Researchers at Notre Dame are using artificial intelligence to improve CT screening efficiency and relieve the burdens of the radiologists in diagnosing patients with COVID-19.
Yiyu Shi, a professor of engineering and computer science, is a principal investigator of this research project.
“Essentially, it’s just an AI software program that computes based on an input, which is an [CT] image and it will produce an output, which is the prediction result being positive or not,” Shi said.
Weiwen Jiang, a postdoctorate in the department of computer science and engineering and co-principal investigator, said there were two main phases involved with developing the AI model.
“The first phase is called the training phase, where we use training datasets online to train our model,” he said. “After the training phase, we will go through the testing process which is the second phase. In this phase, we use the model to make predictions based on a CT image from a real case.”
The shortage of PCR tests when COVID-19 first began motivated Jiang to improve the screening efficiency using his expertise in machine learning algorithms.
“This pandemic made a very bad impact worldwide, so we would like to make some contribution to [combat] such a crisis,” Jiang said.
Shi said he was inspired to start working on this research project after witnessing the complicated process of running a CT screening for radiologists.
“In my research, I work extensively with doctors and radiologists at different hospitals, and one feedback I got from them with the coronavirus is that they’re overloaded with all the CT images to read from different locations and then see if it's positive or not,” Shi said. “Those CT images are 3D and there are multiple slices, so then they have to look at each individual one to make a decision ... to make a diagnosis and that takes a long time.”
Shi and his research team has already finished the AI network, and are currently working to transfer the network to a flash drive.
“We already have a network now that can give pretty accurate predictions,” he said. “We are just trying to put that network into a plug-and-play USB stick so that when a doctor needs to use that assistance, he/she can just plug the USB stick into a computer and then run the program, and it [the network] will give the results in just a few minutes.”
The biggest challenge for Shi and his team has been trying to maximize the computational power of the algorithm while not encroaching on the privacy of patients.
“Tthis whole network takes a lot of computation in order to get to the final prediction,” Shi said. “Now, if this whole thing can be put on Amazon Cloud, or cloud service, then it can be pretty straightforward and easy. But, the problem is that we don’t want any patient’s data to be uploaded to the cloud, which would cause a lot of problems. So ideally, the prediction needs to be done locally on a computer and that will have very limited computational power.”
In addition to faculty and graduate students, five undergraduate students have also contributed to the research, Shi said.
“They are in my computer engineering capstone design class, and they chose to use this project as their capstone project, which is great,” he said. “So it’s part of their coursework and at the same time, getting something very useful for the community.”
Currently, the research team is working on perfecting the current design and making the model efficient.
“We’re basically still working on compressing the model and we will make some adjustments next month,” graduate student Dewen Zeng said.
Shi’s research team has been progressing fast on their timeline, and hopes to have the whole project ready for a wide release later this fall.
“We expect that we’ll be able to run the program on the USB stick by the end of this month or maybe early October,” Shi said. “And we’ll spend maybe a month, or so, trying to optimize the performance and the computation on a USB stick to make it faster and make it more accurate on that USB stick where we have very limited resources. And hopefully, we’ll be able to release the results. It’s going to be open source. Anyone can use it when we release it in mid-November.”
