Segmentation of Lungs via Hybridization of CA and Level Set Algorithm

In this study, we present a fast and robust practical tool for segmentation of solid tumors with minimal user interaction. The lung image is segmented manually by the experts may includes risk and time consuming process. Hence, in this study, the lung image is segmented in order to identify the tumor sector. Initially, the input lung image is applied with the denoising process for removing noises with the aid of multi-wavelet transformation. After this process, the CA algorithm is applied for obtaining the background seeds and foreground seeds (tumor seeds) and then the level set algorithm is applied for acquiring the acute tumor tissues. As a result of the mentioned process, the tumor sector is segmented and the results are depicted. Studies on lung tumor datasets demonstrate 8085% overlap performance of the proposed algorithm with less sensitivity to seed initialization, robustness with respect to heterogeneous tumor types and its efficiency in terms of computation time.


INTRODUCTION
The medical imaging is a gifted technique to be aware of the physical condition of different organs of human body.There are numerous techniques available such as digital radiography (X-ray), Computerized Tomography (CT) and Magnetic Resonance Imaging (MRI), which are being applied on patients 'bodies based on different diseases (Hasan et al., 2012).Analysis of medical images is frequently intricate and time overwhelming, even for skilled physicians.The support of image analysis and machine learning can make this process easier (Agarwal et al., 2010).Computer aided radiological diagnosis frequently involves an analysis of individual anatomical structures, location of pathological regions and extraction of various features describing texture and size.Image segmentation, playing a crucial role in plentiful applications, relies on the extraction of homogeneous and non-overlapping regions (Wieclawe and Pietka, 2012).Image segmentation is the task of splitting a digital image into one or more regions of interest (Mostofi, 2009).Segmentation is defined as partitioning portions of an image.It adds structure to a raw image (Dhole et al., 2013).
Lung disease is one of the foremost causes of death.Some evidence suggests that early detection of lung disease may allow for timely therapeutic intervention which in turn results in a more favorable prognosis of the patients (Krishnan and Robinson, 2013).Numerous computer-aided lung image analysis methods require the segmentation of lung tissue in an initial processing step (Sun et al., 2010).Lung diseases are a few of the most widespread medical conditions worldwide.Tens of millions of people experience from lung disease worldwide.Smoking, infections and genetics are in charge for the majority lung diseases (Mesanovic et al., 2011).Lung cancer is one of the most serious cancers, with the smallest survival rate after the diagnosis, with a gradual increase in the number of deaths every year.Survival from lung cancer is directly related to its growth at its detection time.The earlier the detection is, the higher the chances of successful treatment are.An estimated 85% of lung Cancer cases in males and 75% in females are caused by cigarette smoking (Sharma and Jindal, 2011).For providing solution to this issue, the prognosis of lung cancer is a vital one.In order to make prior diagnosis, the segmentation of lung tumor is a crucial process.
As discussed earlier, in order to identify the lung tumor, segmentation is the vital step and various researches have done in this field.A few of them are listed below.In Hashemi et al. (2013) have meant that at detecting lung nodules automatically throughout Computerized Tomography (CT) image.For that reason, they have offered a technique to develop the efficiency of the lung cancer diagnosis system, through a region growing segmentation method to segment CT scan lung images.Afterwards, cancer recognition was presented by Fuzzy Inference System (FIS) for differentiating amid malignant, benign and advanced lung nodules.They have examined the diagnostic performances of FIS system through Artificial Neural Networks (ANNs).In Sasidhar et al. (2013) have concerned two steps a. Automated Segmentation of lung regions b.Automated Detection of lung cancer.For speeding up the process of detecting lung cancer, segmentation of lung region plays an important role.The proposed algorithm has assisted the radiologist for the better diagnosis of lung cancer.An Automated approach has been exploited to spot lung cancer.Sharma and Jindal (2011) have presented a system for early recognition of lung cancer by analyzing lung CT images using several steps.The approach had been started through extracting the lung regions from the CT image using several image processing techniques, including bit image slicing, erosion, weiner filter.They have established the using of bit plane slicing technique instead of the thresholding technique that was utilized in the initial step in the extraction process to convert the CT image into a binary image.Bit-plane slicing technique was both faster and data-and userindependent compared to the thresholding technique.After the extraction step, the extracted lung regions were segmented using region growing segmentation algorithm.Then, the initial lung candidate nodules resulting from the Region growing segmentation were analyzed to extract a set of features to be used in the diagnostic rules.These rules were utilized to detect cancerous region and also have obtained the result.In Siddique et al. (2014), have presented a method to automatically detect both tumors and abnormal lymph nodes simultaneously in Computerized Tomography (CT) thoracic images.They have performed the detection in a multistage approach, by first detecting all potential abnormalities and then they have differentiated between tumors and lymph nodes.Then lung tumors were categorized as benign and malignant using ANFIS by measuring the parameters like precision, recall and sensitivity.Wide range of researches is carried out in this field and in our study, the input image is selected with initial seeds randomly and then these seeds are applied with the CA algorithm for the computation background seeds and foreground seeds (tumor seeds).Subsequent to this process, level set algorithm is applied in order to obtain the acute tumor.

