Experts at St. Petersburg Polytechnic University have developed a highly accurate neural network capable of detecting COVID-19 from lung X-ray and CT images. The reported accuracy reaches 99.23 percent, a finding that emerged from studies conducted at St. Petersburg Polytechnic University and shared with media outlets including socialbites.ca.
Today, polymerase chain reaction testing stands as the standard method for diagnosing the disease, yet the process presents several drawbacks. Results typically require 4 to 6 hours, and early-stage infections can yield false negatives. This reality motivates ongoing efforts to improve early diagnosis and reduce transmission risk through faster, more reliable methods.
Imaging of the lungs with X-ray or CT scans serves as a valuable complement to PCR testing. These imaging techniques help clinicians distinguish COVID-19 related pneumonia from other respiratory conditions. Nevertheless, interpreting X-ray and CT findings demands the expertise of highly qualified radiologists and clinicians who can integrate imaging results with clinical context.
To streamline image interpretation, researchers at St. Petersburg Polytechnic University have developed a neural network that identifies signs of COVID-19 related pneumonia on computed tomography images. The system is designed to flag specific patterns associated with COVID-19 and related respiratory illnesses, aiding physicians in making quicker and more accurate assessments that can improve patient triage and management.
One developer from the project, an engineer with the world wide Scientific Center Advanced Digital Technologies and the Mathematical Modeling and Intelligent Control Systems program at SPbPU, explained that the model enables rapid and precise detection of COVID-19 or other pneumonias on lung CT scans. This capability can alleviate workload for medical staff by providing dependable preliminary interpretations and enhancing diagnostic consistency across cases, according to the researchers.
For training the neural network, the team assembled a diverse dataset that draws from CT slices produced in multiple countries. The collection includes more than 7,500 images of lungs affected by coronavirus related pneumonia, more than 2,500 images of lungs with community acquired pneumonia, and roughly 7,000 images depicting healthy lungs. This broad spectrum helps the model learn to discriminate among different pulmonary conditions and normal anatomy, which is essential for reducing misinterpretation in clinical practice.
In validation studies conducted through fourfold cross-validation, the neural network demonstrated robust performance, reinforcing confidence in its potential to accurately and swiftly diagnose COVID-19 from CT images. The researchers emphasize the technology’s promise for speeding up diagnostics while maintaining reliability, a balance that can be crucial in busy clinical environments where timely decisions matter most.
Earlier work by the same research teams explored neural networks for other medical image tasks, including the detection of deepfakes in photographs and video. That prior experience with image analysis informs the current project, providing a foundation in pattern recognition and reliability that underpins the ongoing development of AI-assisted radiology tools. The collaboration highlights how advances in artificial intelligence are increasingly translating into practical tools that support frontline clinicians and reduce diagnostic uncertainty.