A new automated system that involves deep learning technology enables the detection of COVID-19 lesion via the analysis of a computed tomography (CT) scan.

The functioning of the system developed by researchers at Universitat de Barcelona (UB; Barcelona, Spain) consists of “a first phase of lung segmentation with the CT scan to reduce the searching area,” said Giuseppe Pezzano, researcher at the UB and the principal researcher of the study. “Then, an algorithm is used to analyze the lung area and detect the presence of COVID-19. If it tests positive, the image is processed to identify the areas that are affected by the disease.” The study “has enabled us to verify the efficiency of the system as a support tool for decision-making for health professionals in their COVID-19 detection task, and for measuring the gravity, the extension and the evolution of the pneumonia caused by SARS-CoV-2, in the mid and long term,” noted Pezzano.

The algorithm has been tested in 79 volumes and 110 sections of CTs which had detected COVID-19 infection, obtained in three open-access image repositories. The researchers achieved an average accuracy for the segmentation of lesions caused by the virus of about 99%, without false positives being observed during the identification. The method uses an innovative way to calculate the mask of segmentation of medical images, which provided good results in the segmentation of nodules in the tomography images.

Some recently published studies “show that deep learning and computing vision algorithms have achieved a better precision than the experts’ cancer detection in mammograms, prediction of strokes and heart attacks,” said Petia Radeva, professor at the Department of Mathematics and Computer Science of the UB. “We could not be left behind and therefore we have worked on this technology to help doctors fight COVID-19 by offering them high-precision data for the analysis of medical images in an objective, transparent and robust way.”

“This type of automated system represents an important tool for health professionals in order to make more robust and accurate diagnoses, since it can provide information a human being could not measure,” added Oliver Díaz, lecturer at the Department of Mathematics and Computer Science of the UB. 

http://www.gzjiayumed.com/en/index.asp