Segmentation of Lung Lobes and Nodules in CT Images

Signal & Image Processing : An International Journal(SIPIJ) Vol.1, No.1, September 2010
DOI : 10.5121/sipij.2010.1101 1
SEGMENTATION OF
LUNG LOBES AND NODULES IN CT IMAGES
Anitha. S1
and Sridhar.S2
1
Department of Computer Science and Engineering, College of Engineering, Guindy
any_kanny@yahoo.co.in
2
Department of Computer Science and Engineering, College of Engineering, Guindy
sridhar@cs.annauniv.edu
ABSTRACT
The objective of this paper is to develop a segmentation system in order to assist the surgeons to remove the
portion of lung for the treatment of certain illness such as lung cancer, and tumours. The fissures of lung
lobes are not seen by naked eyes in low dose CT image, there is a proposal for automatic segmentation
system. The lung lobes and nodules in CT image are segmented using two stage approaches such as
modified adaptive fissure sweep and adaptive thresholding. Initially pre-processing is used to remove the
noise present in CT image using filter, then the fissure regions are located using adaptive fissure sweep
technique, then histogram equalization and region growing is applied to refine the oblique fissure. Lung
Nodules are segmented using thresholding. The comparative analysis of manual and automatic segmentation
for fissure verification has been performed statistically. The analysis is made on 20 set of images.
KEYWORDS
CT, Fissure Sweeping, Lung Segmentation, Lung Nodules.
1. INTRODUCTION
Multisided CT imaging is the primary digital technique for imaging the lung for the detection of
pulmonary (lung) disease such as lung cancer, tumour, and cystic fibrosis. The clinical CT image
of the lung is shown in Figure 1 and its general anatomy is shown in Figure 2. In order to segment
lung lobes, fissures or boundaries are detected based on apriori knowledge from the anatomic atlas,
this method uses ridge operator and graph search [22]. By using 3D watershed transform lung
lobes are segmented, in this method distance map is performed on vessel mask and then combine
the original, refine the fissures using interactive watershed transform [11]. The first step for lung
lobe segmentation is lung segmentation, in this method lung regions are extracted by using
threshold and left and right lungs are segmented using dynamic programming [6]. Fissure
extraction using Vander burg’s linear feature detector and morphological operators for segmenting
lung lobes from 1mm clinical CT image [12], 0.2mm isotropic CT images [10]. Fissure near the
lesions are calculated using surface curvature analysis [9]. Boundary of an object is identified
using spline contours and iterative energy minimization formula, however this method fails when
large gaps are present [2] [3] [16]. Fissures are present in the lung where there are no vascular or
bronchial trees and hence it is necessary to segment the vascular and the bronchial trees by
considering the grey level information, 3D shape constraint, and apriori knowledge [5]. Lung lobes
are segmented using apriori knowledge [23] and 3D shape constraint [18]. In this method fissures
are identified using graph search and fuzzy reasoning. By using the Gaussian, sobel and ridge
operators fissures are detected using shape based curve growing method in 1.25mm CT image [19]
[15]. The pre requisite for the fissure detection to segment the lung lobes is lung segmentation, in
this paper segmentation by registration is used [20]. Lung lobes are segmented in clinical CT
image using adaptive fissure sweeping for finding the fissure region and using watershed
transform to enhance the fissure [17]. Pulmonary fissures are segmented by extracting the vascular

2
trees [8]. 3D visualization of lung cavities is important for surgical planning of treating lung
disease [7] [1]. Methods such as variation approach [21], graph cuts [4], guided random walks [14]
are used for finding the object boundaries, however this method requires manual intervention and
high computation. In order to identify the actual fissure locations and curvatures from the fissure
regions using discrete wavelet transform [5]. Fissure regions are identified using fissure sweeping
technique and then the locations are identified using discrete wavelet transform [24].
Figure1. Clinical CT Image Figure2. General anatomy of Human lung
2. PROPOSED TECHNIQUE
The proposed technique mainly consists of four subdivisions:
2.1 Image Enhancement and Segmentation.
2.2 Fissure Detection for identifying the fissure regions.
2.3 Fissure Refinement using region growing method.
2.4 Nodule Segmentation using Adaptive Threshold
The system architecture is presented in Figure 3.
Lung Enhancement
(using wiener filter)
Lung Segmentation
Imaging System
Lung Lobe
Segmented image
Vessel enhancement
Fissure region
detection
Fissure
Refinement
CT Lung Image
Lung Nodule
Segmentation
Lung Nodule
Segmented image
Lung Lobe
Segmentation
Figure 3.Block Diagram

