Detection of lung cancer nodules using automatic region ... - IEEE Xplore

IEEE - 31661

Detection of lung cancer nodules using automatic region growing method S.SHAIK PARVEEN

C.KAVITHA

Research Scholar Bharathiar University Coimbatore, India

Assistant Professor/Computer Science Thiruvalluvar Government Arts College Rasipuram, India

Abstract—Image Segmentation is an important part of image processing. It is used in medical field to detect and to diagnose the death threatening diseases. Manual readings can be done to analyze the medical images. But still the result leads to misdiagnosis by manual segmentation and the accuracy is not so high. Many Computer Aided Detection systems arise to increase the accuracy and performance rate. In the field of medical diagnosis, imaging techniques is presently available, such as radiography, computed tomography (CT) and magnetic resonance imaging (MRI). Medical image segmentation is more crucial part in analyzing the images. Although the conventional region growing algorithm yields in better result, it lacks with the concept of manual selection of seed points. A new approach is used to segment the images to identify the focal areas in lung nodules. Threat Points Identification is used with region growing method for segmenting the suspicious region. Experiment is carried out using real time images to investigate our method. Keywords—Region growing method, Lung nodule Detection, Image Segmentation, Medical Image Processing, Computer Tomography

I. INTRODUCTION Lung cancer is considered to be the main cause of cancer death worldwide, and it is difficult to detect in its early stages because symptoms appear only in the advanced stages causing the mortality rate to be the highest among all other types of cancer. More people die because of lung cancer than any other types of cancer such as breast, colon, and prostate cancers. There is significant evidence indicating that the early detection of lung cancer will decrease mortality rate. The most recent estimates according to the latest statistics provided by world health organization indicates that around 7.6 million deaths worldwide each year because of this type of cancer. Furthermore, mortality from cancer are expected to continue rising, to become around 17 million worldwide in 2030. Early detection of lung cancer is valuable. The 5-yearsurvival- rate of lung cancer has stagnated in the last 30 years and is now at approximately just 15%. Lung cancer takes more victims than breast cancer, prostate cancer and colon cancer together. This is due to the asymptomatic growth of this cancer. In the majority of cases it is too late for a successful

therapy if the patient develops first symptoms (e.g. chronic croakiness or hemoptysis). But if the lung cancer is detected early (mostly by chance), there is a survival rate at 47% according to the American Cancer Society. The goal of segmentation is to simplify and/or change the representation of an image into something that is more meaningful and easier to analyze. 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. Literature has a wide range of techniques. In which some are histogram based technique, edge based techniques, region based techniques, hybrid methods which combine both the edge based and region based methods together, and goes on [1]. Now-a-days image processing has wide scope in medical image processing for diagnosing the dangerous diseases. Some applications of image segmentation are machine vision, medical imaging, object detection, recognition tasks, and video surveillance [2]. In the medical field, many imaging modalities are in use such as radiography, computed tomography (CT) and magnetic resonance imaging (MRI) [1, 2]. Recently Computer Tomography (CT) is commonly used to identify the lung diseases like lung cancer, tuberculosis (TB) and chronic obstructive pulmonary disease (COPD). In order to assist the radiologists many computer aided detection (CAD) arises to increase the accuracy and performance rate. Still then the detection rate is not high. Region growing is region-based image segmentation method. This method starts with a single pixel, referred to as seed pixel and the neighboring pixels are added to it based on similarity properties like intensities, model, shape and texture, forming a region which is to be segmented in the image. Selection of seed pixel depends mainly on the problem domain. To overcome this problem all the objects in an image are detected without any manual specification of the seed pixel. The remainder of the paper is organized as follows. Section 2 provides an overview on related research works in medical image segmentation.

4th ICCCNT - 2013 July 4-6, 2013, Tiruchengode, India

IEEE - 31661

Section 3 explains the methodology of the proposed system. Section 4 discusses on experimental results for real images. Section 5 concludes the paper with fewer discussions. II.

