project involves quality control procedures for computer tomography (CT) detectors. These tiny digital cameras are combined on a panel to generate a high-resolution CT image of human organs. During production, workers test the detectors both individually and as a panel. When the panel test fails, they physically switch the detectors to different positions, which may improve the overall image quality.
Zhang’s research team develops software that can eventually be used to virtually switch CT detectors, which will be much cheaper and faster than the physical switching. As a first step toward that goal, their algorithm predicts the post-switch panel performance from individual tests and already generated pre-switch panel data.
“Machine learning has experienced a revival after people discovered new and better training algorithms and learned how graphics processing units (GPUs) originally designed for video games can be used to implement them,” Zhang says. “With today’s ultra-fast computers, this is an exciting area to work in.”