A robust and efficient framework for tubular structure segmentation in chest CT images

Abstract

BACKGROUND:

Tubular structure segmentation in chest CT images can reduce false positives (FPs) dramatically and improve the performance of nodules malignancy levels classification.

OBJECTIVE:

In this study, we present a framework that can segment the pulmonary tubular structure regions robustly and efficiently.

METHODS:

Firstly, we formulate a global tubular structure identification model based on Frangi filter. The model can recognize irregular vascular structures including bifurcation, small vessel, and junction, robustly and sensitively in 2D images. In addition, to segment the vessels from JVN, we design a local tubular structure identification model with a sliding window. Finally, we propose a multi-view voxel discriminating scheme on the basis of the previous two models. This scheme reduces the computational complexity of obtaining high entropy spatial tubular structure information.

RESULTS:

Experimental results have shown that the proposed framework achieves TPR of 85.79%, FPR of 24.83%, and ACC of 84.47% with the average elapsed time of 162.9 seconds.

CONCLUSIONS:

The framework provides an automated approach for effectively segmenting tubular structure from the chest CT images.

Keywords

Tubular structure segmentation computer-aided diagnosis lung nodule Frangi filter multi-view discriminating scheme

1. Introduction

Computer-aided diagnosis (CAD) system for lung nodule based on chest CT images is an essential tool for early lung cancer detection, which can improve the 5-year survival rate from 18% to 70% [1, 2, 3, 4]. False positives (FPs) reduction plays a crucial role in the CAD system since it can improve the accuracy and efficiency of the subsequent steps [5]. As most of the FPs are vessels or airways, tubular structures segmentation method can be regard as an effective approach for FPs reduction.

Automatic tubular structure segmentation methods have been widely studied over decades [6], the existing such methods can be divided into two categories: supervised methods and unsupervised methods [7], and the unsupervised vessel segmentation methods can be further divided into tracking-based, model-based, and filter-based methods [8]. It should be mentioned that the filter-based methods are most used in tubular structures segmentation [9], and Hessian matrix is a main approach for designing the filters, such as the Frangi filter [10]. To date, the Hessian-based filters are prevailing in many CAD systems because of their excellent performance in distinguishing tubular structure and blob structure [11], which are the typical characteristic of vessels and nodules, respectively. For example, in the work developed by Gonçalves et al. [12], the authors propose a multi-scale lung nodule segmentation method with a central medialness adaptive principle. This principle uses Hessian matrix to provide good segmentation results in blob structures for the method.

As is mentioned above, the Hessian-based filter is an effective approach for segmenting vessels or airways in chest CT images, but some problematic disadvantages in these methods still cannot be overlooked. Firstly, local blob structures are often formed in irregular vascular cross-section, and low response of these regions on Hessian filters will influence the accuracy of tubular structures segmentation. Moreover, Hessian-based filter cannot suppress nodule regions completely, and this will make it difficult to segment vessels from the JVN. Finally, Hessian-based filters cannot distinguish nodule cross-section and the blob structure vascular cross-section in a single 2D view. Although 3D Hessian-based filters can be used to solve this problem, these methods are often computationally expensive due to their processing of each voxel and its neighborhood in the image at several scales.

To solve these problems, we propose a multi-view tubular structure segmentation framework in this work. Specifically, we firstly formulate a global tubular structures identification model based on the Frangi filter response rate of each region of interest (ROI). This model utilizes the global characteristics of tubular structures to eliminate the influence from the local incompleteness of morphological features caused by the low response to the local blob structure, and thus the complex tubular structure can be recognized robustly in a 2D CT image. Furthermore, focusing on the vessel segmentation of JVN, we design a local tubular structures identification model that uses sliding window to smooth the JVN regions enhanced by Frangi filter, and judges the tubular structure sub-region with an intensity threshold. With the model, the adhesive vessels can be segmented from JVN accurately. Finally, we present a multi-view tube voxel discriminating scheme that synthesizes the tubular structure detection results of three orthogonal views, i.e. sagittal view, coronal view, and transverse view. This scheme is helpful to improve the accuracy of distinguishing nodule and vessel, and as it converts all the processing into 2D calculations from 3D calculations, the computational complexity can be drastically reduced. Compared with the state-of-the-art methods, the experimental results demonstrate that the proposed framework can improve the robustness and efficiency of tubular structure detection.