LITERATURE SURVEY

Implementation of computer aided detection contains various fields such as enhancing the CT images, identifying suspected region, feature extraction from segmented image, classifying the tumors and so on. Many algorithms have been proposed to improve the efficiency of the CAD system in the above mentioned fields. Some of those methods are discussed in this section. Kenji Suzuki, Hiroyuki Abe, Heber MacMahon, and Kunio Doi in [3] introduced a different version in image processing technique using Massive Training Artificial Neural Networks (MTANN). Their technique gives a solution to the problem faced by radiologists as well as computeraided diagnostic (CAD) systems to detect these nodules when the lung nodules overlaps with the ribs or clavicles in chest radiographs. An MTANN is extremely a non-linear filter that can be trained by use of input chest radiographs and the equivalent “teaching” images. They used a linear-output backpropagation (BP) algorithm that was derived for the linear-output multilayer ANN model in order to train the MTANN. The dual-energy subtraction is a technique used in [3] for separating bones from soft tissues in chest radiographs by using the energy dependence of the x-ray attenuation by different materials. Using fuzzy rules [4] proposed a Templatematching technique based on genetic algorithms (GA) template matching (GATM) for detection of nodules existing within the lung area. In their work, GA was used to determine the target position in the observed image efficiently and to select an adequate template image from several reference patterns for quick template matching. [5] proposed a three step segmentation process for the analysis of lung image. In their approach, if the area in the CT image occupied by GGO is large, then it is easy for a medical doctor to extract the features. However, the possibility to overlook the light gray shadow becomes higher when GGO exists as a small area. In the first step of their model, extracting ROI is performed to segment the lung area. To achieve better segmentation accuracy, preprocessing the CT slices is carried out by employing binarization, labeling, shrinking and expansion. In the second step, calculation of characteristics of GGO shadows suchas mean value,

standard deviation, and semi interquartile range have been carried out. In the final step, the GGO shadow’s regions were extracted by linear discriminant function. Suspicious shadows are extracted by Variable N-Quoit (VNQ) filter from GGO. The suspicious shadows are classified into a certain number of classes using feature values calculated from the suspicious shadows. Kanazawa, K., Kubo, M. and Niki, N., [6] explored Computer aided diagnosis system for lung cancer based on helical CT images. In this technique the time complexity is reduced and the diagnosis confidence is raised. This method consists of two stages such as an analysis stage a diagnosis stage. In the analysis stage, the lung and pulmonary blood vessel regions are preprocessed extracted which in turn examine the features of these regions with the help of methods present in image processing. In the diagnosis stage, diagnosis rules are generated according these features, and identify the tumor regions using these diagnosis rules. Penedo, M.G., Carreira, M.J., Mosquera, A. and Cabello.D., [7] described a computer-aided diagnosis scheme which has two-level artificial neural network (ANN) architecture. In first level artificial neural network identifies suspicious regions in a low-resolution image. The curvature peaks calculated for all pixels in each suspicious region is given as the input to the second artificial neural network . The small size tumors are identified by the signature in curvature-peak feature space, where curvature is the local curvature of the image data when sighted as a relief map. The result gives a true positive identification in this network is threshold at a particular level of importance. Okada K, Comaniciu D, Krishnan [8] proposed a multiscale joint segmentation and model fitting solution which extends the robust mean shift-based analysis to the linear scale-space theory. In their paper, an ellipsoidal (anisotropic) geometrical structure of pulmonary nodules in the multislice Xray computed tomography (CT) images was used for target’s center location, ellipsoidal boundary approximation, volume, maximum/average diameters. Regions of interest (ROI) slices were combined to form 3D ROI image and a 3D template by [9]. They used this template to find the structures with similar properties of nodules. In another work by Ozekes S, Osman O, Ucan ON [10] proposed a methodology for segmenting the CT images of lungs using Genetic Cellular Neural Networks (GCNN). In their work, ROIs were specified with


