LETI Researchers Trained a Neural Network to Detect Lung Abnormalities from COVID-19
The system will allow doctors to diagnose patients with suspected COVID-19 more quickly and accurately in the highly congested medical facilities during the pandemic.
Coronavirus, which causes the dangerous COVID-19, has been spreading at a high rate since 2020. Today, over 500 million cases have been reported worldwide. A high number of daily cases results in a colossal burden on the entire healthcare system.
Coronavirus detection requires different technologies because external symptoms (cough, fever, weakness, etc.) are also seen with other diseases and do not allow a definite diagnosis of COVID-19. Because the virus spreads in the lungs, CT scans and chest X-rays are widely used to detect and determine the severity of the disease.
"Today, medical systems are under tremendous pressure to detect coronavirus in patients' lungs. Doctors have to look at a significant number of scans every day. This can lead to decreased concentration and attention, which means an increased likelihood of error. We have developed a special program that will help a physician to focus on the images with pathology in the first place and thus simplify the task of making a diagnosis."
The researchers collaborated with specialists from St. Petersburg City Mariinsky Hospital. The medics provided the LETI researchers with 1,600 X-rays from patients who were diagnosed with coronavirus. To train the neural network, the researchers classified the collected data into three different categories of lung lesions: atelectasis, soft tissue emphysema, and pneumonia, which manifests with COVID-19.
"The neural network instantly recognizes what is on the image: pathology or normal condition. According to tests, its accuracy is about 98%. We hope that in the future, our program can be implemented in hospitals for fast and accurate COVID-19 diagnosis," adds Nikolay Staroverov.
In the future, scientists plan to increase the set of collected data to improve the accuracy of the neural network.