The rest of this paper is organized as follows. Section 2 presents the detailed design on global tubular structure identification model, local tubular structure identification model, and multi-view tube voxel discriminating scheme in our work. Experimental results and discussion are reported in Section 3. Finally, the paper is concluded in Section 4.

Figure 1.

General schema used in this work: (a) Image initialization; (b) Multi-view reconstruction; (c) Global tubular structure segmentation; (d) Local tubular structure segmentation; (e) Tube discrimination.

2. Materials and methods

The overview schema of our work is depicted in Fig. 1. The proposed framework depends on five modules: (1) Image initialization: the original CT image sequence is processed by lung parenchyma segmentation and ROI extraction; (2) Multi-view reconstruction: the processed image sequence is reconstructed into three slice sequences at three views; (3) Global tubular structure segmentation: the ROIs in these slices are enhanced by Frangi filter, and their response rates are computed to identify tubular structure. (4) Local tubular structure segmentation: the enhanced JVN region is smoothed by the sliding window, and the vessels in the JVN are segmented with an intensity threshold. (5) Tube discrimination: the relationship between each voxel and its corresponding tube is judged with all the discriminating results from the three views.

2.1 Image acquisition

The chest CT databases used in this work are provided by Lung Image Database Consortium (LIDC) [13], VESSEL 12 challenge [14] and our co-operator Affiliated Hospital of Liaoning University of Traditional Chinese Medicine (LNUTCM) [15]. As we all know, LIDC is a publicly available reference database of low-dose helical CT images, which consists of 1018 cases [13]. The dataset from VESSEL12 challenge contains 20 CT sets with lung masks [14]. Furthermore, LNUTCM is a grade-A hospital in China [15], and the CT database supplied from LNUTCM contains 48 cases.

To quantitatively evaluate our methods, the data of CT images with tubular structure labels is required in the work. But no gold standard existed in advance for labeling tubular structures in chest CT images. Hence, artificially defined tubular structure labels for our databases are used in the validations. These labels are provided by sixteen radiologists from LNUTCM with 8–10 years of experience in general radiology, and the processed images will be further verified by a thorax surgeon with 30 years of experience. We select 108 CT sets, including 40 CT sets from LIDC, 20 CT sets from VESSEL12 challenge and 48 CT sets from LNUTCM, to label tubular structure regions for the experiment.

2.2 Tubular structure segmentation

In this section, we propose a multi-view tubular structure segmentation framework. To shed light on our work, we firstly introduce the preprocessing including the image initialization and multi-view reconstruction. Then we present a global tubular structure identification model which is used to detect the tubular structure regions in each view. Furthermore, we show a local tubular structure identification model that can be used to remove the vessels in JVN. Finally, we present a multi-view tube voxel discriminating scheme and discuss its availability in our work.

2.2.1 Preprocessing

This stage in our work mainly includes three parts: lung parenchyma segmentation, ROI extraction and multi-view reconstruction. The lung parenchyma segmentation is used to remove the irrelevant regions in the CT image, such as torso and bed board, and the ROI extraction can eliminate noise in the lung parenchyma area. Considering lung parenchyma and ROI have obvious intensity difference with surroundings, we adopt the basic global thresholding method [16] in the two parts. Then, in the multi-view reconstruction, the processed CT set is reconstructed into 3 slice sequences at 3 orthogonal views, i.e., sagittal view, coronal view, and transverse view, and all the subsequent stages are performed on these reconstructed slices.

