By: Eli Freund, Editorial Communications Manager, UConn School of Engineering
Ever since the introduction of IBM’s Watson, companies have been on a voracious quest to use computers and machine learning to help doctors and diagnosticians solve medical quandaries. For a team of five University of Connecticut computer science students, they’re looking to apply those same machine learning principles towards determining whether a patient has COVID-19 or other lung-damaging diseases.
Using a neural network, which is a form of machine learning, a team consisting of computer science students Jamey Calabrese, Jay Chandran, Everett Han, Yuwen Jin, and Adam Veilleux will be inputting x-rays and scans of lungs with COVID-19, tuberculosis and pneumonia, and teaching the computer, based on certain key characteristics of the images, to diagnose which particular lung disease is shown.
“The whole scope of the project is to take images of lung scans of COVID-19, tuberculosis, and pneumonia patients and then train our algorithm to work on a neural network model. Once the system is trained you can input new scans and the neural network will start classifying the scans on its own. This will be a tool that will give doctors and researchers a good confirmation or second opinion,” Chandran said.
To feed their neural net, the team is using scans from different places, including CheXpert, which, according to their website, is a large dataset of chest x-rays and competition for automated chest x -ray interpretation offered by Stanford University.
For the team, their decision to go after building this diagnostic platform comes from a solid batch of research that suggests that using lung scans is one of the best ways to quickly diagnose COVID-19. According to a study done by radiologists at Louisiana State University Health Sciences Center, where they identified common characteristics and compared their diagnosis to a concurrent COVID-19 PCR test, they found that they were able to predict a positive test nearly 84 percent of the time when those common characteristics were used to make a diagnosis.
More specifically, according to one of the radiologists, they found that “the presence of patchy and/or confluent, band-like ground glass opacity or consolidation in a peripheral and mid-to-lower lung zone distribution on a chest radiograph is highly suggestive of SARS-CoV-2 infection and should be used in conjunction with clinical judgment to make a diagnosis.”
But while building this diagnostic platform using a neural net is exciting, the team still faces some challenges in their journey towards Senior Design Demonstration Day.
“One of the biggest obstacles is working with large amounts of data in general and using it in a practical way. One of the things we’ve used used is the UConn cluster. A lot of people are using it, so it’s not much faster than a regular computer. We might have to use the ultra-high-performance cluster at UConn,” Chandran said.
Another pitfall for the team has been finding enough time to work together, especially while balancing classwork and other activities.
“Time management has hit the bottom of my priorities and all my schoolwork takes up the prime real estate of my time, which I’m sure will be elevated next semester. In order to make significant progress, you have to find two workdays in a week, and that’s mostly infeasible,” Calabrese said.
But the team is confident that they can get some significant work done, and are also excited about the range of applications this diagnostic platform may have.
“It’s certainly been interesting creating a platform that can help during this pandemic, but a tool like this is bigger than COVID. I’m a big open-source guy, so if we can advance this towards how we classify diseases, and let other developers build on this, the possibilities for diagnosing illness are endless,” Calabrese said.
This article is part of a multi-part series on engineering students, and their journey through senior design. Part two of this team’s journey will come out in April 2021.
