Tamil Nadu university develops AI-based software for Covid-19 preliminary screening

CHENNAI: Running a chest x-ray through artificial intelligence-based software can help identify the presence of novel coronavirus before a confirmatory swab test is done. The software, developed by Thanjavur-based Sastra University to look for specific features in a chest x-ray of a Covid patient, could be used on a wider population to help reduce demand for RT-PCR tests.

It was tested and validated with samples from three TN district government hospitals. The tool, using machine learning algorithms to identify and differentiate among the x-ray of a healthy patient, one with coronavirus or one with pneumonia in less than a second, is trained on several million images.


"This tool can be a preliminary screening filter based on x-ray images, so we can narrow down the test scope of targeted suspects. This shall reduce the demand for RT-PCR tests under the revised ICMR testing protocols," said S Vaidhyasubramaniam, vice-chancellor, Sastra Deemed University. "Validation was done with x-ray images from government hospitals in Theni, Tirunelveli and Thanjavur and the preliminary results were satisfactory. Thanjavur government hospital is scaling up testing with more images."

Dr Narayana Babu, DME and incharge of Omandurar Government Multi Superspecialty Hospital, said the software would be tested, while the Tirunelveli Medical College and Hospital has requested the varsity to share the software.

Professor R Elakkiya, who developed the tool, said it was modelled on 14 million x-ray images from 1,000 classes and was later trained on 6,000 x-ray images. In about 0.23 seconds, it could scan a chest x-ray and tell if it belonged to a Covid positive patient, a healthy one or one with pneumonia. The deep learning algorithm in the tool uses multiple layers to extract features from x-ray images and create a database. When a new image is fed, the tool compares the new features with those in its database and generates a result.

"We developed VGG16, a network trained on 14 million images. I made a transfer learning from that network to my CNN baseline network and I augmented it with 12,000 images," she said. "We validated our tool with 150 images from the three government hospitals and found it 100% accurate."