Lung Cancer Detection and Classification UsingANN & Multinomial Bayesian Classifier

Abstract— Medical image enhancement & classification play an important role in medical research area. In this paper, we have tried to analyze the texture of Computed Tomography (CT) images taken of abdomen and to find the values of various parameters of texture. Mainly CT lung images are classified into normal and abnormal category. Classification of images depends on the features extracted from the images. In this paper we are focusing on texture based features such as GLCM (Gray Level Co-occurrence Matrix) feature plays an important role in medical image analysis. Totally 12 different statistical features are extracted. To select the required features among them we use sequential forward selection algorithm. Afterwards we prefer multinomial multivariate bayesian for the classification stage which gives perfect classification as compare to Fuzzy C-Mean (FCM) clustering algorithm,. Classifier performance will be analyses further. Then, the abnormal images are projected to segmentation to view the affected portion. The effectiveness of the modified weighted FCM algorithm in terms of computational rate is improved by modifying the cluster center and membership value updating criterion.

Index Terms— Histogram Equalization, feature extraction, sequential forward selection, Image segmentation, fuzzy c-means algorithm.

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 at 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 the 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[1].The early detection of lung cancer in its primary stage is a challenging problem, due to the complicated structure of the cancer cells, where most of the cells are overlapped to each other. It is a computational procedure that sort images into groups according to their similarities. In this paper we used Histogram Equalization for preprocessing of the images and feature extraction process and multivariate multinomial Bayesian classifier to check the state of a patient in its early stage whether it is normal or abnormal. The manual analysis of the sputum samples is a very time consuming, inaccurate and requires well trained person to avoid diagnostic errors. The quantitative procedure is very helpful for earlier detection of lung cancer. Experimental analysis is made with dataset to evaluate the performance of the different classifiers. The

performance is based on the correct and incorrect classification of the classifier.

The reminder of this paper is organized as follows.In section III we have to preprocess the given test image for reducing noise and to enhance the contrast by using Contrast Limited Adaptive Histogram Equalization (CLAHE). Afterwards, in section IV texture features (GLCM) will be extracted from the test image. In feature extraction stage, statistical measurements are calculated from the gray level co-occurrence matrix for different directions and distances. Among the various features extracted we have to select the distinct features that will be utilized for classification purpose. For the selection of features SFS (Sequential Forward Selection) is used. In section V Kernelised bayesian is used to classify whether the test image comes under normal or abnormal.

A. System Architecture

The images obtained from CT scan machine were in DICOM format. Because the images were in 3D format, the commands of MATLAB could not be used on them. Therefore, to work on the images, they have been converted to BMP format, so that they can easily be used in the MATLAB environment. The syntax used for this purpose is:

I1=rgb2gray (I); (1)

Where I is a CT scan image, Rgb2gray converts RGB images to grayscale by eliminating the hue and saturation information while retaining the luminance. The grayscale image is of 2D type. The images were stored in the workspace of the MATLAB software from where they can be processed.

Training Set

Preprocessing

(CLAHE)

Test image

Preprocessing

(CLAHE)

.

Figure1. System architecture

Into the system architecture flow of the work is given into the right side of that architecture one test image is taken after that preprocessing has been done by using The Contrast Limited Adaptive Histogram Equalization (CLAHE) then texture fea-tures (GLCM) will be extracted from the test image, from the feature vector feature subset is obtained and then processed to-wards the Multivariate multinomial Bayesian classifier for classification of test image whether it is normal one or abnor-mal. Same sequence flow is used for the training set images into the left side of the architecture.

II. PRE-PROCESSING OF CT IMAGE FOR TO ENHANCE THE CONTRAST

The Contrast Limited Adaptive Histogram Equalization (CLAHE) is an improved version of adaptive histogram equal-ization. The contrast limited adaptive histogram equalization algorithm partitions the images into contextual regions and ap-plies the histogram equalization to each one. This evens out the distribution of used gray values and thus makes hidden features of the image more visible. The amount of contrast en-hancement for some intensity is directly proportional to the slope of the Cumulative Distribution Function (CDF) at that intensity level. Hence contrast enhancement can be limited by limiting the slope of the CDF. The slope of CDF at a bin loca-tion is determined by the height of the histogram for that bin.