IEEE - 31661

using the 8 directional search and +1 or -1 values were assigned to each voxel. Yang Song, Weidong Cai, Jinman KimDavid Dagan Feng [11] presented a new method to automatically detect both tumors and abnormal lymph nodes based on the low-level intensity and neighborhood features and high-level contrast-type features, with a two-level SVM classification. One level of conditional random field (CRF) is based on unary level contextual and spatial features and pair wise-level spatial features. Other level is based by relabeling the detected tumors as positive or mediastinum by filtering the high-uptake myocardium areas. Xujiong Ye, Xinyu Lin, Jamshid Dehmeshki, Greg Slabaugh, and Gareth Beddoe in [12] presented a new computer tomography (CT) lung nodule computer-aided detection (CAD) method. The method can be implemented for detecting both solid nodules and ground-glass opacity (GGO) nodules. Foremost step of the method is to segment the lung region from the CT data using a fuzzy thresholding technique. The next step is the calculation of the volumetric shape index map and the ―dot‖ map. The former mentioned map is based on local Gaussian and mean curvatures, and the later one is based on the Eigen values of a Hessian matrix. They are calculated for each Voxel within the lungs to enhance objects of a specific shape with high spherical elements. The combination of the shape index and ―dot features provides a good structure descriptor for the initial nodule candidate generation. Certain advantages like high detection rate, fast computation, and applicability to different imaging conditions and nodule types make the method more reliable for clinical applications. By analyzing [13] we proposed automatic region growing method for segmentation to improve the segmentation result. Therefore, it is necessary to propose a new and efficient technique to enhance the accuracy of segmentation. III.

METHODOLOGY

A. Data collection The data used here was got from reputed hospital. It contains CT images of 11 patients which has 3000 images for experiment. B. Preprocessing Computer Tomography (CT) images are difficult to interpret. Preprocessing is essential to

improve the quality of CT images. The image processing methods used for this phase are PSNR calculation, Erosion, Median Filter, Dilation, Outlining, Lung Border Extraction, Flood Fill Algorithm and finally extraction of lung parenchyma. The preprocessing methods are shown in figure 1. Mainly, the CT chest image contains the lung region, background, heart, liver and other organs areas. The main aim of this preprocessing is to detect the lung region and regions of interest (ROIs) from the CT scan image. The first step is application of PSNR (Peak Signal-to-Noise Ratio) calculation to the CT scan images. Considering one patient CT images, it consists of more number of slices. It is important to retrieve one best from all. PSNR is most commonly used to measure of quality. Although a higher PSNR generally indicates that the reconstruction is of higher quality. The best suitable with better accuracy is chosen for the further enhancement of lung region. Original CT Lung Image

Median Filter

Outlining

Lung Border Extraction

Flood Fill Algorithm

Extraction of Lung parenchyma Fig. 1. Methods involved in preprocessing extraction from CT Scan Lung Image

for

lung

Digitization noise and high frequency components are some of the unwanted regions present in the CT scan images. These can be removed by using median filter. The advantage of using median filter is, it can remove the noise without disturbing the edges. By applying the lung border extraction technique the lung region border is then obtained. Finally, to fill the obtained lung border with the lung region flood fill algorithm is used. The lung region is extracted from the CT scan


IEEE - 31661

image after the application of these algorithms. Finally the resultant lung region is used further for the segmentation in order to detect the cancer nodule. Figure 2 shows the preprocessing techniques used in extracting the lung region from computer tomography (CT) scan images from a through e where figure 2(e) shows the extracted lung region.

