•  
  •  
 

Abstract

Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) poses a global threat, impacting millions worldwide. While automated detection of lung infections through Computed Tomography (CT) scans is a promising alternative, segmenting infected regions from CT slices remains challenging due to low-contrast infection boundaries and blurred appearances. To address this challenge, A deep-learning model called ECGANCOVID-Net is proposed for detection and identification of infected regions in chest CT images. Our model utilizes a semantic hierarchical segmenter to detect regions of lung infection caused by Coronavirus in CT medical images. The model consists of two components, namely the U-CGAN-Net models. The initial neural network, UCGAN-Net1, is designed to detect lung parenchyma. Subsequently, the second neural network, UCGAN-Net 2, operates on the segmented lungs to accurately identify the specific regions impacted by COVID-19 lesions. UCGAN-Net comprises a conditional generative adversarial network (CGAN) incorporating an adapted generator and discriminator. Furthermore, our model employs data augmentation techniques to address the issue of limited training data. Through extensive trials, it has been discovered that the suggested methodology exhibits superior performance compared to recently proposed techniques. This is particularly evident in the improved overall performance of our model when accurately determining the location of tiny lesions. The proposed ECGANCOVID net has demonstrated exceptional performance in segmenting COVID-19 lesions, achieving higher localization performance with a Dice Similarity Coefficient (DSC) of 84.5% and Intersection over Union (IOU). Additionally, the suggested model has undergone external validation using an unseen dataset, resulting in Dice Similarity Coefficient of 69.7%.

Keywords

COVID-19 disease, Computed tomography (CT) images, Conditional generative adversarial network (CGAN), Lung and lesion segmentation, Hierarchical segmentation strategy

Subject Area

Computer Science

First Page

706

Last Page

729

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Receive Date

9-9-2023

Revise Date

1-7-2024

Accept Date

1-9-2024

Download

Share

COinS