2.2.2 Global tubular structures detection

In this subsection, we aim to recognize global tubular structure cross-sections in the slices reconstructed from 3 orthogonal views. Focusing on the problem that Frangi filter is often affected by the local blob structure appeared in irregular cross-sections, we formulate a global tubular structure identification model to improve the robustness of Frangi filter on detecting tubular structure with irregular morphological characteristics.

The Hessian matrix of the pixel ( $x, y$ ) with the Gaussian filter at scale $s$ is shown in Eq. (1) and the response result function of Frangi filter is defined using the eigenvalues of the Hessian matrix to measure tubular structure degree shown in Eqs (2)–(5):

$\displaystyle H\left({x,y,s}\right)=\left[\begin{array}[]{cc}{I_{xx}\left({x,y% ,s}\right)}&{I_{xy}\left({x,y,s}\right)}\\ {I_{xy}\left({x,y,s}\right)}&{I_{yy}\left({x,y,s}\right)}\\ \end{array}\right]$ (1) $\displaystyle V_{o}\left({x,y,s}\right)=\left\{\begin{array}[]{ll}0&\textit{if% }\ \lambda_{2}>0\\ \exp\left(-\frac{R_{B}^{2}}{2\beta^{2}}\right)\left(1-\exp\left(-\frac{R_{A}^{% 2}}{2c^{2}}\right)\right),&\textit{otherwise}\\ \end{array}\right.$ (2) $\displaystyle R_{A}=\sqrt{\sum_{j\leqslant 2}\lambda_{j}^{2}}$ (3) $\displaystyle R_{B}=\lambda_{1}/\lambda_{2}$ (4)

where $V_{0}(x,y,s)$ denotes the response result of Frangi filter to pixel ( $x,y)$ at scale $s$ ; $\lambda_{i}(i=1,2)$ are the eigenvalues of $H(x,y,s)$ , $R_{A}$ in Eq. (3) is the measure of second-order structures, and $R_{B}$ in Eq. (4) is the shape measure accounting for the eccentricity of the second-order ellipse; $\beta$ and $c$ in Eq. (2) are thresholds that control the sensitivity of the filter to $R_{A}$ and $R_{B}$ . The method integrates the tubular structure degree measures provided by the filter response at different scales and uses the maximum of $V_{0}(x,y,s)$ to be the final estimation of tubular structure degree in Eq. (5), where $s_{\textit{min}}$ and $s_{\textit{max}}$ are the minimum and maximum scales at relevant structures for covering the range of tube widths.

$\displaystyle V_{\textit{max}}\left({x,y}\right)={\max}_{s_{\min}\leqslant s% \leqslant s_{\max}}V_{o}\left({x,y,s}\right)$ (5)

In the enhancement results of Frangi filter, the rate of local low response regions in the tubular structure is far less than the blob structure. Based on that, we present Eq. (6) that calculates the response rate to distinguish these two structures:

$\displaystyle\textit{RspFgi}\left(\textit{CR}\right)=\frac{\mathop{\sum}% \nolimits V_{\textit{max}}\left({x,y}\right)}{\mathop{\sum}\nolimits I\left({x% ,y}\right)},\left({x,y}\right)\in\textit{CR}$ (6)

where CR is a ROI, $I(x,y)$ is the gray value of pixel ( $x, y$ ) in CR. On the basis of the results of Eq. (6), our global tubular structure identification model is shown in Eq. (7), where threshold $T_{\textit{RF}}$ is set for RspFgi():

$\displaystyle\textit{Tubular}\left(\textit{CR}\right)=\left\{\begin{array}[]{% ll}1,&\textit{RspFgi}\left(\textit{CR}\right)>T_{\textit{RF}}\\ 0,&\textit{otherwise}\\ \end{array}\right.$ (7)

This model eliminates the influence of the local low response in endpoints, vascular junctions, or irregular cross-sections, and improves the robustness of tubular structures detection by utilizing global characteristics of enhancement result.

