Development of a Module for COVID-19 Diagnostics Based on Computed Tomography Images of the Chest Based on Computer Vision Methods

Abstract

The implementation of a module of a medical decision support system for diagnosing COVID-19 using chest CT images is considered. The U-Net architecture is used for segmentation of lung parenchyma and pathological areas in chest CT images, the DSC and IoU values for parenchyma are 0.951 and 0.933, for pathological areas — 0.97 and 0.959, respectively. A method for image pre-processing based on adaptive histogram equalization is described. Methods for segmentation masks post­processing are also proposed. The first of them is necessary to separate masks into masks of the left and right lungs; it is based on the analysis of areas and mutual positions of contours. The second one is needed to eliminate artifacts. In addition to image processing methods, approaches that generate the data necessary for radiologists to make a diagnosis are also implemented (the volumes of both lungs and pathological findings in them are calculated, percentages of parenchymal tissue involvement in the pathological process are determined, the severity of the disease is assessed). The algorithms for generating a processed series of images and a DICOM SR are described. The average time spent by the module on processing one CT study containing about 600 slices, with a video memory limit of 6 GB, is 68 s, and with a limit of 8 GB — 56 s. Considering that the approximate time spent by a radiologist to process a study is about 6 minutes, the developed module can be effectively used in medical practice to reduce the burden on medical personnel.