New AI Use Case for Detecting COVID-19 With CT Scans
American College of Radiology helps in fight against the SARS-CoV-2 coronavirus.
Posted Mar 15, 2020
Testing for COVID-19 is a critical component in the fight against the global pandemic. Knowing the number of persons infected with the SARS-CoV-2 coronavirus that causes the COVID-19 disease can assist private and public sector leaders in making better decisions based on facts, versus conjecture or estimates.
On March 11, 2020, the American College of Radiology (ACR) Data Science Institute (DSI) released an artificial intelligence (AI) use case for the COVID-19 disease caused by the SARS-CoV-2 coronavirus that enables developers to create machine-learning solutions to detect COVID-19 using chest computed tomography (CT) scans.
According to the March 13, 2020 figures from Harvard Medical School, the exact incubation period for COVID-19 is unknown—it is estimated that symptoms generally appear on average around day five, and can range anywhere between an estimated three to 13 days later. Coronaviruses are mostly thought to be spread through droplets produced from the coughs and sneezes of infected persons. It is also believed that it can be transmitted by contact with infected objects and surfaces, followed by touching certain areas of the face, such as the nose, mouth, and perhaps even the eyes.
Having a way to quickly identify patterns in patient imaging can help in determining if the patient has COVID-19. Artificial intelligence deep learning is well suited to detect patterns in complex data that may be overlooked by the human eye.
The ACR Science Institute constructed the AI use case based on data from a retrospective study of the chest CT scans of 121 symptomatic COVID-19 patients conducted by the Icahn School of Medicine at Mount Sinai, New York, New York, along with researchers from various hospitals in China, that was published in Radiology on February 20, 2020.
The study identified certain common characteristics that present in the imaging of COVID-19 over time. Namely, the researchers found that the "hallmarks of COVID-19 infection on imaging were bilateral and peripheral ground-glass and consolidative pulmonary opacities," and that with increased time after the onset of symptoms, the CT's findings occurred more frequently. These findings include consolidation, linear opacities, "crazy-paving" pattern, the "reverse halo" sign, bilateral and peripheral disease, and greater total lung involvement.
Out of 121 COVID-19 patients studied, 78 percent had ground-glass opacities, consolidation, or both, and only 22 percent had neither of those findings. In terms of opacities in the lobe, 27 percent had it in all five lobes, 15 percent in four lobes, 15 percent in one lobe, 12 percent in two lobes, and 9 percent in three lobes.
Interestingly, the study findings suggest that chest CT scans would be better suited as a companion diagnostic test, versus a standalone tool, given that 56 percent of patients imaged within 0-2 days of symptoms presenting had normal CT scans without any of the hallmark ground-glass opacities and consolidation.
Over three years ago, in November of 2017, the American College of Radiology Data Science Institute announced the initiative to provide an open-source framework available to medical institutions and radiology developers to create AI use cases—providing standards for integrating artificial intelligence algorithms in clinical practice.
For the COVID-19 AI use case, expert radiologists provide a framework that contains considerations for dataset development, such as gender and age (15 years or older, mainly because of no noticeable imaging features on younger patients). Comorbid lung disease considerations include emphysema, bronchitis, bronchiolitis, lung fibrosis, other diffuse lung diseases, malignancy, and immunosuppression. The pathologic diagnosis list Positive Real-Time Reverse Transcriptase Polymerase Chain Reaction (rRT-PCR), and negative STAT rapid influenza/RSV PCR tests. The procedures for CTs include chest scans with or without intravenous contrast (most will be without an intravenous contrast), with or without high-resolution protocols, contiguous thin sections (greater than or equal to 1.5 mm) preferred, and low-dose CT scan.
The scan views include axial supine and multiplanar reformats. Patient history may have fever, cough, dyspnea and synonyms, exposure, and travel history. Associated findings include the presence or absence of pulmonary edema, pleural fluid, pulmonary nodules, and lymphadenopathy. Lung tissue involvement may include ground glass, consolidation, cavitary, segmental, patchy, lobar, multilobar, unilateral, bilateral, diffuse, peripheral, central, crazy-paving, reverse-halo, rounded morphology, interlobular septal thickening, and cavitation.
With the availability of the AI use case for COVID-19, clinicians and radiologists can create artificial intelligence models that are specific to their unique patient population, as well as facilitate the collaboration and contribution towards AI models for diagnostic imaging to fight the global pandemic.
Copyright © 2020 Cami Rosso All rights reserved.