Although the initial wave of the SARS-CoV-2 pandemic has abated in many countries, healthcare providers are still looking to identify as many COVID-19 patients as possible and contain the disease. Fast and accurate diagnosis is especially important when unsuspecting patients with a coronavirus infection come to the hospital with health complaints but don’t yet show symptoms of COVID-19.
Nasal swab samples analyzed by RT-PCR are currently recommended for the diagnosis of COVID-19, however, supply shortages, a wait time of up to two days for results, and a false negative rate as high as 1 in 5 mean alternative, large-scale COVID-19 screening tools are still being sought.
SARS-CoV-2 is known to damage lung tissue, and in a distinct way that doctors are now seeking to exploit for new diagnostic approaches. Many COVID-19 patients develop pneumonia, which can progress to respiratory failure and sometimes death. COVID-19 pneumonia is different from more common forms of bacterial pneumonia, and the differences show up in chest CT scans. Most striking are cloudy lesion patterns that resemble shards of glass or reticular lines within the opaque lesions that look like irregular paving tiles, which occur around the peripheries of both lungs. Lesions from bacterial pneumonia are usually concentrated in one lung and may not resemble shards of glass.
In China, CT scans are already used as a COVID-19 diagnostic tool when a patient arrives at a healthcare setting with fever and a suspected infection, though this approach has not been widely adopted in the United States. Two studies, published in Nature Medicine and Cell, advance this idea by using artificial intelligence (AI) trained on CT lung scans as a quick diagnostic tool to look for COVID-19 infection in patients who come to the hospital and require medical imaging.
Writing in Cell, researchers at Macau University of Science and Technology used 532,000 CT scans from 3,777 patients in China to train their AI tools, focusing on the tell-tale lesions seen in COVID-19 patient lungs. In pilot studies at several Chinese hospitals the AI model correctly diagnosed pneumonia caused by the coronavirus at least 85 percent of the time when it was applied to a dataset of 417 patients in four separate cohorts. COVID pneumonia was misdiagnosed as non-COVID pneumonia in 7–12 percent of cases.
“This group does a tremendous job of a deep-dive on external validation: they have this large dataset from China and they looked at how it performed in many hospitals,” says Matthew Lungren, a radiologist at Stanford University Medical Center who was not involved in either study.
Recognizing a very small number of COVID-19 pneumonia cases out of a large number of non-specific general pneumonia cases is important for a diagnostic tool when SARS-CoV-2, the coronavirus behind the pandemic, becomes endemic and is no longer the leading cause of pneumonia, Lungren explains.
“A large dataset with a diverse source of data is crucial to achieve robust and generalizable conclusions in AI based diagnoses,” writes Cell coauthor Kang Zhang, a professor of medicine at Macau University of Science and Technology, in an email to The Scientist. “One of the most challenging issues in AI application in healthcare is poor reproducibility.”
One challenge of using CT scans for COVID-19 diagnosis is that a lot of people infected with SARS-CoV-2 experience severe clinical symptoms such as cough and fever, but have no biomarkers visible in the CT scans. If healthcare professionals are trying to get an accurate COVID-19 diagnosis faster than standard PCR methods, “only basing [COVID-19 diagnosis] on imaging may not be enough,” says Yang Yang, a radiologist at Mount Sinai Hospital.
Yang’s team also trained its COVID-19 AI model on CT chest scans and published the results in Nature Medicine. This model integrated results of the CT scans with clinical findings such as patients’ ages, whether they had a cough or fever, and their white blood cell counts, creating what the authors called a “fusion model” to diagnose patients with COVID-19 based on clinical and imaging data. Their fusion model diagnosed COVID-19 with 83.5 percent accuracy in a test set of 279 patients. When looking at the same set of images, a senior thoracic radiologist diagnosed COVID-19 with 84.6 percent accuracy.
“There are aspects to their methodology which I think are very important for this field in general,” says Lungren, namely, many AI diagnostic models based upon imaging data would benefit from the input of additional clinical data.
Zhang says at least 10 large hospitals in China, and several in the US, India, Iraq, and Ecuador are using his model to diagnose patients suspected of having COVID-19 pneumonia. His team made its algorithms and training datasets publicly available for other researchers to use.
X. Mei et al., “Artificial intelligence–enabled rapid diagnosis of patients with COVID-19,” Nature Medicine, doi:10.1038/s41591-020-0931-3, 2020.
K. Zhang et al., “Clinically applicable AI system for accurate diagnosis, quantitative measurements and prognosis of COVID-19 pneumonia using computed tomography,” Cell, doi:10.1016/j.cell.2020.04.045, 2020.