(a)

Step 1: Computation of Histogram and accumulated histogram are done to get Hlc and AHlc Step 2: Location of peaks in Hlc is found by using histogram gradient changes. (HP1, HP2…..HPi ) where HPj are the gray levels. Step 3: Threat threshold’s candidates are selected by following condition, Tt = { HPj | When the selected threat area ≤ 9% of the entire region of interest, where j=1,2,…,i}, S=r,r+1…z. AHlc can be used to calculate the selected threat area. Step 4: Threat threshold should be among of {Ts; S=r~z} i.e Tt= Tl r ≤ l ≤ z, such that |HPi - HPi1 | is maximum among {| HPk - HPk-1 | ;S=r ~z}.

(b)

Step 5: Pixel is marked at (u, v) as a candidate of threat pixel if p (u, v) > Tt by using Ti(u,v) = 4-y , where y=1,2,3

(c)

Step 6: A pixel at (u, v) is considered as threat pixel Ti (u, v) > 5 if ∑

(d)

2)

(e) Fig. 2. Preprocessing methods for extraction of lung region: a. Original Image b. Filtered Image (Median filter) c. Border Extraction d. Flood fill technique e. Lung region Extraction

C. Segmentation The segmentation is performed for determining the cancer nodules in the lung. This phase will identify the Region of Interest (ROI) which helps in determining the cancer region. Region growing method is used for segmentation. Selection of seed pixel depends mainly on the problem domain. To overcome this problem all the objects in an image are detected without any manual specification of the seed pixel. R. Bellotti, F. De Carlo [14] proposed a CAD system using region growing and active contour model with detection rate of 88.5% which is still has to be improved.

1) Threat Pixel Identification Threat pixels are produced by thresholding the preprocessed image. Threat pixel threshold is determined by histogram analysis. Segmentation through threat pixel identification contains the following steps,

Region growing method

Region growing method seeks group of pixels with uniform intensities. Seeded region growing performs a segmentation of an image with respect to set of points known as seed. Threat pixel generated from the above process can be considered as seed point. Given the seed the region growing method finds the tessellation of the image into regions with property that each connected component of region meets exactly one of Si. Each step of algorithm involves the addition of one pixel into above set. Let z be the unallocated pixel. α= {x ɇ Ui=1 n Si |n(x)∩ Ui=1 n Ai ≠Φ}

(1)

Where, n(x) is set of immediate neighbours of the pixel neighbours of 8 connected pixel x. if x Є α then n(x) meets just one of Ai. Hence i(x) Є{1, 2,…..n} to be the index such that n(X) ∩ Si (X) ≠Φ.

(2)

δ(X) =|g(X) - mean g Є Si (X) [g(y)]|

(3)

Where δ(X) is measure of how different x is from the region it joins and g(x) is the gray value of the pixel x. If N(x) meets two or more values of Ai then Ai will be selected according to the lowest value δ(X)=min x Є T { δ(X)}


(4)

IEEE - 31661

The above process is repeated till all the pixels have been identified.

extend our gratitude to all the reviewers for their critical comments.

D. EXPERIMENTAL RESULTS

References

The experiments are conducted on the computer-aided diagnosis systems with the help of real time lung images. This data set consists of 11 patients which has 3000 slices (one normal case, 2 benign cases and 9 malignant cases) from which best slice is used for further processing. The original CT image is been preprocessed by using different techniques of image processing and it is been segmented using threat pixel identification and region growing method. The detection of lung nodule of a CT scan lung image is given in figure 3.

[1]

(a)

[2]

[3]

[4]

(b)

[5] (c)

(d)

Fig. 3 (a) represents the CT scan image of malignant 3(b) represents the malignant cancer nodule detected image after segmentation, 3(c) represents the CT scan image of benign and 3(d) represents the benign nodule detected image after segmentation

[6]

E. CONCLUSION A new computer aided detection technique has been designed to detect the suspicious region automatically without any human interruption. Preprocessed image is segmented using threat pixel identification process in combination with region growing method. The segmented region from the proposed work can be used for further processing such as feature extraction and classification in future. Hence the proposed method is highly desirable in order to assist the radiologist in the detection of lung nodules and to increase the diagnostic accuracy.

