COVID-19 lung infection segmentation from CT imaging using statistics and edge-region-based active contour

As of October 2024, the number of global confirmed cases of COVID-19 goes beyond 776 million, with over 7 million deaths, according to World Health Organization (WHO) website. This scarring figure has led to an impressive effort from the medical community, in the attempt to early detect the signs of the infection. Whereas the Reverse Transcription Polymerase Chain Reaction (RT-PCR) testing protocol is being used to detect the infection, medical imaging plays an important role to evaluate the level of lung’s damage caused by the presence of the virus. Both computed tomography (CT) and chest radiographs (CXR) have been utilized for laboratory testing by radiologist to identify and measure the affected lung area by isolating the region of interest (ROI). Manual segmentation of ROI is a complex process requiring extensive time and experienced medical staff. Therefore, there is an urgent need of automated assisted medical tools that accurately measure the infected areas and reduce the manual annotation time. An impressive amount of approaches have been proposed to detect the infection or to segment the infected areas, where most of the proposed techniques rely on deep learning (DL). In this work, an alternative to DL is proposed, that is based on several steps, including statistical measures. More precisely, in the first step, the image is coarsely segmented by using an electromagnetism optimization based multilevel thresholding. The multilevels are a priori estimated with the help of Gaussian mixture models (GMM). Next, a morphological skeleton is constructed for the basis of a localized edge-region-based active contour model considering multi-class segmentation. The segmented class is reevaluated and correction step is performed if necessary, i.e. if the number of components is wrongly estimated. The experiments indicate very promising results, the approach performing similar to recent state-of-the-art methods.