2.2.3 Local tubular structures detection

In the CAD system of lung nodules, segmenting vessels from JVN is necessary since JVN is very common and the peripheral vessel may affect the malignancy level classification of the nodule. Thus, in this subsection, we concern to remove the vessel from JVN.

In general, Frangi filter can only suppress the blob structure at the center region, and the boundary region will be remained, which makes the enhanced blob structure become a circular structure. Notably, this remaining boundary region adheres with vessel, and eliminating the region is crucial to separate nodule from vessel. For this purpose, we design a local tubular structures identification model that expands the suppressed region of nodule to the boundary circular structure region with a sliding window. With this model, the vessel can be separated from JVN exactly. There are three main steps in the model: dilating the suppressed regions, smoothing ROI with a sliding window, and intensity threshold segmentation. Their details will be described below.

As we all known, the dilation can expand an area in multiple directions by adding pixels around the boundary with a structuring element. Specifically, we firstly extract the suppressed regions by computing the difference between the original JVN patch and the enhanced patch by Frangi filter, as shown in Fig. 2d. Then we dilate the suppressed region with a structuring element shown in Fig. 2e. The JVN after dilation is presented in Fig. 2f.

Figure 2.

Procedures of the local tubular structures identification model: (a) Original CT image, in which a JVN is marked inside a red box; (b) JVN region; (c) JVN region processed by Frangi filter; (d) Suppressed region; (e) Suppressed region with dilation; (f) JVN region after the suppressed region is dilated, and the sliding window is marked as yellow box which contains 3 $\times$ 3 blocks marked green; (g) JVN region smoothed by sliding window; (h) Detection result of local tubular structures identification model.

After the dilation, we utilize sliding window to smooth the whole JVN patch. This approach can expand the suppressed region to the whole nodule without covering the vessel, and we can take advantage of the difference to separate these two sub-regions. The sliding window designed in our work is shown in Fig. 2f. The JVN patch is divided into $m\times n$ blocks that has green border in the figure. The sliding window labeled yellow border contains 3 $\times$ 3 blocks and stride is set as one block in horizontal and vertical direction, respectively. We smooth the pixels in the sliding window with Eq. (8):

$\displaystyle\textit{SmoothFg}(P_{i}^{k})=\left\{\begin{array}[]{ll}0&P_{i}^{k% }\notin S_{k}\\ \frac{\mathop{\sum}\nolimits_{i=1}^{N}\textit{Grayvalue}\left({P_{i}^{k}}% \right)}{N},&P_{i}^{k}\in S_{k}\\ \end{array}\right.\left(0\leqslant k\leqslant m\times n\right)$ (8)

where $P_{i}^{k}$ denotes $i$ th pixel in the $k$ th block in the patch, $S_{k}$ is the intersection of JVN region and block, $N$ is the number of pixels in $S_{k}$ , Grayvalue( $x$ ) is a function that obtains the intensity of $x$ and SmoothFg( $P_{i}^{k}$ ) is a function that returns the intensity of the pixel $P_{i}^{k}$ after smoothing.

The processing of sliding window expends the suppressed region to the whole nodule region as shown in Fig. 2g, and we can use the difference of intensity between vessel and nodule to separate them. We select $T_{sp}$ to be the intensity threshold and the pixels below $T_{sp}$ are set to 0, and then the remaining high intensity region is judged as vessel that is shown in Fig. 2h.

