Today, computer tomography is usually used to diagnose viral pneumonia. The method has some drawbacks. The key drawback is considerable radiation exposure. Among others are the probability of a mistake by an X-ray technician or the issue of excessive weight, which prevents some patients from being examined. However, there is a way out: Roman Kostuchenko, a bachelor’s student from the Faculty of Applied Mathematics, Informatics, and Mechanics of Voronezh State University, created a program for the automatic identification of the type of pneumonia by X-ray pictures. This development is very timely during the coronavirus pandemic.
“The topic of my bachelor’s graduation paper is “Developing an X-ray pneumonia identification system”. I chose it with the help of my scientific supervisor. From the very beginning, I wanted a medical topic. Irina Kashirina suggested a number of ideas and I had some of my own. Finally, we decided to work with pictures. I had to develop a system which would use X-ray pictures of lungs to diagnose if there is pneumonia, and in case of a positive result, to identify its type: bacterial or viral. When developing the system, I had to study general problems of medical diagnostics and the principle of neural networks’ operation. I also studied the experience of using neural networks in other medical projects. As a result, I managed to create a system based on convolutional neural networks, which uses X-ray pictures to identify pneumonia and its type. It is too early to talk about introducing it into practical use as it takes time to and is a huge responsibility to perfect such support systems for medical decisions. I would like to find new solutions and techniques to improve the forecast precision in my master’s thesis,” said Roman.
Irina Kashirina, professor at the Department of Mathematical Methods of Operations Research and the supervisor of the project, emphasised that the research needs to be continued and the measurements precision needs to be increased:
“Surely, we are not the first to work on this problem. However, it hasn’t been solved yet, and we have a chance to increase the precision of our development and thus make it more efficient. This is just a beginning: the algorithm precision is over 80% now. In the nearest future, we plan to increase it to 95%”.