With the emergence of several complex Covid-19 variants, diagnostic applications are urgently needed to control the epidemic and DTU researchers want to contribute with their “X-ray and CT Application for Covid-19 Detection”.
Dr. Duong Tuan Linh (top) with his research team at the DTU CSE
Seeing how people’s health was so severely impacted in the initial days of virus infestation in Vietnam, Dr. Duong Tuan Linh, a DTU researcher, proposed an application built on artificial intelligence to diagnose Covid-19. With previous experience in the EfficientNet convolutional deep neural network project, he collected X-ray and CT images and developed a work plan.
“I discovered that, if I only used the original EfficientNet network, the model was only efficient enough to distinguish between two groups in the X-ray and CT images: well or unwell,” explained Dr. Duong Tuan Linh. “That was a critical issue. So I looked for ways of replacing the final ‘fully connected layer’, which is more detailed and must also be taken into consideration when replacing, adding, or tuning parameters. Ultimately, I found a way to introduce new layers of neurons into the backbone network to increase the efficiency of distinguishing the different agents of pneumonia.
As a result, our AI-based application can compare X-ray or CT imagery exhibiting signs of Covid-19, pulmonary tuberculosis or pneumonia illnesses to images of healthy lungs and the cause, a bacterial infection, a virus, COPD, SARS-CoV-1, or MERS-CoV. The latter two may be caused by SARS-CoV-2, which is generally called Covid-19)”
Images show whether a patient is infected or not and which agent is involved
Results:
- With X-ray images, an accuracy of 97% was achieved
- With CT chest scans, an accuracy of 99% was achieved
The research group has published several papers related to the use of artificial intelligence and the development of this application in international ISI/Scopus-indexed journals.
Dr. Duong Tuan Linh benefits from the support of DTU Provost Dr. Le Nguyen Bao, who assembled a team to design the application. After surveying local doctors and Dr. Hoang Ha at DTU on the feasibility and requirement, the research group created machine learning models for use on the web and through a mobile app.
- For the web platform, the research group used the Java programming language to create the processing core, and HTML, CSS and JavaScript to create the user interface
- For the mobile platform, the research group used Java for Android and Swift for iOS
(Media Center)