In computed tomography (CT) images, pulmonary lobe segmentation is an arduous task due to its complex structures. To remedy the problem, we introduce a new framework based on lung anatomy knowledge for lung lobe segmentation. Firstly, the priori knowledge of lung anatomy is used to identify the fissure region of interest. Then, an oriented derivative of stick filter is applied to isolate plate-like structures from clutters for lobar fissure verification. Finally, a surface fitting model is employed to complete the incomplete fissure surface for lung lobe segmentation. Compared with manually segmented fissure references, the designed approach obtained a high median
In medical practice, recognition of lung lobes is useful for clinical diagnosis and lung disease assessment [
A schematic diagram is shown in Figure
Schematic diagram of the human lungs.
Pulmonary lobe segmentation algorithms can be divided into three categories. The first approaches consist of methods that search for the fissure surface based on shape features and gray-level information. The second approaches consist of methods that use this information from airways, vessels, and fissures to estimate lobar boundaries. The third approaches consist of methods that construct pulmonary atlases to guide lung lobe segmentation. Although many methods have been presented, no standard method has yet emerged [
Considering the fact that fissures are physical boundaries between different lobes, Pu et al. [
Motivated by the assumption that the information of airways and blood vessels may contribute to segment pulmonary lobes, Ukil and Reinhardt [
In addition, a number of studies constructed pulmonary atlases to guide lung lobe segmentation. Following this strategy, Zhang et al. [
The proposed scheme aims at automatically segmenting pulmonary lobes. In order to identify fissure ROI, many studies [
Fifteen CT images were used to evaluate the validation of our scheme and acquired from the LObe and Lung Analysis 2011 (LOLA11) challenge [
A manual segmentation of lung lobes in 9 coronal slices was made by two human observers. Unfortunately, the ground truth is not publicly available. To make full use of the publicly available data for experimental evaluation, we manually segmented 15 lung lobes and regarded them as the lung lobe ground truth. In addition, lobar fissure segmentation is different from lung lobe segmentation [
Our scheme (developed in MATLAB and C++) uses a new framework for lung lobe segmentation. The designed algorithm employs three stages: (1) fissure ROI identification, (2) lobar fissure verification, and (3) lung lobe segmentation. Figure
The flowchart of lung lobe segmentation.
In this paper, we present a new computerized scheme for fissure ROI identification. Firstly, an effective method based on the graphic processing unit (GPU) is used to extract the centerline of airway trees [
To accurately segment airways, a useful and valuable method [
Normalize
Here,
The purpose of the GVF model is to make the following tube detection filter (TDF) [
The average dot product is
When
The neighbor point with the highest TDF response is chosen as the next centerline point
Different subtrees in the airway tree belong to different lobes [
Subtrees in an airway tree. (a) An airway tree. (b) Graphical representation of a typical skeletonized airway tree (figure taken from [
Airways are tubular structures and less adhered with each other. Furthermore, the GVF model has the advantage that it is feature preserving. Therefore, a method based on the GVF model is used to segment the airway tree, but the approach is not applicable for pulmonary artery and vein separation. As shown in Figure
Pulmonary arteries and veins. (a) Pulmonary arteries always adhered with pulmonary veins, and pulmonary veins may cross lobar fissures. (b) The local image corresponding to (a). As shown in the region labeled with yellow arrows, pulmonary arteries and veins are joined together. Furthermore, pulmonary veins may cross pulmonary fissures.
Pulmonary artery and vein separation is a challenging task due to its complex structure. To solve the problem, the shape prior and graph cut method [
To segment the interwoven 3D tubular trees, a medialness method [
Thus, the response of the final offset medialness is
To suppress the response away from the tube center, Bauer et al. [
The multiscale medialness response is
Here,
Pulmonary arteries may adhere with pulmonary veins in Figure
Pulmonary artery and vein separation. (a) Pulmonary arteries adhered with pulmonary veins, and the adhered regions are marked with yellow arrows. (b) Pulmonary arteries and veins are separated.
Based on the reality that pulmonary arteries are close to bronchial trees [
To illustrate the validation of the proposed scheme, we integrate with bronchial trees, pulmonary arteries, and lobar fissures to describe the distribution map from different angles of view. Bronchial trees and lobar fissures are labeled with yellow color and green color in Figure
The distribution map among bronchial trees, pulmonary arteries, and lobar fissures is displayed in different angles of view, in which bronchial trees, pulmonary arteries, and lobar fissures are marked with different colors.
As a result, we consider different bronchial trees and the corresponding pulmonary arteries belonging to different pulmonary lobes as a whole in Figure
Lobe shapes. (a) Bronchial trees and the corresponding pulmonary arteries belonging to different lobes are displayed in different colors. (b) Lobe shapes corresponding to (a). (c) The relationships among lobe shapes, fissures, and bronchial trees. (d, e) The relationships between lobe shapes and fissures.
