Two-Stage COVID-19 Lung Segmentation from CT Images by Integrating Rib Outlining and Contour Refinement

Abstract

In the early 2020, the corona virus disease (COVID-19) has become a global epidemic, resulting in a profound impact on the lives of billions of people, from both of safety and economic perspective. Rapid screening is of great significance to control the spread of disease. In the clinic practice, computed tomography (CT) is widely utilized by the doctors to detect and evaluate novel coronavirus pneumonia and lung segmentation is a necessary initial step for lung image analysis. However, the segmentation task is still confronted with several challenges, such as high variations in intensity, indistinct edges and noise due to data acquisition process. To address aforementioned challenges, in this paper, we present a novel two-stage-based COVID-19 lung segmentation method. First, we design a coarse segmentation method combining threshold and rib outline, which can remove most background while retaining complete lung shapes. Then, a contour refinement algorithm was introduced to refine the coarse segmentation based on local information including intensity, shape and gradient. The proposed method was tested on 20 sets of COVID-19 CT cases. Quantitative evaluations are conducted with different methods (including the deep learning-based approach), and the results demonstrate that our method can provide superior performance with few-shot samples. Our method achieves an average symmetric surface distance (ASD) of $0.0101 \pm 0.0147$mm and dice similarity coefficient (DSC) of $99.22 \pm 0.99\%$ on lung CT image segmentation compared with ground truths. To better promote the research in this field, we have released our code to facilitate the research community(https://github.com/qianjingw/Two-Stage-COVID-19-Lung-Segmentation).

Qianjing Wang is a graduate student.