IEEE Xplore - Login
1 of 1
http://ieeexplore.ieee.org/Xplore/login.jsp?url=http%3A%2F%2Fieeex...
Home | Login | Logout | Access Information | Alerts | Purchase History |
Login You are not logged in.
Username Password
» Forgot your password? Please remember to log out when you have finished your session. You must log in to access: • Advanced or Author Search • CrossRef Search • AbstractPlus Records • Full Text PDF • Full Text HTML
BROWSE
SEARCH
Cart | Sitemap | Help
IEEE XPLORE GUIDE
SUPPORT
Article Information A Hybrid Fuzzy Based Algorithm for 3D Human Airway Segmentation Rizi, F.Y.; Ahmadian, A.; Sahba, N.; Tavakoli, V.; Alirezaie, J.; Fatemizadeh, E.; Rezaie, N. Bioinformatics and Biomedical Engineering, 2008. ICBBE 2008. The 2nd International Conference on Volume , Issue , 16-18 May 2008 Page(s):2295 - 2298 Digital Object Identifier 10.1109/ICBBE.2008.906 Summary:Segmentation of the human airway tree from volumetric computed tomography images is an important stage for many clinical applications such as virtual bronchoscopy. The main challenges of previously developed methods are to deal with two problems namely, leaking into the surrounding lung parenchyma during segmentation and the need to manually adjust the parameters. To overcome these problems, a multi- seeded fuzzy based region growing approach in conjunction with the spatial information of voxels is proposed. Comparison with a commonly used region growing segmentation algorithm shows that the proposed method retrieves more accurate results by achieving the specificity and sensitivity of 98.81% and 85.18%, respectively. The proposed algorithm needs no manually adjustment of parameters as well as any pre-filtering process, while leading to deliver the clinically accepted segmentation result with no leakage. » View citation and abstract IEEE Members Log in by entering your IEEE Web Account Username and Password.
Access this document » Buy this document now » Learn more about subscription options » Learn more about purchasing articles and standards
IEEE Communications Society members: If you subscribe to the IEEE Electronic Periodicals Package or IEEE Electronic Periodicals Package Plus, you must access your subscription at www.comsoc.org. Users at Subscribing Institutions Check with your librarian, information professional, or system manager to determine if you need to log in. Please complete the online Technical Support Form if you need assistance. Already Purchased This Article? Select the Purchase History link to access the document. You will have 5 Days after purchase to access the Full Text PDF. Please complete the online Technical Support Form if you need assistance. Guests • Search and access Abstract records free of charge • Register for table of contents alerts • Purchase Full Text PDF documents » Learn more about subscription options or how to become an IEEE Member.
Learn more about IEEE Subscriptions
Help
Contact Us
Privacy & Security
IEEE.org
© Copyright 2009 IEEE – All Rights Reserved
2009/09/05 11:44 ظ.ق
IEEE Xplore - Login
1 of 1
http://ieeexplore.ieee.org/Xplore/login.jsp?url=http%3A%2F%2Fieeex...
Home | Login | Logout | Access Information | Alerts | Purchase History |
Login You are not logged in.
Username Password
» Forgot your password? Please remember to log out when you have finished your session. You must log in to access: • Advanced or Author Search • CrossRef Search • AbstractPlus Records • Full Text PDF • Full Text HTML
BROWSE
SEARCH
Cart | Sitemap | Help
IEEE XPLORE GUIDE
SUPPORT
Article Information A Hybrid Fuzzy Based Algorithm for 3D Human Airway Segmentation Rizi, F.Y.; Ahmadian, A.; Sahba, N.; Tavakoli, V.; Alirezaie, J.; Fatemizadeh, E.; Rezaie, N. Bioinformatics and Biomedical Engineering, 2008. ICBBE 2008. The 2nd International Conference on Volume , Issue , 16-18 May 2008 Page(s):2295 - 2298 Digital Object Identifier 10.1109/ICBBE.2008.906 Summary:Segmentation of the human airway tree from volumetric computed tomography images is an important stage for many clinical applications such as virtual bronchoscopy. The main challenges of previously developed methods are to deal with two problems namely, leaking into the surrounding lung parenchyma during segmentation and the need to manually adjust the parameters. To overcome these problems, a multi- seeded fuzzy based region growing approach in conjunction with the spatial information of voxels is proposed. Comparison with a commonly used region growing segmentation algorithm shows that the proposed method retrieves more accurate results by achieving the specificity and sensitivity of 98.81% and 85.18%, respectively. The proposed algorithm needs no manually adjustment of parameters as well as any pre-filtering process, while leading to deliver the clinically accepted segmentation result with no leakage. » View citation and abstract IEEE Members Log in by entering your IEEE Web Account Username and Password.
