Skip to main content

Duke Students Taught a Computer to Detect COVID-19 in Lung Scans

A lot of illnesses can look like COVID-19 in a CT scan. Duke students built an AI tool to spot telltale signs of the virus and flag cases that swab tests miss.

Duke students Anmol Warman ’22 and Pranav Warman ’20 have trained a computer to spot the telltale signs of COVID-19 in lung scans and rule out other infections that look similar to the human eye.
Duke students Anmol Warman ’22 and Pranav Warman ’20 have trained a computer to spot the telltale signs of COVID-19 in lung scans and rule out other infections that look similar to the human eye.

Coronavirus may have shut down campuses and closed labs, but that hasn’t stopped some Duke students from brainstorming ways to improve COVID-19 testing while working from home -- with help from artificial intelligence.

A student-led team is building a machine learning system that could help doctors analyze CT scans of people’s lungs and diagnose COVID-19 more quickly and accurately than nasal swab tests.

Their system is able to distinguish COVID-19 from other common infections, such as pneumonia. And unlike many AI tools where it’s unclear how they work, the Duke tool also shows what it “saw” that led to the diagnosis.

Two undergraduates, Anmol Warman ’22, and Pranav Warman ’20, led the effort. The brothers say they first got the idea for the project in early March. By then, the first positive cases of coronavirus were being reported in their home town on the outskirts of Tampa.

But testing bottlenecks made it difficult to determine how widespread the virus was. Supplies were limited, and getting results could take days.

What’s more, the most common coronavirus test -- which looks for viral RNA in cells swabbed from a person’s nose -- identifies only 66% to 80% of people who have the virus.

In other words, it misses up to a third of cases. The test gives these people false reassurance that they don’t have the virus when actually they do.

“That's probably one of the worst things in a pandemic,” said Anmol Warman, a rising junior majoring in biology. “You don't want people thinking that they're safe and actually not being safe.”

Many states have since seen improvements in swab testing supplies, but lingering questions about the accuracy of such tests have pushed researchers to study alternatives, including CT scans.

While normal lungs look black on a CT scan, coronavirus scans often show whitish-gray patches called “ground glass opacities.”

A cross-sectional image of the damaged lungs of a COVID-19 patient on a CT scan. Case courtesy of Dr Fateme Hosseinabadi , Radiopaedia.org, rID: 74868. CT scans were widely used to diagnose COVID-19 patients in China, but they’re not currently recommended as a first-line test in the U.S. The problem is that, to the human eye, these lighter patches on a CT scan can look a lot like other illnesses too, such as tuberculosis or viral pneumonia

The Duke team wondered if artificial intelligence might help. “Could AI help us use all the testing options we have?” said Pranav Warman, who just graduated with a double-major in computer science and biology.

If they could train a computer to tell the difference between COVID-19 and other common infections in CT scans, they reasoned, perhaps researchers could get a better handle on how many people are infected.

The team’s first major hurdle was wrangling enough data. Artificial intelligence needs data to learn from. Most published studies teach AIs to identify patterns based on hundreds of thousands of examples.

Over spring break the students started collecting CT images that had been posted online in the public domain on medical websites such as Radiopaedia.org. Realizing they were going to need expert help, they also asked radiologists at the James A. Haley Veterans' Hospital in Tampa to outline suspicious areas in each image.

They eventually compiled some 900 CT images from 338 patients with known diagnoses. Some patients had COVID-19, some were normal and others had lung damage from other illnesses.

The team fed the images to a neural network, and then set it free to find features common to COVID-19 patients but not in the others.

Other teams have developed AI tools to detect COVID-19 from CT scans. “But what's really cool about our system is that it doesn't just spit out a diagnosis,” Anmol Warman said. It also shows the locations of the features it relied on to arrive at the answer.

The research hasn’t gone through peer review yet; however, the Duke tool performed just as well, if not better, than other AI models that don’t have this transparency.

When they tested it on data from countries and hospitals it had never seen before, the system correctly identified those carrying the virus 98% of the time. Unlike swab tests, actual cases were rarely overlooked. And 95% of the time it also ruled out people with hazy patches in their lungs due to other conditions.

Importantly, their system worked equally well on symptomatic patients and “silent spreaders” -- infected people who don’t have cough or fever but could still be spreading the virus unknowingly, and could become a bigger risk as states reopen.

“I’m proud of our students for taking the initiative to work on this,” said Duke computer vision and medical imaging expert Guillermo Sapiro, who mentored the team.

The students will be presenting their work to representatives from Duke Health this summer. They hope to find out what steps they would need to take to make it useful in the clinic.

“We know that testing is going to be extremely important, especially now that some epidemiologists are suggesting a second wave,” said Pranav Warman, who will be entering Duke’s medical school this fall. “We need to know who has the virus to prevent further spread.”

CITATION: "Interpretable Artificial Intelligence for COVID-19 Diagnosis from Chest CT Reveals Specificity of Ground-Glass Opacities," Anmol Warman, Pranav Warman, Ayushman Sharma, Puja Parikh, Roshan Warman, Narayan Viswanadhan, Lu Chen, Subhra Mohapatra, Shyam Mohapatra and Guillermo Sapiro. medRxiv, May 22, 2020. https://doi.org/10.1101/2020.05.16.20103408