Health IT, Artificial Intelligence

AI system detects Covid-19 in lungs faster than radiologists, study finds  

Northwestern University researchers developed an AI system that analyzes patients' chest X-rays to identify Covid-19. A study shows it can classify the images faster and with slightly higher accuracy than radiologists.

A new artificial intelligence system, developed by researchers at Northwestern University, was able to detect Covid-19 in the lungs by analyzing X-ray images faster and slightly more accurately than thoracic radiologists, according to a new study.

Published in the journal Radiology, the study details the use of the system, including how it was developed and tested. The system, DeepCOVID-XR, utilizes a machine-learning algorithm to detect Covid-19 on chest X-rays.

“The system learned the characteristic features of Covid-19 on its own, without explicit instructions or programming,” said Dr. Aggelos Katsaggelos, an AI expert at Northwestern University and senior author of the study, in an email. “DeepCOVID-XR consists of the ensemble of six specific types of deep neural networks, referred to as convolutional neural networks. CNNs have proven to be very powerful in learning patterns in images.”

The algorithm was trained and validated using 14,788 chest X-ray images, of which 4,253 were from patients with Covid-19. The images were gathered from patients seen at sites across Chicago-based Northwestern Medicine from February to April. The algorithm was then tested using 2,214 chest X-ray images.

Researchers found that the system was able to classify the 2,214 test images, of which 1,194 were Covid-19-positive, with an accuracy of 83%.

“A pleasant surprise was that the errors made by the algorithm were explainable,” Katsaggelos said. Further, the researchers found that DeepCOVID-XR was able to hone in on the classic features of COVID-19-associated pneumonia seen on chest X-rays.

Researchers also compared the system’s performance to that of five experienced thoracic radiologists. The radiologists were asked to examine 300 chest X-ray images and classify them, and then the system was used to do the same. They found each radiologist took approximately two-and-a-half to three-and-a-half hours to examine the set of images, but the AI system took around 18 minutes.

In addition, DeepCOVID-XR outperformed the radiologists slightly with respect to accuracy, with the system displaying 82% accuracy versus the radiologists’ accuracy, which ranged from 76% to 81%.

“We feel that this system has the potential to provide significant benefit to overburdened healthcare systems in mitigating unnecessary exposure to the virus by serving as an automated tool to rapidly flag patients with suspicious chest imaging for isolation and further testing,” said Katsaggelos.

The Covid-19 pandemic is surging in the United States, with more than 13 million Americans having contracted the disease and at least 268,600 having died from it as of Dec. 1, according to a New York Times database.

The AI system is currently in a clinical research phase, Katsaggelos said. AI systems need to be validated repeatedly using diverse datasets before they can be moved out of the research setting and into clinical use.

The Northwestern researchers have made the model and source code used to develop the system freely available for other researchers worldwide to further test and fine-tune the algorithm, he added.

Photo credit: Egor Kulinich, Getty Images

 

 

Shares0
Shares0