Access this document » Buy this document now » Learn more about subscription options » Learn more about purchasing articles and standards
IEEE Communications Society members: If you subscribe to the IEEE Electronic Periodicals Package or IEEE Electronic Periodicals Package Plus, you must access your subscription at www.comsoc.org. Users at Subscribing Institutions Check with your librarian, information professional, or system manager to determine if you need to log in. Please complete the online Technical Support Form if you need assistance. Already Purchased This Article? Select the Purchase History link to access the document. You will have 5 Days after purchase to access the Full Text PDF. Please complete the online Technical Support Form if you need assistance. Guests • Search and access Abstract records free of charge • Register for table of contents alerts • Purchase Full Text PDF documents » Learn more about subscription options or how to become an IEEE Member.
Learn more about IEEE Subscriptions
Help
Contact Us
Privacy & Security
IEEE.org
© Copyright 2009 IEEE – All Rights Reserved
2009/09/05 11:44 ظ.ق
A Hybrid Fuzzy Based Algorithm for 3D Human Airway Segmentation Fereshteh Yousefi Rizi, Alireza Ahmadian, Nima Sahba, Vahid Tavakoli Dept. Biomedical Engineering, Medical science/University of Tehran & Research Center for science and technology in Medicine, Tehran, Iran
[email protected],
[email protected] [email protected],
[email protected], Javad Alirezaie Dept. Electrical Engineering, Ryerson University Toronto, Canada
[email protected]
Abstract— Segmentation of the human airway tree from volumetric computed tomography images is an important stage for many clinical applications such as virtual bronchoscopy. The main challenges of previously developed methods are to deal with two problems namely, leaking into the surrounding lung parenchyma during segmentation and the need to manually adjust the parameters. To overcome these problems, a multiseeded fuzzy based region growing approach in conjuction with the spatial information of voxels is proposed. Comparison with a commonly used region growing segmentation algorithm shows that the proposed method retrieves more accurate results by achieving the specificity and sensitivity of 98.81% and 85.18%, respectively. The proposed algorithm needs no manually adjustment of parameters as well as any pre-filtering process, while leading to deliver the clinically accepted segmentation result with no leakage. Keywords- Airway tree segmentation, fuzzy, region growing, spatial information
I. INTRODUCTION The lungs exchange air with the external environment via the airways. The airways consist of a series of branching tubes which become narrower, shorter, and more numerous as they penetrate deeper into the lung [1]. With the development of Xray computed tomography(CT), acquisition of 3D images has become possible. Thus, it is critical to develop effective volumetric segmentation algorithms. The segmentation of airways from CT images is a critical first step for numerous virtual bronchoscopic(VB) applications [2]. An automated extraction method can be used as a diagnostic tool to help physicians pinpoint diseases [3]. Since extraction of the airway tree can provide information about the size of the airways in areas where disease is located, this procedure can be used in stent design to treat those pathological conditions. Due to this, knowledge about the airway tree structure is very important in planning surgical interventions. Even though the extraction of the human airway tree can be performed manually by experts, the 3D complexity of the airway tree makes the extraction tedious and time consuming. Manual
Emad Fatemizadeh Department of Electrical Engineering, Sharif University of Technology, Tehran, Iran
[email protected]
Nader Rezaie Pneumologist Consultant, Iran University of Medical Sciences, Firoozgar Hospital, Tehran, Iran
[email protected]
identification of the airway tree by an expert may take up to 3 hours [3]. Segmentation in medical imaging is generally considered a very difficult problem [4]. This difficulty mainly arises due to the sheer size of the datasets combined with the complexity and variability of the anatomic organs. This task becomes more difficult by the shortcomings of imaging modalities, such as sampling artifacts, noise, etc. Despite improvements in high spatial image acquisition systems, identification of the airway tree from 3D CT images is still a challenging problem because of the limitations inherent to quality of CT image acquisition. Airway-tree segmentation is a complex task for several reasons. Airway voxels are generally near -1,000 Hounsfield Units (HU), noise and partial volume effects make it impossible to use a simple threshold to identify all airway voxels within an image. Whenever mixtures of different tissue types comprise a voxel, intermediate gray level values are the result. Moreover, due to the size of the voxel, thin or stenosed airways can appear broken or discontinuous. Finally, image reconstruction artifacts, such as those introduced when a sharp high-frequency kernel is used, may cause the airways to appear discontinuous [2]. Such discontinuities may cause problems during the segmentation, resulting in both underand over segmentation errors. Furthermore, the limited intensity contrast between the participating materials (air, blood, and tissue) increases the segmentation difficulties. Airway segmentation methods can be categorized into four main groups namely knowledge-based techniques, region growing, central-axis analysis, and the mathematical morphology. The technique proposed by Sonka uses an anatomic knowledge base describing structural relationships between airways and neighboring pulmonary vessels. The knowledge-based rules are applied to the image on a sectionby-section basis. A fuzzy logic approach was later added by Park et al to improve specificity [2]. These approaches make the method very knowledge-dependent. The two later methods are told as time consuming and complex approaches.