Fissure ROI identification is useful in lung lobe segmentation. As shown in Figure
Lobe shapes. (a) Right fissure ROI, in which the estimated right upper, middle, and lower lobe distribution maps are marked with
To effectively verify pulmonary fissures, an oriented derivative of stick (ODoS) filter [
To efficiently detect the weak and thin pulmonary fissures, Xiao et al. [
The DoS filtering response can be formulated as
Here,
A vector representation is
Therefore, the response of the ODoS filter can be described as follows:
To suppress noise and other structures, Peng and Xiao [
In 3D partitioned images
Plate-like structure extraction in the right lung. After using an orientation partition scheme, the enhanced images are divided into a number of planar structures.
From the above processing, some plate-like structures are easily acquired, including lobar fissures and accessory fissures. As is shown in Figure
Lobar fissure verification in the right lung, in which fissure ROI is used to limit the search area for fissure verification.
After acquiring the right fissure (RF) and the right oblique fissure (ROF), we use a simple but effective criterion to verify the right horizontal fissure (RHF):
As shown in Figure
Right oblique fissure and horizontal fissure separation. (a) Right fissure (RF). (b) Right oblique fissure (ROF). (c) Right horizontal fissure (RHF).
Lung lobe segmentation is an arduous task due to incomplete, deformed, and disrupted fissures. To remedy the problem, a surface fitting model [
Once oblique fissures and horizontal fissures are verified, we modeled a fissure surface in the form of
A matrix expression is
At every point of the surface, the surface fitting model attempts to force the partial derivatives of the surface in neighboring cells to be equal. A mathematical expression in the second set of linear equations is of the form
To construct a smooth fissure surface, a Laplacian function is treated as the regularization term. It can be described as
Therefore, the surface fitting model is employed to construct a complete fissure surface, as shown in Figure
Surface fitting model. (a) Right oblique fissure. (b) The surface fitting results corresponding to (a). (c) Right horizontal fissure. (d) The surface fitting results corresponding to (c).
To segment pulmonary lobes, we present a cutting model to subdivide human lungs into five lobes. As shown in Figure
Pulmonary lobe separation. (a) Right complete oblique fissure surface divides the right lung into two parts. (b) Right complete horizontal fissure surface divides the special part into two parts. (c) The flowchart of the right lung lobe segmentation.
The proposed scheme is tested on 15 clinical CT images. All experiments are performed in 64 bit Windows 10 operating system, running on a computer with an Intel (R) Xeon (R) E5-1607 v2 3.00GHZ CPU and 20 GB RAM. The runtime of the proposed method for a 400 × 512 × 512-size 3D image is approximately 1580 s. In addition, the 3D visualization results are rendered by MeVisLab software [
The manual reference and its adjacent 3 mm region are labeled with
As shown in Figure
Fissure segmentation. (a) The proposed method. (b) ODoS method. (c) DoS method. The weak fissures are marked with red ellipses.
The proposed method is validated on 15 CT examinations in Figure
Segmentation validation on 15 CT examinations. (a) Left lung. (b) Right lung.
The volumes of the lung lobe ground truth and the segmented lung lobes are, respectively, marked with
To provide a global impression, four representative segmentations are chosen from the LOLA11 dataset. As shown in Figure
The segmented lung lobes are marked with different colors. The upper, middle, and lower lobes are rendered in green, red, and yellow, respectively. The first, second, and third rows represent the visual lobar boundaries, the proposed method, and the constrained interpolation profile method, in which the distinct areas of pulmonary lobes are marked with red arrows.
Table
Volumetric overlapped results. In this section, LUL, LLL, RUL, RML, and RLL are used to represent the left upper lobe, left lower lobe, right upper lobe, right middle lobe, and right lower lobe, respectively. We divide all cases into two parts: normal groups and abnormal groups.
