Techniques for the Automated Segmentation of Lung in Thoracic Computed Tomography Scans

Introduction

Computed Tomography (CT) of the thorax is used to diagnose and assess numerous thoracic diseases such as mesothelioma, emphysema, pulmonary embolism, and lung nodules (Armato & Sensakovic, 2004; Morgan, 1992; Mastuani, MacMahon, & Doi, 2002; Sensakovic,, et al., 2011; Zhou, et al., 2005). The availability of multidetector-row, fan-beam, and cone-beam CT scanners has improved image resolution and decreased scanning time, thus expanding the role of CT in the diagnostic assessment of patients. The improved resolution has resulted in substantially larger amounts of image data that must be analyzed by the radiologist or clinician. It is the tedious and time-consuming nature of this data analysis that has motivated the development of Computer-Aided Diagnostic (CAD) methods to assist radiologists in the evaluation of CT scans.

CAD methods that are applied to CT scans of the thorax usually require identification of scan voxels that comprise the lung (i.e., lung segmentation). The purpose of lung segmentation is threefold. First, many thoracic diseases such as nodules and emphysema are located in the lung. Lung segmentation allows a CAD method or image-guided therapy system to focus on analyzing only those voxels that comprise the lung, thus reducing the potential for errors (e.g., misdiagnosis of aortic calcification as a lung nodule). Further, limiting processing to the lungs greatly reduces computation time, because the lungs occupy only a fraction of the total thoracic CT scan. Second, the lungs are prominent, high-contrast structures centrally located within the thorax and thus may be used as guides for the segmentation of other structures of interest such as the heart or pleura (Sensakovic, 2010). Finally, lung segmentation may be useful for image data visualization. The three-dimensional (3D) display of CT scans is an area of rapid development, and lung segmentation would be required in a situation where, for example, a volume-rendered version of the lung parenchyma is desired as a visual aid for surgical planning (Wei, Hu, MacGregor, & Gelfand, 2006).

Lung segmentation must be accurate because abnormalities such as lung nodules may exist at the extreme periphery of the lungs. If the entire lung is not segmented, such abnormalities will be lost to subsequent analyses (Armato & Sensakovic, 2004). Quantitative assessment of lung volume (Sensakovic, Armato, Starkey, Kindler, & Vigneswaran, 2010) is a means of assessing several diseases, and inaccurate measurements could result in incorrect assessment. Finally, segmentation of pleural disease will often use the lung segmentation to establish the boundary of the pleura (and thus the disease boundary). Inaccurate segmentation of the lung will result in an incorrect segmentation (and assessment) of pleural disease (Sensakovic, 2010).

Lung segmentation in CT scans is one of the most extensively researched areas in medical image processing (Sluimir, et al. 2006). The low density of lung parenchyma compared with surrounding soft tissue translates into high contrast on CT, which in turn facilitates the use of several image processing techniques such as histogram thresholding (Armato & Sensakovic, 2004), active contours (Cohen, 1991; Kass, Witkin, & Terzopoulos, 1988), models (Brown, 1997), and graph cuts (Massoptier, 2009). This chapter describes an intensity-based lung segmentation method. The segmentation method begins with simple thresholding, and several image processing modules are included to improve segmentation accuracy and robustness. Common segmentation difficulties are discussed and motivate the inclusion of each module in the lung segmentation method.