Lung’s Segmentation Using Context-Aware Regressive Conditional GAN

Applied Sciences

Zakir Khan¹  Arif Iqbal Umar¹  Syed Hamad Shirazi¹  Assad Rasheed¹  Waqas Yousaf¹  Muhammad Assam²  Izaz Hassan²  Abdullah Mohamed³ 

[1] Department of CS & IT, Hazara University Mansehra, Mansehra 21120, Pakistan;Department of Software Engineering, University of Science and Technology Bannu, Bannu 28100, Pakistan;Research Centre, Future University in Egypt, New Cairo 11835, Egypt;

关键词: generative adversarial network (GAN);    COVID-19;    lung segmentation;    deep learning;   

DOI  :  10.3390/app12125768

来源: DOAJ

【 摘 要 】

After declaring COVID-19 pneumonia as a pandemic, researchers promptly advanced to seek solutions for patients fighting this fatal disease. Computed tomography (CT) scans offer valuable insight into how COVID-19 infection affects the lungs. Analysis of CT scans is very significant, especially when physicians are striving for quick solutions. This study successfully segmented lung infection due to COVID-19 and provided a physician with a quantitative analysis of the condition. COVID-19 lesions often occur near and over parenchyma walls, which are denser and exhibit lower contrast than the tissues outside the parenchyma. We applied Adoptive Wallis and Gaussian filter alternatively to regulate the outlining of the lungs and lesions near the parenchyma. We proposed a context-aware conditional generative adversarial network (CGAN) with gradient penalty and spectral normalization for automatic segmentation of lungs and lesion segmentation. The proposed CGAN implements higher-order statistics when compared to traditional deep-learning models. The proposed CGAN produced promising results for lung segmentation. Similarly, CGAN has shown outstanding results for COVID-19 lesions segmentation with an accuracy of 99.91%, DSC of 92.91%, and AJC of 92.91%. Moreover, we achieved an accuracy of 99.87%, DSC of 96.77%, and AJC of 95.59% for lung segmentation. Additionally, the suggested network attained a sensitivity of 100%, 81.02%, 76.45%, and 99.01%, respectively, for critical, severe, moderate, and mild infection severity levels. The proposed model outperformed state-of-the-art techniques for the COVID-19 segmentation and detection cases.

【 授权许可】

Unknown