Training Algorithms to Identify COVID-19 in CT Scans

A physician in Wuhan, China, reads a lung CT scan of a Covid-19 patient.

As COVID-19 cases in the United States continue to surge, there is an ongoing search for quick and reliable testing. Though there are conflicting opinions on using chest computerized tomography (CT) scans as a method of screening in the United States, the technology has been successfully used to identify the disease in other parts of the world.

The University of Pittsburgh’s Jingtong Hu received funding from the National Science Foundation (award no. 2027546) to improve CT scan screening by training a computer to analyze images and do the diagnostic heavy lifting.

“Several works have recently demonstrated the potential of deep neural networks in identifying typical signs or partial signs of COVID-19 pneumonia,” said Hu, assistant professor of electrical and computer engineering at Pitt’s Swanson School of Engineering. “Our goal is to train an algorithm to be able to tell the difference between regular pneumonia and the type of pneumonia caused by COVID-19.”

Deep neural networks are sets of algorithms, inspired by the human brain, that recognize patterns and learn to complete a task by training on sample data — in this scenario, CT scans with confirmed pneumonia caused by COVID-19.

Jingtong Hu, assistant professor of electrical and computer engineering at the University of Pittsburgh Swanson School of Engineering

This approach has the potential to drastically speed up the screening process and reduce the burden on radiologists, who are challenged to accurately screen the volume of incoming images. Hu will collaborate with Yiyu Shi, associate professor of computer science and engineering at the University of Notre Dame, to address some of the current limitations with this technology.

“Though deep neural networks work well on 2D images, they are not well equipped to handle the amount of data associated with 3D images,” said Hu. “We will explore various hardware and software solutions that will allow the technology to adapt to these larger files.”

Since deep neural networks require significant processing power, the research group will use a field-programmable gate array (FPGA), which is a high performance integrated circuit that is designed to be reconfigured by the user. This type of device also provides the flexibility needed to handle the evolving structure of neural networks.

“The first step of this project is to create the algorithm and allow it to learn from the current data,” Hu explained. “We will use FPGAs to make the computation faster and more energy efficient. The ultimate goal of this work is to develop a mobile scanning device that can screen for signs of COVID-19 in highly populated places, like an airport or university.”

The diagnostic standard in the U.S. is the RT-PCR test kits, which are not widely available and can take days to deliver results, many of which are inaccurate. Hu and his team hope that the results of this project can help effectively address some of the issues associated with testing in the U.S.

This work will be made open source so that the developed techniques can be applied beyond COVID-19 where neural networks need to handle large volumetric data.

No Comments Yet

Leave a Reply

Your email address will not be published.