3
2.1. Image Enhancement and Segmentation
Pre-processing is done to remove the noise present in the CT image. In order to remove the
Gaussian noise, Wiener filter is used. Since Wiener filter gives an estimate of the original
uncorrupted image with minimum mean square error, and hence Wiener filter is used for removing
the noise present in the lung CT image. Bronchial and the vascular trees are enhanced in the CT
image using the morphological Hat operation. The morphological operations are based on simple
mathematical concepts from set theory. These operators are particularly useful for the analysis of
binary images and common usage include edge detection, noise removal and image enhancement.
Segmentation refers to the process of partitioning a digital image into multiple segments. The goal
of segmentation is to simplify and/or change the representation of an image into something that is
more meaningful and easier to analyse. Image segmentation is typically used to locate objects and
boundaries (lines, curves, etc.) in images. More precisely, image segmentation is the process of
assigning a label to every pixel in an image such that pixels with the same label share certain
visual characteristics. In this paper adaptive threshold method is used to segment the left and right
lungs.
2.1.1. Steps for lung enhancement
Pre-processing of image is done using wiener filter. Wiener filter of size 3/3 pixel is optimum.
Resulting image contains less speckle noise. Vascular and Bronchial trees are enhanced using the
morphological operators such as the bottom and top hat filters.
2.1.2. Steps for lung Segmentation
Left and right lung are segmented using the adaptive thresholding. Lung is segmented using
thresholding and connected component labelling. Mean of the CT image is used as the threshold
value.
2.2. Fissure Detection for identifying the fissure regions
The fissure sweeping method is used for finding the fissure region, in fissure sweeping predefined
fissure angles are used to find the fissures. Fissure angle is defined as the angle between the fissure
and the flat horizontal line. For finding the fissures in the superior region of the lungs positive
fissure angle is used and to detect the fissure in the inferior region of the lung negative fissure
angle is used.
2.2.1. Steps for Fissure Region Sweeping
It uses predefined fissure angle, fissure angle is the angle between the fissure and flat horizontal
line. At each angle the technique scans the CT image to detect the empty region where fissure
could be located. Euclidean distances are used to detect the search region.
DE = DBW-DW
Where DE = Distance between the boundary and the centre point of the empty region.
DBW = Distance between the boundary and the centre point of the reference region.
DW = Distance between the boundary and centre point of the non empty region.
2.3. Fissure Refinement
Fissure locations and curvatures are identified within a fissure region. The contrast of the lobar
fissure is poor, histogram equalizer is used to increase the contrast and then fissure are refined
using region growing method. The region growing method determines the empty regions between
vascular and bronchial trees of lung lobes. The centre of empty region provides fissure’s location.

