Artificial intelligence (AI) can diagnose COVID-19 from CT scans, researchers in China claim. At least two teams have released studies that they say demonstrate that deep learning can analyse radiological features for accurate COVID-19 diagnosis faster than current blood tests, saving critical time for disease control.
COVID-19 first appeared in Wuhan in China at the end of last year and has since spread across the globe. By early March the World Health Organization had declared the outbreak a pandemic, and there have now been over 130,000 confirmed deaths worldwide.
Presentation of the viral disease ranges from asymptomatic to severe pneumonia with acute respiratory distress and multiple organ failure. COVID-19 is typically diagnosed by a reverse-transcription polymerase chain reaction (RT-PCR) test on a blood sample, but there are concerns about the sensitivity and availability of tests, and turnaround times for results.
CT scans are reported to be able to detect characteristic manifestations of COVID-19 in the lungs, with claims that this can lead to faster diagnosis than with current RT-PCR tests. But COVID-19 also shares similar imaging features with other types of pneumonia, making it difficult to differentiate. Could deep-learning rapidly analyse images and identify features of COVID-19?
To develop an AI tool to detect COVID-19, researchers led by Bo Xu of the Tianjin Medical University Cancer Institute and Hospital took CT images from 180 individuals who were diagnosed with typical viral pneumonia before the COVID-19 outbreak and 79 patients with confirmed COVID-19. They randomly assigned images from the patients to train or test the deep-learning algorithm.
In results published on medRxiv, the researchers claim that their model identified COVID-19 from CT images with an accuracy of 89.5%. Two radiologists who also assessed the images achieved an accuracy of around 55%. The team say the results demonstrate that AI can offer accurate diagnosis from a CT scan (medRxiv 10.1101/2020.02.14.20023028).
In other work, published in Radiology, another team from China, led by Jun Xia of the Wuhan Huangpi People’s Hospital, trained a deep-learning model to detect COVID-19 using chest CT scans from 400 patients with COVID-19, almost 1400 people with community acquired pneumonia and more than 1000 people without pneumonia (Radiology 10.1148/radiol.2020200905).
When they tested their AI on CT images from 450 patients, 20% with COVID-19, it achieved an accuracy of around 90%. Again, the researchers say that this shows that deep learning can differentiate COVID-19 from community acquired pneumonia and other lung diseases.
Not everyone agrees, however. Michael Lu, a radiologist and expert in AI for imaging at Massachusetts General Hospital, tells Physics World that you first need to consider whether CT is good for diagnosing COVID-19. “The current US stance is not to do CT as the primary diagnostic test,” he explains.
“There was some initial data out of China saying that CT had high sensitivity, was accurate for detection of COVID-19,” Lu explains. “Subsequent to that, there’s been a lot of discussion of that paper, concern that there may have been selection bias involved. With any of these tests, the pre-test probability, sort of the prevalence or what stage of the COVID-19 pandemic you’re in, can actually have a major impact on the apparent performance of the test.”
Europe seeks to harness power of AI in COVID-19 crisis
Although he is keen to stress that the researchers are working rapidly with available data, Lu is also concerned that the mix of COVID-19 cases and historic lung images used to train and test the AI models does not reflect what a doctor dealing with suspected COVID-19 patients would see. “It’s an artificial distribution of patients,” he explains.
There are also issues around the practicalities and expense of mass testing using CT scans. “One of the issues is that, at least currently in the US, the scanners have to be cleaned after a patient comes in,” Lu says. In China, they have reportedly started covering people with plastic bags to prevent virus transmission between patients.
Lu believes that PCR tests will improve and will remain the gold standard for COVID-19 testing. He adds, however, that we should still pay attention to work on CT scans coming out of China, as they are ahead of the rest of the world in dealing with the pandemic, and things could change.