Improving Lung Region Segmentation Based on Lazy Snapping and Clustering for Aiding COVID-19 Diagnosis

Abstract

The COVID-19 global pandemic, brought on by the rapid spread of the new coronavirus (SARS-CoV-2), has developed into one of the healthcare industry's most significant challenges in recent memory. Early detection of positive patients is essential to prevent the further spread of the COVID-19 virus. Chest x-ray (CXR) images of patients reporting shortness of breath initially led clinicians to suspect the presence of this novel virus. On CXR images, among the alterations detected in the lungs are indications of cloud region, also known as Ground-Glass Opacity. Consequently, the primary objective of this study is to develop a robust segmentation and to acquire an accurate segmented lung region in a CXR image, as this is a necessary step for accurate diagnosis using computer-aided diagnostic systems (CADS). The proposed methodology employs a multi-level segmentation strategy to improve the performance of lung region segmentation, where Lazy Snapping is utilized as pre-segmentation step to automatically remove the bone of the chest area, followed by clustering to achieve the complete segmentation. Furthermore, the advantage of fast k-means (FKM) clustering has also been utilized to obtain the desired lung region. The proposed strategy using Lazy Snapping and FKM was experimented on 150 CXR images and has achieved an average accuracy, sensitivity of and specificity of 92.38%, 85.23% and 96.27%, respectively. Based on the results obtained, this approach demonstrated efficacy in lung segmentation in chest x-ray images and has a significant potential for clinical use.