4
2.3.1. Steps for Fissure Refinement
Fissure locations are identified using region growing method. Initially the contrast is enhanced
using histogram equalization. Region growing approach is used to detect the fissure. In region
growing method fissure’s centre line is used as the seed point. The seed grows in an orthogonal
direction to that of the centre line of the fissure region found from the fissure sweeping technique.
2.4. Nodule Segmentation
Nodule is a mass of tissue located in the lung. It appears as round, white shadows on CT
image. Mass of 25mm or larger can cause cancer. Therefore it is necessary to segment lung
nodules from blood vessel at an earlier stage. Thresholding is used to segment lung nodules, since
it takes less time and provides an efficient method to segment nodules.
2.4.1. Steps for Nodule Segmentation
Thresholding method is used to automatically detect the nodules present in lung. Mean pixel
intensity of the image is taken as the threshold.
3. IMPLEMENTATION
3.1. Pre-processing for the removal of noise
The CT Images normally contains artefacts, noise which will not be suitable for further processing
and hence it has to be pre-processed to reduce the noise using Wiener filter. Wiener filter is used
because of following advantages.
Advantages of Wiener filter
i. It takes very short time to find the optimal solution as the apriori knowledge about the
image is taken into consideration.
ii. It controls the output error as it is optimal.
iii. It is straightforward to design and it is faster.
iv. It is a least mean square filter.
The disadvantages of Wiener filter is that the results are too blurred and it is spatially invariant, but
these are not observed in the experimentation result and hence Wiener filter works better when
compared with respect to other. The main constraint in the use of Wiener filtering is that signal
and noise should be Gaussian processes for optimality. The noise in the image can be modelled as
Gaussian. Gaussian is the natural way of modelling noise. Therefore Wiener filter is used for
removal of noise in CT image. Wiener filer gives an estimate of the original uncorrupted image
with minimum mean square error and the estimate is the non linear function of the corrupted
image. Wiener filters of size 3*3 are used to remove the e noise present in the CT image, two of
the error metrics are used to evaluate the filter with mean square error (MSE) and the peak signal
to noise ratio (PSNR). The mean square error is the cumulative squared error between the
compressed and the original image, whereas peak signal to noise ratio is a measure of the peak
error. The mathematical formulae for the two are
MSE = 1/MN (I(x,y)- I'(x,y)) ---- (1)
PSNR =20*log10 (255/sqrt(MSE)) ---- (2)
Where I(x,y) is the original image, I'(x,y) i the approximated version (which is actually the
decompressed image) and M,N are the dimensions of the images. A lower value for MSE means
lesser error, and as seen from the inverse relation between the MSE and PSNR, this translates to a
high value of PSNR. Logically, a higher value of PSNR is good because it means that the ratio of
Signal to Noise is higher. Here, the 'signal' is the original image, and the 'noise' is the error in
reconstruction. According to equation 2 PSNR is calculated in Table 1 for the CT images.

5
Patient # PSNR (db)
1 42.1595
2 42.884
3 37.0637
4 37.1547
Table1. Peak signal to noise ratio for the CT image
According to the equation (1) and (2) the value of mean square error and the peak signal to noise
ratio is given for the Figure 4 as 42.1595
Input CT image
Preprocessing
(using wiener filter)
Filter output image with
PSNR value = 42.1595
Figure 4. Wiener Filtered Image
3.2. Vessel Enhancement
Vascular and Bronchial tress are enhanced using the morphological operations. Morphology is a
technique of image processing based on shapes. The value of each pixel in the output image is
based on a comparison of the corresponding pixel in the input image with its neighbours. The first
step is to create structure element. Morphological operations like top hat and bottom hat filters are
used. Top-hat filtering is used to correct uneven illumination when the background is dark. Top hat
operation contains peaks of object that fit the structuring element. Bottom hat operations are used
to fill the gap between the objects.
Vessel Enhanced = Input image +top hat (Input image) - bottom hat (Input image) ----- (3).
The output is obtained from equation (3), which is given the Figure 5.

6
Input CT image
Morphological Operation
Vessel enhanced image
Figure 5.Vessel Enhanced Image
3.3. Lung Segmentation
Here the left and the right lung are segmented using adaptive thresholding. Threshold values are
selected based on the mean value of the CT image. The output of lung segmentation for CT images
is shown in Figure 6
Input CT image
Thresholding
Lung segmented image
Figure 6.Lung Segmented Image
3.4. Fissure Region detection
Fissure regions are identified using the fissure sweeping technique and the fissures are enhanced
using region growing method. The fissure sweep technique finds the fissure regions in the pre-
processed CT images. The fissure sweep technique finds the fissure regions in the pre-processed
CT images. The general anatomic structure of the lung is shown in Figure1, in which the lobar
fissures separate the lung lobes where no major vascular or bronchial trees cross. This signifies
that in the pre-processed binary images, fissures are represented by large amounts of air or space
extending from the middle to the lateral side of the lungs. The fissure sweep technique uses this
knowledge to coarsely define void regions where fissures could be present. The fissure sweep