978-1-4244-1748-3/08/$25.00 © 2008 IEEE
Region growing algorithms have been used extensively because of its simplicity, speed and flexibility. Its flexibility in terms of adding flexible rules based on the structural properties of the object has made it an efficient segmentation method. 3D Region growing based methods simply use voxel connectivity and a threshold to identify regions, usually through 26-connectivity [2]. The speed of region-growing algorithm provides acceptable segmentations for most VB applications. To overcome the leakage problem, a two seeded fuzzy based region growing approach in combination with spatial information of voxels is proposed. The two seeded fuzzy region growing method takes the idea of growing two regions(airways wall and airway lumen) silmulataneously and letting them compete for voxels. The method guarantees that both resulting regions are connected in themselves due to applying the spatial membership function. This approach not only takes the advantage of region growing algorithm but also it has led to segment the airway with no significant leaks as visually inspected. II.
MATERIALS
A.
Fuzzy Image Segmentation Medical images are fuzzy as they can be are characterized as a composition of signal intensities specific to different tissue types, noise, blurring, background variation, partial volume. Segmentation approaches should utilize this fact to extract information from images and retain fuzziness as realistically as possible. One way to achieve this is to use fuzzy clustering algorithm to generate corresponding membership functions for different regions in the image [5]. Fuzzy clustering algorithms are the most popular and extensively used segmentation methods to process such fuzziness concept [6]. The two most well known features used by clustering methods are the pixel intensity and the spatial location of the pixels [7], [8], [9], [10]. Clustering cannot separate image regions which have similar pixel intensities by considering only their pixel intensities, though they may be able to do by exploiting information on the location of pixels. In the same way, adjacent regions having different pixel intensities cannot be segmented by only considering pixel locations. B.
Fuzzy C-means Algorithm The fuzzy C-means algorithm (FCM) has been utilized in a wide variety of image processing applications such as medical imaging [5]. Its advantages include a straightforward implementation, fairly robust behavior, and the ability to model uncertainty within the data. A major disadvantage of its use in imaging applications, however, is that FCM does not incorporate information about spatial context. Mathematically, FCM is derived to minimize the following objective function with respect to the membership functions u jk and centroids ν k
C
J FCM = ¦¦ u qjk y j − v k
2
(1)
j∈Ω k =1
Where y j is the observation at pixel j, C is the number of clusters or classes, and Ω is the image domain. The membership functions are constrained to be positive and to satisfy C
¦u
jk
=1
(2)
k =1
Here, the objective function is minimized when high membership values are obtained in areas where the observations are close to the centroid, and low membership values are obtained where observations are distant from the centroid. The parameter q is a weighting exponent that satisfies q 1 and controls the degree of ‘‘fuzziness’’ in the resulting membership functions. As q approaches unity, the membership functions become crisper, and approach binary functions. As q increases, the membership functions become increasingly fuzzy [11]. By iteratively updating the cluster centers and the membership grades for each data point, FCM iteratively moves the cluster centers to the “right” location within a data set [5]. C. The Spatial Membership Function Most of the methods on fuzzy pixel classification are based on the fuzzy c-means (FCM) algorithm. It does not lead to a properly so-called segmentation, since it only classifies the pixels into fuzzy classes and does not create fuzzy regions [12]. There exists a high degree of correlation among the neighboring pixels of an image which possess similar feature values belong to the same cluster. This spatial relationship can impact the performance of clustering in FCM algorithm. To exploit this effect we use H as the uniform window of size 3×3 to obtain spatial membership function of the current stage that is generated by convolution the gray level membership function. D. Region Growing In region growing methods, the issue is to find the seeds of regions, and the function linking region homogeneity, and membership grades. Region growing methods is carried out by fuzzy rules involving fuzzy criteria such as region homogeneity, region size or gradient sharpness [12]. This method mainly suffers from partial volume effects and noise due to the global threshold used during segmentation. The “optimal” thresholds differ for large versus small airways because of these factors. The segmentation result tends to become less fine in details of the airways, in particular where small branches are located, and contains rough edges. Many contributions have been done to overcome the drawbacks of region growing. Summers et al. [10] used 3D seeded region growing and a manually selected threshold to segment airways for rendering. The algorithm proposed by A.P. Kiraly et al. [2] used a modified adaptive 3D region growing algorithm to extract lung regions from CT.