2.2.4 Multi-view voxel discrimination

The previous two subsections focus on the 2D tubular structure detection. However, many of the blob structure cross-sections are vessels and they may be missed in these methods. As we all known, a nodule is sphere that presents blob structure at any view, and a vessel is tube that may also show blob structure at some certain views. This means that the entropy of the vascular cross-section presenting blob structure is low, and it’s the main reason of high FPs rate in nodule detection. Generally, 3D filters are the effective approaches to solve this problem by obtaining spatial information. But the methods often suffer the high computational complexity since processing voxels is much more complex than pixels. In fact, most of the spatial structure information obtained by 3D filters is redundant, and simplifying 3D information can reduce the computational complexity dramatically. Therefore, we present a multi-view voxel discriminating scheme that evaluates a voxel with the tubular structure detection results from the three orthogonal views mentioned in 2.2.1. This method ensures the emergence of high entropy cross-section in distinguishing nodule and vessel. Simultaneously, as 3D calculations are converted into 2D calculations, the method costs low computation for generating an accurate result.

In our scheme, the view selection is the main problem. Specifically, the views selected in our work are sagittal view, coronal view, and transverse view. The reason for that is these views can maximize the entropy of the cross-sections. As we all known, the intersection angle between the views has positive correlation with the difference of the objects in these views, and bigger difference usually generates higher entropy. Sagittal view, coronal view and transverse view are orthogonal each other, and thus they can produce the greatest intersection angles among all view combinations. In addition, as these views are parallel or perpendicular to the main elongation direction of vessel tree and airway tree, their tubular structure features are more typical. Therefore, the entropy of the cross-sections obtained from our view selection is the highest comparing with other views combinations.

With the view selection above, our multi-view tube voxel discriminating scheme is shown in Eqs (9)–(12), where $g_{x_{i}}\left({y,z}\right)$ , $g_{y_{i}}\left({x,z}\right)$ , and $g_{z_{i}}\left({x,y}\right)$ are the decision result of a pixel at the position ( $y, z$ ), ( $x, z$ ) and ( $x, y$ ) in the reconstructive plane $x_{i}$ , $y_{i}$ , and $z_{i}$ , respectively. CR ${}_{x}$ , CR ${}_{y}$ and CR ${}_{z}$ are connected regions in the corresponding views that contain pixels ( $y, z$ ), ( $x, z$ ) and ( $x, y$ ), respectively. The $g(x_{i},y_{i},z_{i})$ denotes the final discriminating result of the pixel at position ( $x_{i}$ , $y_{i}$ ) in $z_{i}$ th slice of CT sequence.

$\displaystyle g_{x_{i}}\left({y,z}\right)=\left\{\begin{array}[]{ll}1,&CR_{x}% \ \textit{is tubular structure}\\ 0,&\textit{otherwise}\\ \end{array}\right.$ (9) $\displaystyle g_{y_{i}}\left({x,z}\right)=\left\{\begin{array}[]{ll}1,&CR_{y}% \ \textit{is tubular structure}\\ 0,&\textit{otherwise}\\ \end{array}\right.$ (10) $\displaystyle g_{z_{i}}\left({x,y}\right)=\left\{\begin{array}[]{ll}1,&CR_{z}% \ \textit{is tubular structure}\\ 0,&\textit{otherwise}\\ \end{array}\right.$ (11) $\displaystyle g\left({x_{i},y_{i},z_{i}}\right)=g_{z_{i}}\left({x_{i},y_{i}}% \right)\cup g_{y_{i}}\left({x_{i},z_{i}}\right)\cup g_{x_{i}}\left({y_{i},z_{i% }}\right)$ (12)

3. Experimental results and discussion

This section presents and analyzes the experimental results of the proposed framework for evaluating the performance in tubular structure segmentation. The experiment consists of two parts: (i) results illustration; (ii) quantitative evaluation. The CT sets are obtained from LIDC [13], VESSEL 12 challenge [14] and LNUTCM [15]. All the experiments are conducted on 8 GB RAM, Intel Core i7 processor with 3.60 GHz, and Windows 10 operating system. The method is implemented using the MATLAB R2018b Win64.

Figure 3.