7
technique first calculates a set of predefined fissure angles. It is defined by angles between the
fissures and a flat horizontal line using knowledge from the general lung anatomy.
ALGORITHM
Input: Threshold image
Output: Fissure detected image
STEPS
1. Take the input image.
2. Draw a horizontal reference line and the fissure line
3. Keep the horizontal line is fixed and varies fissure line from the top to bottom which scans the
input image
4. Determine the void region by checking the continuous number of black pixel from the scanning
process.
5. The algorithm for fissure sweeping is implemented in the Figure 7 and the fissure swept result is
shown in the Figure 8.
Input CT image
Fissure Sweeping
Fissure Sweep Image
Figure 7. Fissure Sweeping
Figure 8.Fissure Swept Result

8
3.5. Fissure Refinement
Fissures are enhanced in the lung CT image using Region growing method, initially histogram
equalization is used to enhance the contrast of the fissure region, and then the Region growing
method is applied to refine the fissure. Fissures are generally situated between the vascular and the
bronchial trees of the two adjacent lung lobes. In Region growing method it finds the empty
regions between the vascular and bronchial trees of the two lung lobes. The centre of these empty
regions gives approximate fissure locations. In conventional region growing technique the seed
point goes in all directions whereas in this region growing technique the centre of the fissure
region is used as the seed point. This prevents searching the fissure from deviating too far from the
fissure angle. The seed grows in a direction outwards is perpendicular to the middle of the fissure
region and the output of fissure refined image is shown in Figure 9.
Input CT image
Region Growing Method
Fissure Refined Image
Figure 9.Fissure refined image
3.6. Lung Nodule Region Segmentation
Lung Nodules are segmented using adaptive thresholding. Mean of CT image is used as threshold.
The Lung Nodule Segmented Image is shown in Figure10.
Thresholding
Input CT image Nodule detected image
Figure 10.Nodule Segmented Image

9
4. RESULTS AND DISCUSSION
4.1. Fissure Verification
In order to test the segmentation algorithm, real time clinical CT images from twenty patients have
been processed and the experimental results are shown in Figure 11 and 12. The resolution of all
CT images is 512*512 pixels with a thickness of 5mm to 10mm. The results were obtained for
eight patients.
The performance of the segmentation algorithm is evaluated using a 1.6Ghz Intel(R) Core(TM)
2Duo CPU with 1GB of RAM running Matlab7.0 (R14). The Matlab is a Matrix laboratory and it
is a high-performance language for technical computing integrates computation, visualization, and
programming in an easy-to-use environment. Normally for surgeons it takes around 15 to 25
minutes to analyze the clinical CT images whereas the run time of segmentation algorithm is only
3 minutes and this indicates that the automatic segmentation is better suited for busy clinical
settings.
The identified fissure results are verified by comparing the manual and the automatic segmentation
and the accuracy is calculated statistically. Manual segmentation is carried out by the surgeons. In
Figure 11, fissures are located manually by surgeons and the fissure distances are identified from
two extreme boundaries. In Figure 12, oblique fissures are identified automatically by
segmentation algorithm using fissure sweeping technique and its accuracy is found and it is shown
in Table 2.
The automatic segmentation is better than manual segmentation because of the following reasons:
i. Automatic method is faster.
ii. Automatic method is easy to segment lung lobes with large number of images whereas it is
difficult for manual method to segment lung lobes as the number of images is larger.
iii. Automatic method takes lesser time, whereas the manual method is clumsy and subjective.
Input CT Image Manually segmented right and left oblique fissures
Figure11. Result for the Manual Segmentation

10
Input CT Image Automatically segmented right and left oblique fissures
Figure12. Result for the Automatic Segmentation
Patient # Manual
Segmentation
Ma
Automatic
Segmentation
Aa
Mean
Difference
Md =abs(Ma-Aa)
Accuracy%
(100* Md)/(Ma+Aa)
1 220 161 59 87
2 191 161 30 92
3 1654 3661 2007 72.58
4 1408 4263 2855 66.5
5 639 3274 2635 60
6 161 214 53 88
7 84 85 31 82
8 1270 4133 2863 65.3
Table 2.Accuracy for Segmenting the Oblique Fissure
In Table 2, Ma represents the total number of pixels from the edge to the centre of the fissure
region in Manual Segmentation, Aa represents the total number of pixels from the edge to the
centre of the fissure region in Automatic Segmentation, Md represents the Mean difference
between Manual and the Automatic Segmentation and finally accuracy is calculated using the
following formula:
Accuracy = (100*Md)/ (Ma+Aa)
5. CONCLUSION AND FUTURE WORK
Fissures cannot be seen by naked eyes because CT images are taken by low resolution scanners
and as the number of image increases, efforts taken by the medical experts to analyze the image
consumes much time. Because of this reason, there is a proposal for developing an automated
segmentation system to assist the medical experts, thereby reducing the consumption of time, so
the project has been developed.
In this project initially pre-processing of a CT image is done to remove the noises present in it.
Then the vessels are enhanced using the morphological operators followed by lung segmentation.
The fissure regions are identified, enhanced and verified. Lung nodules are segmented using

