As one of the first radiologists in the country to examine CT scans from patients in China infected with the virus now known as SARS-CoV-2, Michael Chung, MD, was as intrigued by what he did not see in those images as he was by the lung characteristics he observed.
“Although pneumonia is a key driver in this disease, the characteristics we typically see—lung cavitation, discrete pulmonary nodules, pleural effusions, and lymphadenopathy—were absent, especially in the early stages of COVID-19,” says Dr. Chung, Assistant Professor of Radiology, and Medicine (Cardiology), at the Icahn School of Medicine at Mount Sinai. “Coupled with the characteristic features we did see—bilateral and peripheral ground-glass and consolidative pulmonary opacities—that was really beneficial in helping to formulate our differential diagnosis and in discriminating COVID-19 from other infections.”
From the outset of the pandemic, Mount Sinai radiologists such as Dr. Chung have played an invaluable role in advancing our understanding of lung involvement in COVID-19. But even more important, they have demonstrated through several studies that thoracic imaging—radiology and CT scans—can play a complementary role to real-time reverse transcriptase polymerase chain reaction (rRT-PCR) in the diagnosis, prognosis, and triaging of patients with COVID-19, even more so when augmented by artificial intelligence.
“There has been some question as to, for example, how much value the initial emergency department chest radiograph holds and does it dictate, or correlate, with prognosis or patient outcomes,” says Adam Bernheim, MD, Assistant Professor of Diagnostic, Molecular, and Interventional Radiology at the Icahn School of Medicine at Mount Sinai. “Through our studies, we have been able to demonstrate that thoracic imaging is foundational in the management of patients with COVID-19 and in evaluating disease progression and complications.”
The first study, “CT Imaging Features of 2019 Novel Coronavirus (2019-nCoV),” sourced CT scans from confirmed COVID-19 patients from three hospitals in different provinces of China. Three of the patients initially presented with a normal CT scan, 15 (71 percent) had involvement of more than two lobes, 12 (57 percent) had ground-glass opacities, 7 (33 percent) had opacities with a rounded morphology, 7 (33 percent) had a peripheral distribution of disease, 6 (29 percent) had consolidation with ground-glass opacities, and four (19 percent) had crazy-paving pattern. The results were published in the February 4, 2020, edition of Radiology.
These hallmarks of infection and distribution were subsequently confirmed in a second, larger, retrospective study, “Chest CT Findings in Coronavirus Disease-19 (COVID-19): Relationship to Duration of Infection,” which was published in the February 20, 2020, edition of Radiology. Looking at the relationship between symptom onset and initial CT scan, defined as early (0 to 2 days), intermediate (3 to 5 days), and late (6 to 12 days), Drs. Bernheim and Chung noted that bilateral and peripheral ground-glass and consolidative pulmonary opacities were less frequent among the early cohort than among the intermediate and late cohort. Also notable was that one patient in the early group presented with an initially negative rRT-PCR, which suggests that a chest CT can be falsely negative in some patients who go on to test rRT-PCR-positive, particularly if performed very shortly after symptom onset.
“We were able to show that CT scans provide helpful information in regards to COVID-19 infection, but they have limited sensitivity at the onset of the disease, which means you could have a false negative,” Dr. Bernheim says. “Furthermore, over time, the disease became more extensive and likely to affect both lungs, and the hallmarks we look for as radiologists presented themselves more readily in that later phase.”
For the third study, “Clinical and Chest Radiography Features Determine Patient Outcomes in Young and Middle Age Adults with COVID-19,” Drs. Bernheim and Chung assessed the prognostic value of a chest radiology (CXR) severity scoring system. Using a cohort of younger patients (median age 39) who presented at Mount Sinai emergency departments with rRT-PCR confirmation of COVID-19 infection, Drs. Bernheim and Chung divided each patient’s CXR into six zones, examined them for opacities, and collated scores into a total concordant lung zone severity score. After adjusting for demographics and comorbidities, they found that a CXR ≥ 2 was a predictor of hospital admission and, among those admitted, a CXR ≥ 3 was a predictor of intubation. The results were published in the May 14, 2020, edition of Radiology.
“Because chest radiography is less expensive, faster, and safer to obtain than CT scans, it was the main method used nationwide to gauge patients who presented with a COVID-19 infection,” Dr. Chung says. “This study demonstrated that a chest X-ray is an easy, quick way to stratify disease severity and thus it has considerable utility in the setting of pandemic management.”
Drs. Bernheim and Chung subsequently collaborated with other Mount Sinai researchers on a fourth study, “Artificial Intelligence-Enabled Rapid Diagnosis of Patients with COVID-19,” to see if AI could enhance the diagnostic value of thoracic imaging, specifically CT scans. Three AI models—CT scan, clinical information, and a joint model—were developed and evaluated using a test set of COVID-19-positive and negative cases from China. Zahi Fayad, PhD, Professor of Radiology, and Medicine (Cardiology), at the Icahn School of Medicine at Mount Sinai, and Director of the Mount Sinai BioMedical Engineering and Imaging Institute, says the joint model achieved an area under the curve of 0.92 and had comparable sensitivity to that of a senior thoracic radiologist. The joint model also improved detection of patients who were COVID-19-positive via rRT-PCR yet presented with normal CT scans, correctly identifying 17 of 25 patients (68 percent), all of whom were classified as COVID-19-negative by radiologists. The results were published May 19, 2020, in Nature Medicine.
“The bottom line is that we were able to demonstrate that we can use AI to combine the CT data with basic clinical information and achieve improvements in sensitivity and specificity,” Dr. Fayad says. “Our hope is that this approach can be deployed in resource-challenged environments during the peak of a pandemic when the hospital system is overwhelmed.”
Michael Chung, MD
Assistant Professor of Diagnostic, Molecular and Interventional Radiology, and Medicine (Cardiology)
Adam Bernheim, MD
Assistant Professor of Diagnostic, Molecular and Interventional Radiology
Zahi Fayad, PhD
Professor of Diagnostic, Molecular and Interventional Radiology, and Medicine (Cardiology)