In this paper, we proposed a two seeded fuzzy region growing method that utilizes gray level membership functions obtained automatically from applying FCM algorithm. To extract airway regions more accurately the spatial information of voxels is also used in conjunction with the gray level information. III.
METHOD
a. Segmented branches with no leakage
At the first stage, the gray level membership function µ is generated by applying the FCM algorithm on the preprocessed volume to cluster the whole dataset. Then, the airway class which determines the airway region can be accepted if the outer region, made by dilation of the corresponding airway region, belongs to the lumen airway cluster. After that, the total membership function µG is generated. The following equations show the iterative process that updates the airway regions: i G
µ si −−1updated = H * µ si −1
(3)
µ si = max( µ si −−1updated , µ Ti − 1 )
(4)
µTi = min( µ si , µ Gi )
(5)
Where µ si −−1updated is obtained by convolving the spatial membership function of the stage i − 1 , µ si −1 with H that is explained before. This function is used to obtain the spatial membership function of the current stage, µ si , recursively as explained in equation (4) which is the max operation in fuzzy setting. µTi −1 is in fact the total membership function of the airway region obtained by applying min operation on gray level membership function µ Gi and the spatial membership function, µ si in each stage (equation 5). The equations (4) and (5) describe the 3D region growing process by which candidate voxels are being added to the airway segmented region as far as µ G is greater than µ s . Finally, the region growing process terminates when this condition fails which happens when the growing enters to the lung parenchyma. µ
G
IV.
RESULT
In this work, the dataset of size 512×512×430 with 0.6 mm z resolution, obtained from multi-detector CT scanner at Imaging Center of Emam Complex Hospital, Tehran was used. We first applied our algorithm on slices of 3D dataset, individually. As it shows in Fig. 1, the segmentation results are fairly acceptable but the connectivity of the branches is not satisfied and leaking becomes appeared while reaching lower airway branches. This connectivity can be cured by applying the algorithm in 3D domain in which all information of spatial orientations is involved. In order to show the robustness of our developed hybrid method, the traditional region growing method and the
Leakage b. Segmented branches appeared with leakage shown by red arrow Fig.1 The results of the segmentation method in 2D
proposed fuzzy based method were applied separately to segment the left and right bronchial tree. The results are shown in Fig. 2. As it shows the left branch which is segmented by RG method is appeared with severe leakage throughout the lung parenchyma. The result of segmentation on the right branch proves the ability of our proposed method to complete a visually acceptable segmentation. In order to evaluate our results in a quantitative manner, at first a manual segmentation was carried out by an expert radiologist to obtain a clinically accepted airway tree. Then, a quantitative evaluation was carried out to obtain the sensitivity, specificity and the resulting accuracy. The quantitative results are shown in Table 1. Sensitivity is to identify the ability of the algorithm to segment the correct airway regions as defined below: Sensitivity =
TP TP + FN
(9)
Specificity is the ability of the algorithm to detect nonairway regions, correctly. Specificity =
TN TN + FP
(10)
Accuracy specifies the ratio of the correctly detected regions to all detected regions as follow: Accuracy =
TP + TN TP + FN + TN + FP
(11)
The TP, TN, FN and FP quantities in the above equations are defined as below: TP: Pixels that are airway and they are detected correctly. TN: Pixels that are not airway and they are not detected. FP: Pixels that are not airway but they are detected as airway. FN: Pixels that are airway, but they are not detected by algorithm.