LUL | LLL | RUL | RML | RLL | Group | |
---|---|---|---|---|---|---|
Case 1 | 0.994 | 0.970 | 0.985 | 0.981 | 0.990 | Normal |
Case 2 | 0.937 | 0.999 | 0.999 | 0.998 | 0.977 | Normal |
Case 3 | 0.996 | 0.994 | 0.998 | 0.969 | 0.986 | Normal |
Case 4 | 0.992 | 0.990 | 0.981 | 0.954 | 0.960 | Normal |
Case 5 | 0.986 | 0.992 | 0.984 | 0.918 | 0.985 | Normal |
Case 6 | 0.990 | 0.972 | 0.992 | 0.928 | 0.987 | Normal |
Case 7 | 0.933 | 0.990 | 0.994 | 0.944 | 0.993 | Normal |
Case 8 | 0.999 | 0.994 | 0.984 | 0.969 | 0.991 | Normal |
Case 9 | 0.834 | 0.992 | 0.990 | 0.947 | 0.992 | Abnormal |
Case 10 | 0.998 | 0.986 | 0.982 | 0.864 | 0.979 | Abnormal |
Case 11 | 0.989 | 0.995 | 0.891 | 0.956 | 0.989 | Abnormal |
Case 12 | 0.998 | 0.992 | 0.948 | 0.915 | 0.990 | Abnormal |
Case 13 | 0.988 | 0.999 | 0.999 | 0.845 | 0.980 | Abnormal |
Case 14 | 0.999 | 0.989 | 0.883 | 0.942 | 0.995 | Abnormal |
Case 15 | 0.957 | 0.980 | 0.985 | 0.663 | 0.985 | Abnormal |
Average | 0.960 | 0.989 | 0.973 | 0.920 | 0.985 | |
Maximum | 0.999 | 0.999 | 0.999 | 0.998 | 0.995 | |
Minimum | 0.834 | 0.970 | 0.883 | 0.663 | 0.960 | |
Median | 0.990 | 0.992 | 0.985 | 0.944 | 0.987 | |
Standard deviation | 0.058 | 0.009 | 0.037 | 0.082 | 0.009 |
The limitations of our scheme. (a) Airways. (b) Fissures, in which the disrupted bronchus is marked with red arrows, and the incomplete fissure is marked with yellow arrows.
In this paper, a unique scheme based on lung anatomy knowledge is presented for lung lobe segmentation. The described scheme has many benefits. First, the priori knowledge of airways and pulmonary arteries is used to identify fissure ROI. Many methods [
Compared with manually segmented fissure references, our scheme achieved a median
Our scheme is validated on 15 CT examinations. The average percentages of the segmented lung lobes in the lung lobe ground truth are 0.960, 0.989, 0.973, 0.920, and 0.985 for the left upper, left lower, right upper, right middle, and right lower lobes, respectively. The computerized scheme has a good performance in lung lobe segmentation. This is attributed to an ingeniously designed fusion of the surface fitting model, identified fissure ROI, and the oblique and horizontal fissure verification.
In spite of the aforementioned benefits, the described scheme has many limitations. First, our scheme is heavily depending on airway segmentation, which is an arduous task and sensitive to the image quality. Second, the lung anatomy knowledge is employed to segment pulmonary fissures, which is time consuming. Third, our scheme is derived from the ODoS filter and inherits some drawbacks. For example, some deformed fissures cannot be detected by the computerized scheme due to their orientation distribution. Fourth, poor segmented airways and pulmonary arteries may cause parts of fissures to be undetected in some cases. Finally, the fitting accuracy of the estimated lobe boundaries is largely determined by verified fissures; a better segmented method of fissure detection and lung lobe segmentation is needed to pursue.
In this paper, we present a new lung lobe segmentation scheme. Motivated by the fact that pulmonary veins may cross pulmonary fissures, the priori knowledge of airways and pulmonary arteries is used to identify pulmonary fissure ROI. The second contribution of our scheme emphasizes on pulmonary fissure segmentation. Using an ingeniously designed framework, adhering clutters are reasonably removed. Finally, the surface fitting model, fissure line strength, and orientation are fused to complete the incomplete fissure surface for lobar boundary estimation. The proposed method is validated on 15 CT examinations. Experimental results showed that the described method has a good performance in pulmonary fissure detection and lung lobe segmentation.
The LOLA11 dataset can be downloaded from
The authors declare no conflicts of interest.
Yuanyuan Peng and Lan Peng conceptualized the study, developed the methodology, and contributed to data curation and software. Hualan Zhong, Zheng Xu, and Hongbin Tu validated the study. Xiong Li performed formal analysis and supervised the study. Yuanyuan Peng wrote the original draft. Yuanyuan Peng and Hongbin Tu reviewed and edited the article. Hualan Zhong and Zheng Xu administered the project. Yuanyuan Peng, Hongbin Tu, and Xiong Li were responsible for funding acquisition. All authors read and agreed to the published version of the manuscript.
This research was funded by the Jiangxi Provincial Natural Science Foundation (nos. 20202BAB212004, 20204BCJL23035, 20192ACB21004, and 20181BAB202017), the Educational Science Research Project of China Institute of Communications Education (no. JTYB20-33), the Scientific and Technological Research Project of Education Department in Jiangxi Province (no. GJJ190356), and the Science and Technology Project of Changsha City (no. kq2001014).