Chinese Researchers Develop & Share AI Model that Can Diagnose COVID-19 over Influenza & Pneumonia in 3 Seconds

Chinese Researchers Develop & Share AI Model that Can Diagnose COVID-19 over Influenza & Pneumonia in 3 Seconds

Researchers from Tsinghua University and Wuhan-based Union Hospital (associated with Huazhong University of Science and Technology) have developed an artificial intelligence (AI)-based system that augments physician detection of COVID-19 from other viruses such as the flu or other respiratory diseases in as little as three seconds. Published recently in the journal Nature Communications, this investigational tool exhibits great potential to help providers differentiate SARS-CoV-2, the virus behind COVID-19, and influenzas as the influenza season approaches and the next wave of COVID-19 sweeps through large parts of the world. This system can help medical providers almost immediately differentiate COVID-19 from other illnesses with great accuracy by picking out the deltas in patients’ CT scans. In the spirit of world collaboration targeting COVID-19, the researchers shared the data behind the research as well as the code for the AI system in GitHub.

The AI System

The Chinese research team utilized a large dataset including over 11,000 CT volumes from COVID-19, influenza, and non-viral community-acquired pneumonia and non-pneumonia reported CGTN News. This data was collected from February to March and involved three hospitals in Wuhan, the original epicenter of the COVID-19 pandemic.

Based on a “deep convolutional neural network-based system,” the AI model transforms experts’ detection experiences into actual algorithms powering the system. After rigorous testing, the study authors revealed that this AI model can differentiate four (4) respiratory diseases (including COVID-19, influenzas, non-pneumonia) to an accuracy of 97.8 percent.

Moreover, the AI model beat out five experienced radiologists in performance. Hence this tool, if validated and approved by regulatory authorities, could greatly aid providers’ workload. The study revealed that for a radiologist, the average reading time of 6.5 minutes was beaten by the AI model, coming in at 2.73 seconds!

Humans still win when it comes to distinguishing pneumonia from non-pneumonia in CT scans.

The AI Code Available on GitHub

In a bid to help combat COVID-19, the researchers have made this code available to other developers. It can be accessed via GitHub.

Lead Research/Investigator

Jianjiang Feng, Department of Automation, Beijing National Research Center for Information Science and Technology, Tsinghua University, Beijing, China

Heshui Shi, Department of Radiology, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China