[7]

[8]

ACKNOWLEDGMENT Foremost, we would like to thank our almighty in the success of completing this work. We would

[9]

K. Haris, “Hybrid Image Segmentation using Watersheds and Fast regionMerging”, IEEE Transactions on Image Processing, vol. 7, no. 12, pp. 1684-1699, 1998. W. M. Wells, W. E. LGrimson, R. Kikinis and S. R. Arrdrige, “Adaptive segmentation of MRI data,” IEEE Transactions on Medical Imaging, vol. 15, pp. 429-442, 1996. Kenji Suzuki, Hiroyuki Abe, Heber MacMahon, and Kunio Doi, “ImageProcessing Technique for Suppressing Ribs in Chest Radiographs by Means of Massive Training Artificial Neural Network (MTANN),” IEEE Transactions on medical imaging, vol. 25, no. 4, pp. 406-416, 2006. Lee Y, Hara T, Fujita H, Itoh S, Ishigaki T, “Automated detection of pulmonary nodules in helical CT images based on an improved template-matching technique”, IEEE Trans. Med. Imaging, Vol. 20, pp.595–604, 2001. Hyoungseop Kim, Seiji Mori, Yoshinori Itai, Seiji Ishikawa, Akiyoshi Yamamoto and Katsumi Nakamura,” Automatic Detection of Ground-Glass Opacity Shadows by Three Characteristics on MDCT Images”, World congress on medical physics and biomedical engineering, IFMBE Pro2, 2007. Kanazawa, K., Kubo, M. and Niki, N., “Computer Aided Diagnosis System for Lung Cancer Based on Helical CT Images”, International Conference on Pattern Recognition, Vol. 3, Pp. 381 - 385, 1996. Penedo, M.G., Carreira, M.J., Mosquera, A. and Cabello,D., “Computer-Aided Diagnosis: A Neural-Network-Based Approach to Lung Nodule Detection”, IEEE Transactions on Medical Imaging, Pp: 872 – 880, 1998. Okada K, Comaniciu D, Krishnan, “A Robust Anisotropic Gaussian Fitting for Volumetric Characterization of Pulmonary Nodules in Multislice CT”, IEEE Trans. Med. Imaging, Vol. 24, No. 3, pp. 409– 423, 2005. Osman O, Ozekes S, Ucan ON, “Lung nodule diagnosis using 3D template


IEEE - 31661

[10]

[11]

[12]

[13]

[14]

matching”, Comput. Biol. Med, Vol. 37, Issue 8, pp. 1167–1172, 2007. Ozekes S, Osman O, Ucan ON,” Nodule detection in a lung region that’s segmented with using genetic cellular neural networks and 3D template matching with fuzzy rule based thresholding”, Korean J. Radiol., Vol. 9, pp.1–9 , 2008. Yang Song, Weidong Cai, Jinman KimDavid Dagan Feng,” A Multistage Discriminative Model for Tumor and Lymph Node Detection in Thoracic Images”, IEEE transactions on Medical Imaging, vol. 31, no. 5, May 2012 Xujiong Ye, Xinyu Lin, Jamshid Dehmeshki, Greg Slabaugh, and Gareth Beddoe, ―Shape-Based Computer-Aided Detection of Lung Nodules in Thoracic CT Images,” IEEE Transactions on Biomedical Engineering, vol. 56, no. 7, pp. 1810-1820, 2009. S.Shaik Parveen, Dr.C.Kavitha,” A Review on Computer Aided Detection and Diagnosis of lung cancer nodules,” International Journal of Computers & Technology, Volume 3 No. 3, Nov-Dec, 2012 R. Bellotti, F. De Carlo,” A CAD system for nodule detection in low-dose lung CTs based on region growing and a new active contour model”, Med. Phys. Vol. 34,pp. 4901-4910, December 2007
