Global optimal hybrid geometric active contour for automated lung segmentation on CT images

Highlights

• A global optimal active contour model is proposed for automated lung segmentation.

• The combination of region and edge information leads to high segmentation accuracy.

• The global optimality improves the segmentation robustness and stability.

• The proposed model requires fewer iterations and leads to high efficiency.

Abstract

Lung segmentation on thoracic CT images plays an important role in early detection, diagnosis and 3D visualization of lung cancer. The segmentation accuracy, stability, and efficiency of serial CT scans have a significant impact on the performance of computer-aided detection. This paper proposes a global optimal hybrid geometric active contour model for automated lung segmentation on CT images. Firstly, the combination of global region and edge information leads to high segmentation accuracy in lung regions with weak boundaries or narrow bands. Secondly, due to the global optimality of energy functional, the proposed model is robust to the initial position of level set function and requires fewer iterations. Thus, the stability and efficiency of lung segmentation on serial CT slices can be greatly improved by taking advantage of the information between adjacent slices. In addition, to achieve the whole process of automated segmentation for lung cancer, two assistant algorithms based on prior shape and anatomical knowledge are proposed. The algorithms not only automatically separate the left and right lungs, but also include juxta-pleural tumors into the segmentation result. The proposed method was quantitatively validated on subjects from the publicly available LIDC-IDRI and our own data sets. Exhaustive experimental results demonstrate the superiority and competency of our method, especially compared with the typical edge-based geometric active contour model.

Introduction

Lung cancer is one of the most lethal diseases in clinical medicine. Early diagnosis of lung cancer by computer-aided diagnosis (CAD) systems is of major importance. It is greatly helpful to improve the effectiveness of treatment and increase the patient's survival rate [1]. In most cancerous lungs, pulmonary nodules and tumors are the main lesions. In order to efficiently reduce the search space of such pathological changes, most CAD systems automatically segment the lung fields on thoracic CT images (usually used as a sensitive tool for non-invasive detection) as the first step [2]. However, designing effective lung segmentation method on chest CT scans is challenging, especially for high-throughput applications with a numerous number of data sets to be processed [3], such as the screening of early stage lung cancer.

Inspired by numerous medical image analysis techniques, such as neural networks [4], [5], [6], gradient projection algorithms [7], active contour models [8], [9], [10], and graph cuts [11], [12], a wealth of lung segmentation methods used for chest CT images have been proposed in recent years. According to the different application scenarios, most existing techniques can be classified into two categories. (1) Methods for cancerous lungs with nodules and/or tumors [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28]. (2) Methods for pathological lungs suffered from chronic obstructive pulmonary disease, interstitial lung disease, etc [29], [30], [31], [32]. This paper mainly focuses on the former segmentation methods for lung cancer.

Most existing segmentation techniques for lung cancer on chest CT scans can be classified into four categories: methods based on signal thresholding [13], [14], [15], [16], active contour models (ACMs) [17], [18], [19], Markov-Gibbs random field (MGRF) models [20], [21], [22] and shape models [23], [24], [25].

Thresholding methods segment lung fields from the other tissues based on the fact that lung regions have lower densities compared with the other structures of the chest. Hu et al. [13] computed an iterative threshold to get an initial lung region, followed by a refinement process using opening and closing morphological operations. Korfiatis et al. [14] combined 2D wavelet with 3D optimal threshold to segment lung volumes, further refined by 3D morphological processing. To improve the segmentation accuracy for lung cancer, a robust method based on the curvature of ribs was presented in Ref. [15]. To refine the segmentation and include juxta-pleural nodules, Pu et al. [16] proposed an adaptive border marching algorithm to march along the lung borders after initially segmenting the lung regions by thresholding. The main problem of the threshold-based method is that its accuracy is affected by many factors, including image acquisition protocol, scanner type and inhomogeneity of densities in the lung region. Therefore, some intensive postprocessing steps should be applied to further refine the initial segmentation result [1].

The ACMs are utilized to segment lung tissues based on the distinction of densities between lung fields and the surrounding structures. To some extent, the ACMs can overcome the inhomogeneities in the lung region compared with the threshold-based techniques. Itai et al. [17] proposed a method for automated segmentation of lung areas by employing the parametric snakes model presented in Ref. [33]. Silveira et al. [18] used a geometric active contour initialized at the boundary of the chest region, which was then automatically split into two regions representing the left and right lungs. The segmentation techniques above are both 2D methods, which limit the utility of prior information consisting in adjacent CT slices. Thus, Li et al. [19] proposed a semi-3D segmentation method using a geometric active contour that was firstly presented in Ref. [34]. Nevertheless, this technique suffers from the drawback that the accuracy is sensitive to the initial position of zero level contour. As a result, it needs manual intervention to re-initialize the level set function (LSF) at every certain number of slices, which is excessively time consuming.