Therefore the height limitation of the histogram results in lim-iting the slope of the CDF and hence the amount of contrast enhancement

Algorithm 1

Following is the overview of the algorithm for this function:1. Calculate a grid size based on the maximum dimension of the image. The minimum grid size is 32 pixels square.2. If a window size is not specified chose the grid size as the default window size.3. Identify grid points on the image, starting from top-left corner. Each grid point is separated by grid size pixels.4. for each grid point calculate the histogram of the region around it, having area equal to window size and centered at the grid point.5. If a clipping level is specified clip the histogram computed above to that level and then uses the new histogram to calculate the CDF.6. after calculating the mappings for each grid point, repeat steps 6 to 8 for each pixel in the input image.7. For each pixel find the four closest neighboring grid points that surround that pixel.8. Using the intensity value of the pixel as an index, find its mapping at the four grid points based on their cdfs.9. Interpolate among these values to get the mapping at the current pixel location. Map this intensity to the range [min:max) and put it in the output image.

III. FEATURE EXTRACTION AND SELECTION

A. Extraction of Texture FeatureAmong the four approaches (Structural, Statistical, model

based and Transform) here we are using Statistical approach to texture analysis The statistical approach does not presume in term of primitive but it draws on the general set of statisti-cal tool. It is the most widely used and more generally applied method because of its high accuracy and less computation time.

A gray level co-occurrence matrix (GLCM) contains infor-mation about the positions of pixels having similar gray level values. The GLCM, C, is defined with respect to given (row, column) displacement h. And element (i , j), denoted cij, is the number of times a point having gray level j occurs in position h relative to a point having gray level i. Let Nh be the total number of pairs, then C ij = cij /Nh is the elements of the nor-malized GLCM, C. The probability measure can be defined as:

Pr ( x )={Cij|δ ,ɵ|}(2) Where Cij (the co-occurrence probability between grey levels i and j) is defined as:

MVMNBayesian

Training

Feature Subset Matrix

FeatureSelection(SFS)

FeatureExtraction(GLCM)

Kernel-Structure

MVMNBayesian

Testing

Feature Subset

Classification Result

Feature Vector

Abnormal

Normal

FeatureExtractio

n(GLCM)

C ij=

Pij

∑i , j=1

G

Pij

(3)

Where Pij represents the number of occurrences of grey lev-els i and j within the given window, given a certain (δ-inter pixel distance, θ-orientation) pair. G is the quantized number of grey levels. The result of a texture calculation is a single number representing the entire window. This number is put in the place of the centre pixel of the window, then the window is moved one pixel and the process is repeated of calculating a new GLCM and a new texture measure. In this way an entire image is built up of texture values.From the co-occurrence matrix obtained, we have to extract the 12 different statistical features.

1. ContrastContrast is a measure of the local variations present

in an image. C ( k ,n )=∑

i∑

j(i− j)k Pd [ i , j ]n(4)

If there is a large amount of variation in an image the P [i, j]’s will be concentrated away from the main diagonal and con-trast will be high (typically k=2, n=1).

2. Sum of Squares, Variance

VARIANCE=∑I=0

G−1

∑j=0

G−1

(i−µ )2 P (i , j )(5)

This feature puts relatively high weights on the elements that differ from the average value of P(i, j).

3. CorrelationCorrelation is a measure of image linearity

C c=∑

i∑

j[Cij Pd [i , j ] ]−µi µ j

σ i σj(6)

Whereµi=∑ i Pd[ i , j] σ i

2=∑ i2 Pd [ i . j ]−µi2

4. EnergyOne approach to generating texture features is to use

local kernels to detect various types of texture. After the con-volution with the specified kernel, the texture energy measure (TEM) is computed by summing the absolute values in a local neighborhood:

Le=∑1

M

∑1

N

¿C (i , j )∨¿(7)¿

If n kernels are applied, the result is an n-dimensional feature vector at each pixel of the image being analyzed. 5. Maximum Probability

This is simply the largest entry in the matrix, and cor-responds to the strongest response. This could be the max-imum in any of the matrices or the maximum overall.

Cm=MAX Pd[ i , j ](8)

6. Dissimilarity

∑i , j=1

G

Cij∨i− j∨¿(9)¿

7. Autocorrelation

Other statistical approaches include an autocorrela-tion function, which has been used for analyzing the regularity and coarseness of texture by Keizer. This function evaluates the linear spatial relationships between primitives. The set of autocorrelation coefficients shown below are used as texture features Large value of MD indicates test image is of poor quality. Ideally it should be zero.

C ( p , q )= MN(M −p )(N−q)

∑i=1

M−p

∑j=1

N−q

f ( i , j ) f (i+ p , j+q )

∑i=1

M

∑j=1

N

f 2(i , j)

(10)

8. Inverse different MomentIDM is also influenced by the homogeneity of the

image. Because of the weighting factor IDM will get small contributions from inhomogeneous areas. The result is a low IDM value for inhomogeneous images, and a relatively higher value for homogeneous images.