Illustrative tubular structure segmentation results using the global tubular structure identification model in each view. Each group contains three images sequences that display the tubular structure segmentation results in three orthogonal views, respectively. From top to bottom: transverse view, sagittal view, and coronal view. From left to right: original images, binary image, Frangi filter response result images, tubular structure region determination (marked red), and tubular structure segmentation results. The arrows point to nodules, and the diameter of the nodules in (a) and (b) are 6.25 mm and 23.97 mm, respectively.

3.1 Results illustration

In the Fig. 3, we present the process and the results of the global tubular structure identification model at each view, respectively. In the first column, the nodule in Fig. 3a is a small juxta-pleural nodule (JPN), and in Fig. 3b is a JVN with large diameter. The second column shows that the small nodule in Fig. 3a is remained, and the most of the noise is removed, simultaneously. It can be observed in the third column that the response rates of tubular structure regions are obviously higher than the blob structure regions. The fourth column shows that our method can separate vessels or airways from nodules accurately, and this proves that the response rate is a remarkable characteristic to distinguish vessels from nodules. Finally, the segmentation results are shown in the fifth column. Notably, the small nodule region in Fig. 3a is remained in all the three views, which means that it will not be segmented in the final result.

Figure 4.

Results of the multi-view tube voxel discriminating scheme. Each column presents a sample group. From top to bottom: original image, discriminating results of three orthogonal views, final detection results of our scheme, and manual annotation. It should be note that, the second row presents the discriminating results with different colors, that a region marked cyan represents it is judged as tubular structure only in transverse view, yellow represents in sagittal view, pink represents in coronal view, light green represents in transverse view and sagittal view, blue represents in transverse view and coronal view, orange represents in sagittal view and coronal view, and dark green represents in transverse view, sagittal view and coronal view, simultaneously.

Figure 4 presents the detection result from each view and their corresponding final results in the second and third rows, respectively. From these images, we can find that the proposed scheme has high sensitivity for most tubular structure regions labeled manually in the fourth row. And it should be mentioned that the proportions of the regions marked cyan, yellow, and pink in the second row are close. This indicates that the information redundancy of our view combination is low, and the information gain of each view is balanced. However, we can also find that many endpoints of small vessels are missed by our method. This is because that these regions are short, and thus their tubular structure characteristic is not obvious.

Figure 5 shows the results of applying different tubular structure segmentation methods to 2 samples. Three other state-of-the-art tubular structure segmentation methods are selected for comparison: Minimum Spanning Superpixel Tree (MSST) [17], Multi-scale Superpixel Chain Tracking (MSCT) [18], and Hierarchical Image Matting Model (HIMM) [19]. For a fair comparison, the parameters of these methods are optimized for the best performance. MSST has two parameters that $\theta$ is varied iteratively from (0,170) and $a$ is set to 1.5. MSCT is involved with $s_{1}=$ 28, $s_{2}=$ 16, $s_{3}=$ 5 for three SLIC processes, respectively, and the superpixel filter parameters $r_{v}=$ 220, $r_{\textit{var}}=$ 1000. HIMM uses four threshold parameters for region features that $e_{1}=$ 0.3, $e_{2}=$ 0.18, $e_{3}=$ 2.5, $e_{4}=$ 0.53. Our framework has three parameters. We set $\beta=$ 0.50 and $c=$ 15, which are referred to Frangi filter. The response rate threshold $T_{\textit{RF}}=$ 0.90.

Figure 5a shows the tubular structure segmentation results of these related methods. It can be seen from the figures that the proposed framework can not only detect the bifurcations and disconnected vessel, but also avoid the influence of the low-intensity regions in the main vessel. The reason for the performance is that our method focuses on the global tubular characteristic, and thus is less affected by the local deformation. This embodies the robustness and sensitivity of the framework.

Besides, Fig. 5b shows the results of these methods on the vessels in the JVN. All the 3 vessels adhering with the nodule are detected by the proposed framework, and this indicates that it has the best performance compared with other methods. The reason for the results is that the sliding window can merge the nodule subregion with the attenuation region locally, and this improves the precision of the method.