MGRF-based methods can be used to segment lung tissues due to the advantage that they incorporate 3D information existing in spatially adjacent images. El-Baz et al. [20] proposed an unsupervised method for multimodal medical image segmentation. In this method, expectation-maximization algorithm and linear combination of Gaussians is combined with joint MGRF. The application to precisely segment lung regions using this method is detailed in Ref. [21]. Furthermore, they extended their work by applying iterative MGRF framework on different scale spaces, followed by a Bayesian approach to fuse the segmentation results [22]. However, these MGRF models suffer from an intrinsic defect that the prior model has a significant effect on segmentation accuracy.

Shape-based techniques, similar to MGRF models, add the prior information of lung shape to image signals. Pu et al. [23] described a shape break-and-repair strategy which was utilized to segment lungs with juxta-pleural nodules. Sun et al. [24] segmented the lungs through two main processing steps. First, a 3D active shape model (ASM) matching method is used to roughly define the initial lung borders. Second, the segmented lungs are refined using a global surface optimization method developed by Li et al. [35]. To further extend their work in Ref. [24], Gill et al. [25] separated left and right lungs by means of a classification approach based on 3D sheetness filter. The main limitation of the shape-based techniques is that the accuracy depends strongly on how accurately the prior shape model is registered with respect to the CT image [1].

In addition to the above four main types of methods, some other algorithms have been applied to segment lung regions on chest CT images. For instance, Shen et al. [26] adopted a bidirectional chain coding method incorporated with a support vector machine to smooth the lung border of adjacent regions. Dai et al. [27] used a graph cuts-based method to segment the lung fields. In recent years, nonnegative matrix factorization has been applied to 3D lung segmentation [28].

Among the existing techniques above, the ACMs are superior in lung segmentation because of the following reasons. Firstly, they have obvious geometrical significance, which matches the geometric features of lung regions in CT scans. Secondly, they can be easily remedied by additional regularization terms, which can adapt to different requirements. Thirdly, the parameters adjustment is flexible. Thus, we focus on the lung segmentation method based on ACMs in this paper.

In the existing ACMs, the level set method is widely used due to its advantages in handling numerical computation and topological changes. The level set based ACMs are usually called geometric active contour (GAC) models, which can be mainly classified into two typical modalities: edge-based GAC (EGAC) and region-based GAC (RGAC). The EGAC is suitable for images in which the object and background have a significant difference in gray levels, in other words, with obvious object boundaries. In contrast, the RGAC is suitable for images in which the gray levels of the object and background are more uniform. Each modality has its own strength and weakness in different situations. (1) The EGAC is more efficient in computation speed than RGAC. (2) The RGAC is more applicable to objects with weak boundaries or narrow bands than EGAC. (3) The RGAC is more robust to the initial position of zero level contour than EGAC.

The densities of pulmonary parenchyma and its surrounding tissues usually have a significant difference on chest CT images. Furthermore, the gray levels in lung regions are not uniform because of the vessels and lesions with relatively high densities. Simultaneously considering the computational efficiency, the EGAC model can be a good choice to tackle this task. However, the EGAC only uses the information of edge gradient, leading to low accuracy on the object with weak boundaries or narrow bands. Moreover, the conventional EGAC is sensitive to the initial position of LSF due to the local optimality of energy functional, resulting in cumulative transmission of segmentation error and low robustness. In addition, it is difficult for GAC models to take advantage of the prior spatial knowledge between adjacent CT slices as MGRF and shape models. As a result, the accurate, stable, and efficient lung segmentation on serial CT images using GAC is a nontrivial problem.

To overcome the above difficulties, a global optimal hybrid GAC model based on global region and edge information is proposed, named region- and edge-based geometric active contour (REGAC). Firstly, the combination of both region and edge information can improve the segmentation accuracy in lung regions with weak boundaries or narrow bands. Furthermore, due to the global optimality of REGAC, it is more robust to the initial position of LSF and requires fewer iterations. Thus, the stability and efficiency of lung segmentation on serial CT slices can be greatly promoted by taking advantage of the prior spatial information between adjacent images.

In addition, to achieve the whole process of automated segmentation for lung cancer, two assistant algorithms using prior shape and anatomical knowledge are proposed. (1) The strategy for left and right lungs separation, which can avoid the connection between both lungs near the mediastinum during the iteration of REGAC. (2) The segmentation of juxta-pleural tumors, which is of great importance for the analysis of pathological changes in lung cancer. With the methods above, the effectiveness of automated lung segmentation on serial chest CT images can be greatly improved. The framework proposed in this paper can provide a valuable basis for early detection, diagnosis and 3D visualization of lung cancer.

The remainder of this paper is organized as follows. Section 2 presents the REGAC model, analyzes its global optimality, discusses the parameters setting, and compares it with the conventional EGAC model. Section 3 describes the whole process of automated lung segmentation method using REGAC on serial CT images. The assistant algorithms of left and right lungs separation and juxta-pleural tumors segmentation are also presented in this section. Section 4 compares the performance of REGAC with the typical EGAC in accuracy, robustness, and efficiency, respectively, followed by presenting the lung segmentation results obtained from the publicly available LIDC-IDRI (Lung Image Database Consortium and Image Database Resource Initiative) and our own datasets. Finally, Section 5 concludes this paper.