Detecting Pneumonia for COVID 19 Patients using Multi-Image Fusion for CT Images

Abstract

Early in 2020, the global spread of Coronavirus Disease 2019 (COVID-19) triggered an existential health crisis. Automated lung infection diagnosis using Computed Tomography (CT) images has the potential to significantly improve the current healthcare approach to combat COVID-19. But segmenting infected regions from CT slices is difficult due to the wide variety in infection traits and the weak contrast between infected and healthy tissues. Additionally, gathering a lot of data quickly is impractical, which hinders the training of a deep model. This study proposes COVID-SegNet, a convolutional-based deep learning technique for automatically segmenting COVID-19 infection areas and the whole lungs from chest CT images. The suggested deep CNN includes a feature variation (FV) block that adaptively modifies the global properties of the features for segmenting COVID-19 infection. This can improve its capacity to express features in various situations efficiently and adaptively. To deal with the complex shape variations of COVID-19 infection zones, additionally recommend the use of PASPP, a progressive atrous spatial pyramid pooling. After a simple convolution module, PASPP generates the final features using multistage parallel fusion branches. In order to cover a variety of receptive fields, PASPP uses atrous filters with an acceptable dilation rate in each atrous convolutional layer. For the segmentation of COVID-19 and the lungs, the dice similarity coefficients are 0.987 as well as 0.726, respectively. Experiments carried out on data gathered in the scan centre demonstrate that effectively produce good performance.