Table 1

Comparison of the method with published studies of TPR, FPR, ACC and runtime

Method	TPR (%)	FPR (%)	ACC (%)	Runtime (s)
MSST	81.22	16.00	81.57	257.3
MSCT	77.31	33.50	75.97	159.6
HIMM	72.56	28.73	72.40	221.3
Proposed method	85.79	24.83	84.47	162.9

Figure 5.

Illustrative tubular structure segmentation results on two samples. The first row is global images, the second row is local enlarged images. The first column is original images, the second column is manual annotation, the third to sixth columns are the segmentation results of MSST, MSCT, HIMM, and the proposed framework, respectively.

3.2 Quantitative evaluation experiment

True Positive Rate (TPR), False Positive Rate (FPR) and Accuracy (ACC) are adopted as evaluation metrics to verify the effectiveness of the framework. They are defined in Eqs (13)–(15), where TP, FN, TN, FP are the numbers of true positive pixels, false negative pixels, true negative pixels, and false positive pixels, respectively. We present the TPR, FPR and ACC of the all related methods in Table 1. Notably, there are 126 JVNs, including 637 JVN patches, with different nodule types used in this experiment.

$\displaystyle\textit{TPR}=\frac{\textit{TP}}{\textit{TP}+\textit{FN}}$ (13) $\displaystyle\textit{FPR}=\frac{\textit{FP}}{\textit{FP}+\textit{TN}}$ (14) $\displaystyle\textit{ACC}=\frac{\textit{TP}+\textit{TN}}{\textit{TP}+\textit{% FP}+\textit{TN}+\textit{FN}}$ (15)

As can be seen from Table 1, compared with the other three methods, i.e., MSST, MSCT and HIMM, the proposed framework has the highest TPR and ACC, which proves the effectiveness of this method to detect more general lung tubular structures. The FPR of the proposed framework is 8.83% higher than MSST. The main reason for this is the instability of our method on small nodule. Small area of nodule with narrow shape will increase the fluctuation of response rate, and thus they may be judged as tubular structures in certain views.

In order to quantitatively evaluate the computation of the proposed framework, we also show the runtime of each method in Table 1. In the experiment, the runtime is the average elapsed time of each method on 108 CT sets. As can be seen from Table 1, the proposed framework achieves the shortest runtime compared with other three methods. This embodies its high computational efficiency on processing 3D tubular structure characteristic.

4. Conclusions

This work presents a robust and efficient framework to segment the lung tubular structure regions. In this framework, we formulate a global tubular structure identification model to recognize more general tubular structure cross-sections robustly. Moreover, we design a local tubular structure identification model to segment the vessels from JVN. Finally, we propose a multi-view voxel discriminating scheme based on the previous two models to improve the accuracy and the efficiency of the framework in distinguishing nodule and vessel. In the experiment, the proposed framework achieves promising results in TPR (85.79%), FPR (24.83%), and ACC (84.47%) with the average elapsed time of 162.9 seconds.

In summary, the proposed framework is robust and efficient in tubular structure segmentation, and has great significance of malignancy level classification for lung nodules. This framework is still under development and future works should be directed at improving the specificity of tubular structure segmentation and reducing the influence of binarization.

Footnotes

Acknowledgments

The authors would like to thank the support of the National Key R&D Program of China [grant number: 2019JSJ12ZDYF01].

Conflict of interest

The authors declare that they have no conflict of interest.

Ethical approval

Ethical approval was given by the Institution Review Board (IRB) of Affiliated Hospital of Liaoning University of Traditional Chinese Medicine with the following approval notice number: 2017111CS (KT)-040-01.

References

Setio

AAA

Ciompi

Litjens

Gerke

Jacobs

van Riel

Wille

MMW

Naqibullah

Sánchez

van Ginneken

. Pulmonary nodule detection in CT images: False positive reduction using multi-view convolutional networks. IEEE Transactions on Medical Imaging. 2016; 35(5): 1160-1169.