TABLE I.
Accuracy % 92.92
RESULT OF QUANTITY EVALUATION
Specificity %98.81
Sensitivity %85.18
Figure2. Left Branch is segmented by RG method and the right branch is obtained by applying our method with 700 iterations.
The quantitative analysis of the results shows the sensitivity and specificity of 85.18% and 98.81%, respectively. The result of our algorithm in segmenting the airway tree is shown in Fig.2. As shown the main branches are clearly segmented with no leakage. The iteration number of the algorithm can be increased to detect more branches while no leaks. V. CONCLUSION Fuzzy property is the nature of images. When the images are segmented by fuzzy methods, the inaccuracies property of the elements is adopted. The performance of the proposed algorithm was found very effective in segmenting the upper airway. The new segmentation algorithm proves to be considerably more robust than region-grow-based airway segmentation in terms of detecting accurate branches with no leakage. In addition, no initial parameters need to be tuned by the user and no preprocessing is needed. In many cases, the new algorithm outperforms traditional region growing based segmentation methods due to exploiting the combination of the spatial information with voxel intensities in a fuzzy based manner. It can be seen that our proposed method could not only automatically select initial features, but has led to a desirable accuracy We are currently working on the optimization of the method to improve the segmented results in order to reach more generations where very small branches are present. ACKNOWLEDGMENT We would like to thank Research Centre for Science, Technology in Medicine (RCSTIM), Medical Informatics Group, of Tehran University of Medical Sciences, for providing valuable resources. Thanks everyone at RCSTIM for help, good ideas and being part of a pleasant work environment.
Figure2. Result of airway segmentation by our method with 550 iterations
REFERENCES [1]
Deniz Bilgen, “Segmentation and Analysis of the Human Airway Tree from 3D X-ray Ct Images,” Master Thesis, Graduate College, The University of Iowa, 2000. [2] Atila P. Kiraly, MS William E. Higgins, PhD, Geoffrey Mclennan, MD, PhD, Eric A. Hoffman, PhD, Joseph M. Reinhardt, PhD, “Threedimensional Human Airway Segmentation Methods for Clinical Virtual Bronchoscopy,”Academic Radiology, Vol 9, No 10, October 2002. [3] Juerg Tschirren, Eric A. Hoffman, Geoffrey Mclennan, Milan Sonka, “Intrathoracic Airway Trees: Segmentation and Airway Morphology Analysis From Low-Dose CT Scans,” IEEE Transactions on Medical Imaging, vol.24, No. 12, December 2005. [4] Sarang Lakare,”3D Segmentation Techniques for medical Volumes,” Research Proficiency, Department of Computer Science State University of New York at Stony Brook, 2000. [5] James C. Bezdek, James Keller, Raghu Krisnapuram, Nikhil R. Pal, “Fuzzy Models And Algorithms For Pattern Recognition And Image Processing,” 2005. [6] Songul Albayrak, Fatih Amasyali, “ Fuzzy C-means Clustering on medical diagnostic systems.” International XII, Turkish Symposium on Artificial Intelligence and Neural Network-TAINN, 2003 [7] J.C. Bezdek, Pattern Recognition with Fuzzy Objective Function algorithm, New York: Plenum Press, 1981 [8] R. Krishnapurnam, J.M. Keller, “ A Possibilistic Approach to Clustering.” International Journal of Fuzzy Systems,2|(2), 98-110, 1993. [9] Y. A. Tolias and M. Panas, On applying Spatial Constarints in Fuzzy Image Clustering Using a Fuzzy Rule-Based System, IEEE Intern. Con. On Signal Processing Letters, 1998, 5(10), 245-247. [10] A. W. C. Liew, S. H. Leung and W. H. Lau, Fuzzy image clustering incorporating spatial continiuty, IEE Proc.-Vis on Image Signal Process, 200, 147(2), 185-192. [11] Dzung L. Pham, “Spatial Models for Fuzzy Clustering,” Computer Vision and Image Understanding 84, 285-297, 2001. [12] Summers RM, Feng DH, Holland SM, Sneller MC, Shelhamer JH., “Virtual bronchoscopy: segmentation method for real time display.” Radiology 1996; 200:857-862.
