RSNA2021 Redefining Radiology
Daily Bulletin

AI Shows Promise in Assessing COVID-19 Pneumonia

Tuesday, Nov. 30, 2021

By Richard Dargan

A growing body of research shows that algorithms using artificial intelligence (AI) are effective supplements to medical imaging in the assessment of COVID-19 pneumonia, according to a presentation at RSNA 2021.

Rintaro Ito, MD, PhD


Pneumonia related to severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), the virus that causes COVID-19, was first reported at the end of 2019. Since then, there have been scores of cases around the world. Research has shown that COVID-19 pneumonia lasts longer and causes more damage to the lungs than the more common community-acquired version.

AI has shown considerable promise as a tool for analyzing lung images, for instance, in the assessment of CT scans of people with a high-risk for lung cancer. The pandemic presented a unique opportunity to expand on this research.

“COVID-19 pneumonia is the first case in which data was compiled early in the discovery process for unknown diseases and AI was applied to it,” said study lead author Rintaro Ito, MD, PhD, assistant professor in the Department of Radiology at Nagoya University Hospital in Nagoya, Japan. “I think that the use of AI to fight unknown diseases has reached a new stage with COVID-19 pneumonia.”

Dr. Ito and his Nagoya University colleagues, Shinji Naganawa, MD, and Shingo Iwano, MD, performed a meta-analysis of studies that used AI to assess medical imaging performed on COVID-19 pneumonia.

They found that many of the AI studies were focused on differentiating between COVID-19 pneumonia and community-acquired pneumonia and other lung conditions, an important distinction considering COVID-19 pneumonia’s more destructive pathway. The next most common AI studies were focused on extracting regions of COVID-19 pneumonia from the images and quantifying the pneumonia.

“I believe that the imaging AI studies for COVID-19 pneumonia have been very useful,” Dr. Ito said. “These studies have done extremely well with the data provided.”

The researchers found some common limitations in the AI studies. Many were written too quickly and lacked information on how to select, use and analyze datasets. Adapting information learned from published AI to other data did not necessarily produce similar results.

Tools, Data Access Key to Success in AI

Deploying an AI study checklist, such as the Checklist for AI in Medical Imaging (CLAIM) published by Radiology: Artificial Intelligence, would help remedy these shortcomings, according to Dr. Ito.

“The next generation of radiologists will need to use AI properly,” he said. “Knowing about the current AI research trends, development speed and problems will help them become more familiar with AI.”

The evaluation of AI research has been limited by a scarcity of publicly available data, Dr. Ito said. He noted that in AI competitions, AI models are developed using publicly available datasets, but their performance must be evaluated using datasets that are not publicly available.

“If there is an organization that can evaluate AI using data sets that are not publicly available, it will be possible to evaluate AI more fairly,” Dr. Ito said. “Such an organization would be able to provide a fair evaluation of medical AI.”

The education exhibit, “How Should Radiologists Use AI To Fight COVID-19 Pneumonia?” (INEE-8) can be viewed in the Learning Center and online at