Asija

Manickam

Aronow

Chandy

. Pulmonary Nodule: A comprehensive review and update. Hospital Practice. 2014; 42(3), 7-16.

Shen

Zhou

. Yang

Dong

Yang

Zang

Tian

. Multi-crop Convolutional Neural Networks for lung nodule malignancy suspiciousness classification. Pattern Recognition. 2017; 61(61): 663-673.

Huang

Yue

Wang

. Optimal three-dimensional reconstruction for lung cancer tissues. Technology and Health Care. 2017; 25: S423-S434.

Agam

Armato

. Vessel tree reconstruction in thoracic CT scans with application to nodule detection. IEEE Transactions on Medical Imaging. 2005; 24(4): 486-499.

Xiao

Staring

Shamonin

Reiber

JHC

Stolk

Stoela

. A strain energy filter for 3D vessel enhancement with application to pulmonary CT images. Medical Image Analysis. 2011; 15: 112-124.

Feng

Xie

Liang

Zhang

Wang

. A cross-modality learning approach for vessel segmentation in retinal images. IEEE Transactions on Medical Imaging. 2016; 35: 109-118.

Zhang

Dashtbozorg

Bekkers

Pluim

JPW

Duits

Romeny

BMTH

. Robust retinal vessel segmentation via locally adaptive derivative frames in orientation scores. IEEE Transactions on Medical Imaging. 2016; 35(12): 2631-2644.

Kovacs

Hajdu

. A self-calibrating approach for the segmentation of retinal vessels by template matching and contour reconstruction. Medical Image Analysis. 2016; 29: 24-46.

10.

Frangi

Niessen

Vincken

Viergever

. Multiscale vessel enhancement filtering. Lecture Notes in Computer Science. 1998; 1496: 130-137.

11.

Korfiatis

Kalogeropoulou

Karahaliou

, et al. Vessel tree segmentation in presence of interstitial lung disease in MDCT. IEEE Transactions on Information Technology in Biomedicine. 2011; 15(2): 214-220.

12.

Gonçalves

Novo

Campilho

. Hessian based approaches for 3D lung nodule segmentation. Expert Systems with Applications. 2016; 61: 1-15.

13.

McNitt-Gray

Armato

Meyer

Reeves

McLennan

Pais

Freymann

Brown

Engelmann

Bland

Laderach

Piker

Guo

Towfic

Qing

DPY

Yankelevitz

Aberle

van Beek

EJR

MacMahon

Kazerooni

Croft

Clarke

. The Lung Image Database Consortium (LIDC) data collection process for nodule detection and annotation. Academic Radiology. 2007; 14: 1464-1474.

14.

Rudyanto

Kerkstra

van Rikxoort

Fetita

Brillet

, et al. Comparing algorithms for automated vessel segmentation in computed tomography scans of the lung: the VESSEL12 study. Medical Image Analysis. 2014; 18: 1217-1232.

15.

Affiliated Hospital of Liaoning University of Traditional Chinese Medicine, Shenyang, LN, CHN. Intranets: Liaoning University of Traditional Chinese Medicine. Updated 2018 Nov 19. [Online]. Available: http//www.lntcm.com.cn/en/index.htm.

16.

Chaubey

. Comparison of the local and global thresholding methods in image segmentation. World Journal of Research and Review. 2016; 2(1): 1-4.

17.

Sheng

, et al. Retinal vessel segmentation using minimum spanning superpixel tree detector. IEEE Transactions on Cybernetics. 2019; 49: 2707-2719.

18.

Zhao

Yang

Danni

, et al. Automatic retinal vessel segmentation using multi-scale superpixel chain tracking. Digital Signal Processing. 2018; 81: 26-42.

19.

Fan

Wei

, et al. A hierarchical image matting model for blood vessel segmentation in fundus images. IEEE Transactions on Image Processing. 2019; 28(5): 2367-2377.