MATERIALS AND METHODS
The impact of Digital Image Processing (DIP) is on various fields such as medical, surveillance, engineering applications, astronomy and more.Besides other, in the medical field the exploitation of digital image is bowled over.The digital images are utilized as the initial step of analyzing the diseases that supports the medical practitioners for providing treatment to the patients.In medical field, the digital image plays a crucial role by diagnosing and identifying the stages of diseases in order to heal them.In the proposed method, the digital image is utilized to segmenting the lungs region in order to separate the affected regions.Segmentation is defined as partitioning portions of an image.It adds structure to a raw image (Chen, 2008).In order to detect the pathological abnormalities presented in the lungs, the segmenting its anatomical parts is the vital step.
In the proposed methodology, the hybridization of CA and the level set algorithm is utilized.The introduction of both the methods is detailed in the section.Initially, the digital image of lungs is obtained through the modern acquisition techniques.It may have noises either because of the devices or the acquisition techniques and hence the digital images are applied with the denoising technique.After this process, the denoised digital lung image is applied with the CA algorithm for obtaining the clear-cut view of the background and foreground particles.Then the tumor strength is computed and after that to obtain the acute tumor tissue, the level set algorithm is applied in it and hence the segmented regions are obtained.Some of the researches are available in this lungs segmentation for examining the tumors, rather than the other researches the proposed method obtains the required output without the aid of manual perception.
Let Im be the digital image of lungs with the size of M×N and may acquired through the modern acquisition techniques, Im mn where 0≤m≤M-1 and 0≤n≤N-1.This image Im is utilized for segmenting the tumor region and this is further exploited for the classification and hence the image must be clear without the presence of noises.For removing the noises in the image, the multi-wavelet transformation (Mohideen et al., 2011) is applied and the steps are below: Step 1: The image Im is decomposed via multiwavelet transformation (Velayudham and Kanthavel, 2012) and the decomposing level is δ.
Step 2: The energy distribution of every small sub bands are computed statistically.
Step 3: Initial threshold is computed.
Step 4: The threshold of each sub band is fixed and the wavelet coefficient of every level is computed.
Step 5: By performing the inverse multi-wavelet transform through the above process repeated using the high and low frequency coefficients and finally the denoised image I g is obtained.
Then the denoised image I g is applied with the CA algorithm in order to obtain the background and the foreground regions.For this process, let Gr be the graph that contains vertices vr and edges ed, a weight wt ij is assigned to each edge between the vertices.
Here the image I g pixels are the vertices and edges are the similarity measure (i.e., color/gray level intensity values) between the neighboring pixels.Each vertex vr has an attribute z that is an indicator of the probability of a label, here label is the background and the foreground label and the initial seeds are chosen randomly.The randomly selected seeds are compared with the reference values (an array of tumor seed's value).These initial seeds are selected and then the foreground and background seeds are computed based on the reference value.Then the foreground and background seeds are applied with the CA algorithm through the region around the selected seeds.Once the region is selected then the CA algorithm is running twice on this region in order to identify the tumor presented in the lungs.CA algorithm has its own transition rules for spotting the tumor area.According to these transition rules, the pixels changes its state synchronously that depending on their neighbor pixels.After completing this process, the 2 sets of strength maps are computed and they are background and foreground strength maps.Once the strength maps are combined then the probability map is determined.Initially the seed value is initialized and then the previous state is copied to a new variable.Once the state has been changed then the process is applied.The following algorithm shows the aforementioned process.After that the equation for combining strength map is showed in ( 1

End for End do End for
After the CA algorithm, the obtained strength maps (i.e., foreground and background seed's strength maps) are combined as shown in Fig. 1 and 2d in order to obtain the probability.The probability map is computed because the tumor affected seed is proportional to its distance normalized to the neighboring background seed.Hence this prefers the tumor seeds that are having higher probability and its distance to the background seeds is high and vice versa: Subsequent to this process, the obtained surface is smoothed with the intention of obtaining acute tumor volume.For the smoothing process the level set algorithm is applied.This algorithm defines acute shape and for smoothing process.In general let Γ (v, t) be the level set function, v be the position vector and t be the time of the level set evolution and initially the evolving surface is set to zero value that is Γ (v, t) = 0.The segmentation is obtained once the RMS difference between Γ (v, t) = 0 and Γ (v, t -1) = 0 is less than a predefined minimum RMS value.
For this aforesaid process, the obtained surface is initialized which is considered both the inside and outside region of the surface sf.The vital idea of the level set surface is to represent the evolving contour by means of a signed function and its zero level is corresponded to the actual contour or outline of the tumor surface.The level set embedding is done over the tumor surface via the following equation: Here c and d are the inside and outside surface's means and N is the unit normal vector to surface sf.The level set function's zero level set indicates an initial estimation of the tumor surface sf and its developed on P Tr with a piecewise constant region, though by exploiting a local Gaussian kernel to describe the inner and outer regions around the growing surface for estimating the regional statistics of the map.Once the surface's growth converges, then the final segmented tumor is attained.

RESULTS AND DISCUSSION
The proposed lung segmentation process was implemented with the aid of the working platform of MATLAB ((version 7.12).The proposed system was examined by means of the publicly available lung images Fig. 1 and 2a.With the selected pixels as initial seeds, the CA algorithm was applied on the image for obtaining the foreground pixels i.e., here tumor pixels and background pixels as shown in Fig. 1 and 2b and c.In this algorithm, the threshold value (0.5) was utilized for the CA algorithm.The obtained output segments are given as input to the level set algorithm for extracting the apt tumor segment as shown in Fig. 1 and 2e.The stepwise output of the input image is depicted below.

Fig. 1 :
Fig. 1: (a) Input lung image, (b) foreground strength of the input image, (c) background seed strength, (d) combined strength of background and foreground strength, (e) segmentation using CA (red), segmentation using level set (yellow), (f) segmented lung tumor Fig. 2: (a) Input lung image 2, (b) foreground strength of input image 2, (c) background strength of image 2, (d) combined strength of image 2, (e) segmented tumor of image 2 segmentation using CA (red), segmentation using level set (yellow), (f) segmented tumor segment of image 2

Fig. 3 :
Fig. 3: Chart of epsilon parameter of CA algorithm

Fig. 6 :
Fig. 6: Chart of spatial overlap parameter of level set algorithm

Table 1 :
Parameters of CA algorithm and level set algorithm