Home> News

AI model can tell COVID-19 from flu and other diseases

Updated: 2020-10-20 (China Daily)

5f8ce473a31024adbd9aa608.jpg

A visitor experiences artificial intelligence equipment at an exhibition in Yangzhou, Jiangsu province, on April 28. [Photo by Meng Delong/For China Daily]

Chinese researchers published a paper in the journal Nature Communications this month proposing an artificial intelligence model that can help doctors quickly differentiate between COVID-19, influenza and pneumonia with high accuracy.

Since the COVID-19 outbreak, numerous AI systems have been developed and used for front-line detection and diagnosis, such as analyzing chest X-rays and CT scans. However, with flu season approaching, if COVID-19 and influenza were to break out together, causing the CT diagnosis workload to skyrocket, differentiating between the two respiratory illnesses would prove challenging for doctors.

A new AI model may provide the answer. Researchers from Tsinghua University and Union Hospital in Wuhan, Hubei province, which is affiliated with Huazhong University of Science and Technology, have developed and evaluated an AI system using a large data set with more than 11,000 CT volumes from cases of COVID-19, influenza, non-viral community-acquired pneumonia, and non-pneumonia.

According to the paper, CT volumes of COVID-19 patients were collected mainly from February to March at three hospitals in Wuhan, once the epicenter of the COVID-19 pandemic in China.

The AI model, known as a deep convolutional neural network-based system, turned detection experiences accumulated by experts into algorithms. Test results showed it can differentiate between four respiratory diseases, including COVID-19, influenza and non-pneumonia, with a high degree of accuracy.

In further studies, the research team compared the diagnostic performance of the CT-based AI system with that of five radiologists, and results showed the system performed better than its human counterparts.

The AI model will help lessen the workload of doctors. The study showed that the average reading time for radiologists was six and a half minutes, while that of the AI system was 2.73 seconds.

The paper said the AI system was only slightly worse at distinguishing pneumonia from nonpneumonia than radiologists.

Using CT lung screening to differentiate COVID-19 from other forms of pneumonia is difficult due to the many similarities of the different types of pneumonia, especially in the early stages, and large variations in different stages of the same type. Therefore, developing an AI diagnosis algorithm specific to COVID-19 was necessary, said co-author Feng Jianjiang at Tsinghua University, also an expert in fingerprint recognition and computer vision.

Though doubts about using CT scans to detect COVID-19 remain, Feng said CT scans play a vital role in severity assessments and patient management. Related applications of the AI system have been used by doctors in Wuhan hospitals.

The AI diagnosis algorithm also has the advantages of suitability for high repeat usage and easy large-scale deployment, showing its potential to become a new tool to help control the spread of COVID-19, Feng said.