11
adaptive threshold. Through the project we have developed a segmentation algorithm for
identifying the fissures and nodules from CT images. Using the statistical analysis such as mean
difference between the automatic and manual segmentation, the results are verified. The results
indicate a potential for developing an automatic algorithm to segment lung lobes for surgical
planning of treating lung disease.
This project may be extended in the future, by focussing on the following aspects: Improving
segmentation of horizontal fissures, further evaluating the algorithm under severe pathological
cases, 3D visualization of the segmented lung lobes and nodules using volume rendering method
and fissure detection, verification is carried out with the increased number of images.
6. REFERENCES
[1] AubertG and KornprobstP.(2006),’Mathematical Problems in Image Processing:PartialDifferential
Equations and the Calculus of Variations (AppliedMathematical Sciences 147).’ New York:
Springer-Verlag.
[2] Aykac, D. Hoffman E. A, McLennan G. and. Reinhardt J.M, (2003) ‘Segmentationand analysis of
the human airway tree from three-dimensional X-rayCT images’, IEEE Trans. Med. Imag., vol. 22,
no. 8, pp. 940–950, Aug.
[3] Graham M. W, Gibbs J. D., and. William E. HRobust( 2008 ),’system forhuman airway-tree
segmentation’,Proc. SPIE (Med. Imag.), vol. 6914, pp. 69141J-1–69141J-18.
[4] Grady L. (2006), ‘Random walks for image segmentation’, IEEE Trans. Pattern Anal. Mach. Intell.,
vol. 28, no. 11, pp. 1768–1783.
[5] Haralick R. M,. Stenberg, S. R and Zhuang, X.(1987) ,’Image analysis using mathematical
morphology’,IEEE Trans. Pattern Anal. Mach. Intell.,vol. 9, no. 4, pp. 532–550.
[6] Hu S HoffmanE.A and. Reinhardt J.M, (2001), ‘Automatic lung segmentation for accurate
quantitation of volumetric X-ray CT images’, IEEE Trans.
[7] Hemminger B.M,. Molina, P. L. Egan T. M.Detterbeck, Muller, K. E.Coffrey, and Lee, J. K.
(2005), ‘Assessment of real-time 3D visualization for cardiothoracic diagnostic evaluation and
surgery planningDigit. Imag., vol. 18, pp. 145–153.
[8] Kass M,.Witkin A and.Terzopoulos D.(1987), ‘Snakes:Active contour models,’Int. J. Comput. Vis.,
vol. v1, pp. 321–331.
[9] Kubo M,. Niki N. Eguchi K,. Kaneko M, and. Kusumoto M. (2000) ‘Extraction of pulmonary
fissures from thin-section CT images using calculation of surface-curvatures and morphology
filters’,in Proc. IEEE Int. Conf. Imag.Process. pp. 637–640.
[10] Kubo M,. Kawata, Y. Niki, N Eguchi K., Ohmatsu H.,. Kakinuma R,Kaneko M.,. Kusumoto M,.
Moriyama N,. Mori K, and. Nishiyama H.(2001)‘Automatic extraction of pulmonary fissures from
multidetector-row CTimages’, in Proc. IEEE Int. Conf. Imag. Process. pp. 1091–1094.
[11] Kuhnigk J.-M. Hahn H. Hindennach M. Dicken V. Krass S, and Peitgen H.O.(2003) ‘Lung lobe
segmentation by anatomy-guided 3-D watershed transform’, Proc. SPIE (Med. Imag.), vol. 5032,
pp. 1482–1490,Med. Imag., vol. 20, no. 6, pp. 490–498.
[12] Kubo M.. Niki N, Nakagawa S., Eguchi K.,. Kaneko M, Moriyama N, Omatsu H, Kakinuma R and
Yamaguchi.N. (1999) ,’Extraction algorithm of Pulmonary fissures from thin-section CT images
based on linear feature Detector method’, IEEE Trans. Nucl. Sci., vol. 46, no. 6, pp. 2128–2133.
[13] Kuhnigk J.-M,. Dicken V,. Zidowitz S, Bornemann L.,. Kuemmerlen B,Krass, S Peitgen, H.-O.
Yuval, . S. Fend, H.-H Rau, W. S. and. Achenbach, T(2005),’Newtools for computer assistance in
thoracic CT. Part 1. Functional analysis of lungs, lung lobes, and bronchopulmonary segments’,
Radio-Graphics, vol. 25, pp. 525–536.
[14] Sonka, M. Hlavac V, and Boyle R.(, 2008.),’ Image processing, Analysis, and Machine Vision’, 3rd
ed. Toronto, ON: Thomson Learning.

