3D lung segmentation on CT images using region-based method

Abstract

Computed Tomography (CT) images are becoming a valuable aid in advanced lung cancer investigation. Automated segmentation of lung structures from CT images is the elemental step in many lung cancer diagnosis systems, where the malignant nodules are identified at its initial stage. Lung  segmentation is a challenging task because of the variant density distribution amongst the lung region, similar densities in the pulmonary structures, and scanning protocols. This work intended to recognize the efficient region-based segmentation framework to segment lung volumes. The standard region-based approaches, such as watershed and region growing, are adapted to segment the lung region. The specified segmentation methods segment lung in the axial space and achieve promising segmentation results. The performance of the system is assessed by computing the DSC, Hausdorff distance and AVD between the automatically segmented lung volumes and the volumes manually outlined by radiological experts. The marker-controlled watershed method followed by a sequence of morphological operations segments the lung parenchyma accurately with a dice score of 92.5 and the region growing method with a dice score of 93.8 concerning the expert-traced contours. The obtained results confirm the effectiveness of the system. It is suggested that the performance of the described framework could be further enhanced by fusing shape-based features.