12
[15] SluimerI.ProkopM and GinnekenB.V (2005), ‘Toward automated segmentation of the pathological
lung in CT’, (IEEE Trans. Med. Imag., vol. 24, no. 8, pp. 1025–1038.
[16] Tschirren J. Hoffman, E. A. McLennan G, and. Sonka M.(2005),’Intrathoracic airway trees:
Segmentation and airway morphology analysis from lowdose CT scans’,IEEE Trans. Med. Imag.,
vol. 24, no. 12, pp. 1529–1539.
[17] Ukil S. Sonka M, and Reinhardt F.C (2006) ‘Automatic segmentation of pulmonary fissures in X-
ray CT image using anatomic guidance’, Proc SPIE (Med. Imag.), vol. 6144, pp. 213–223.
[18] WangJ, BetkeM and. Ko, J. P. (2006).’Pulmonary fissure segmentation on CT’ Med. Imag. Anal.,
vol. 10, pp. 530–547.
[19] Wang J Betke M and. Ko J. P. (2002), ‘Segmentation of pulmonary fissures on diagnostic CT—
Preliminary experience’ in Proc. Int. Conf. Diagn. Imag.Anal, pp. 107–112.
[20] Wei Q Hu Y. Macgrego J. H r, and Gelfand G. (2008), ‘Segmentation of lung lobes in clinical CT
images’, Int. J. CARS, vol. 3, pp. 151–163.
[21] WuZ and. LeahyR (1999),’an optimal graphs theoretic approach to data clustering: Theory and its
application to image segmentation’, IEEE Trans.Pattern Anal. Mach. Intell., vol. 15, no. 11, pp.
1101–1113.
[22] Zhang .L. Hoffman E. A and. Reinhardt J. M. (2006) ‘Atlas-driven lung lobe segmentation in
volumetric X-ray CT images’, IEEE Trans. Med. Imag.,vol. 25, no. 1, pp. 1–16.
[23] Zhang, L. E. Hoffman A. and. Reinhardt J. M.(2001),’Lung lobe segmentation by graph search with
3D shape constraints’ Proc. SPIE (Med. Imag.),vol. 4321, pp. 204–215.
[24] Wei, Q., Hu, Y., Gelfand, G., & MacGregor, J.H.,“Segmentation of Lung Lobes in High
resolution Isotropic CT images” IEEE Trans. Med. Imag.,vol. 56, no. 5, pp. 1–11, may 2009.
Authors
[1] S. Anitha is pursuing her Post Graduation in Multimedia Technology,
Her Area of interest include Image Processing, Multimedia and Mobile
Networks.
[2] Dr. S. Sridhar is an Assistant Professor in Computer Science and
Engineering Department at College of Engineering, Guindy, Chennai.
His Research interest is in the area of Image Processing,
Pattern Recognition, Data Mining, Artificial Intelligence.

Segmentation of Lung Lobes and Nodules in CT Images

More Related Content

What's hot (20)

Similar to Segmentation of Lung Lobes and Nodules in CT Images (20)

Recently uploaded (20)

Segmentation of Lung Lobes and Nodules in CT Images