IDM=∑i=0

G−1

∑j=0

G−1 11+(i− j)2 P (i , j )(11)

9. EntropyEntropy is a measure of information content. It mea-

sures the randomness of intensity distribution.

C e=−∑i∑

jPd [ i , j ] ln Pd [i , j ](12)

10. HomogeneityA homogeneous image will result in a co-occurrence matrix with a combination of high and low P [i,j]’s.

Ch=∑i∑

j

Pd [i , j ]1+¿ i− j∨¿

(13)¿

11. Cluster Prominence

PROM=∑i=0

G−1

∑j=0

G−1

{i+ j−µx−µ y}4 × P(i , j)(14)

12. Cluster Shade

SHADE= ∑I=0

2 G−2

(i−2 µ)3 H s (i|Δx , Δy )(15)

Where

µ=12 ∑

i=0

2 G−2

i H s(i∨Δx , Δy)

B. Feature selectionFeature selection algorithms are important to recognition

and classification systems because, if a feature space with a large dimension is used, the performance of the classifier will decrease with respect to execution time and to recognition rate. The execution time increases with the number of features because of the measurement cost. The recognition rate can de-crease because of redundant features and of the fact that small number of features can alleviate the course of dimensionality when the training samples set is limited, leading to overtrain-ing. On the other hand, a reduction in the number of features may lead to a loss in the discrimination power and thereby lower the accuracy of the recognition system.

In order to determine the best feature subset for some crite-rion, some automatic feature selection algorithm can be ap-plied to the complete feature space, varying the number of se-lected features from 1 to m. The basic feature selection algorithm is shown in the follow-ing.

Algorithm 2

1. Start with the empty set Y0= {φ}2. Select the next best feature X+=argmax [J (Yk+X)]; x¢Yk

3. Update Yk+1=Yk+ X+; k=k+14. Goto 2

SFS performs best when the optimal subset has a small number of features. ¬When the search is near the empty Set, a large number of states can be potentially evaluated. Towards the full set, the region examined by SFS is narrower since most of the features have already been selected. The search space is drawn like an ellipse to emphasize the fact that there are fewer

states towards the full or empty sets. As an example, the state space for 4 features is shown. Notice that the number of states is larger in the middle of the search tree. The main disadvantage of SFS is that it is unable to remove features that become obsolete after the addition of other features.

IV. CLASSIFICATION BY MULTINOMIAL MULTIVARIATE BAYESIAN REFERENCE

It is a generative (model based) approach, which offers a useful conceptual framework for Glaucoma images. Any kind of abnormalities can be classified, based on a probabilistic model specification Features that describe data instances are conditionally independent given the classification hypothesis. Multivariate multinomial distribution for discrete data that fit assumes each individual feature follows a multinomial model within a class. The parameters for a feature include the probabilities of all possible values that the corresponding feature can take.Understanding Baye’s ruleA prior probability is an initial probability value originally obtained before any additional information is obtained.A posterior probability is a probability value that has been revised by using additional information that is later obtained.Bayes Rule is stated as follows,

P (h/d) = P (d/h) P (h) / P (d) (16)

Whered= data, h= hypothesis (model). By rearranging the above equation

P (h/d) P (d) = P (d/h) P (h) (17)

P (h/d) probability of instance d being in class h, this is what we are trying to compute. P (d/h) probability of generating instance d given class h, we can imagine that being in class h, causes you to have feature d with some probability. P (h) probability of occurrence of class h This is just how frequent the class h, is in our database, p (d) = probability of instance d occurring, This can actually be ignored, since it is the same for all classes

Max [P (h/d)] = Max [P (d/h) P (h)]

P (d, h) = P (d, h) (18)

The same joint probability on both sides. This gives the proper classification of CT scan image into the normal or abnormal

V. RESULTS AND DISCUSSIONS

A. Experimental Setup The images have been acquired from the CT scan de-

partment of Shrimati Kashbai Nawale hospital and medical

college, Narhe, Pune. To avoid any effects on the parameters of images like change in dimensions, change in texture val-ues due to different sources, we have taken all the images from the same source and same CT scan machine. The CT scanner consists of a couch upon which the patient is placed and a circular gantry through which the couch with patient is passed. Within the gantry is a rotating ring with an X-ray source opposed to a linear array of detectors. That’s why the images obtained from CT scan machine were in DICOM format. A typical CT scan machine is shown in figure 2. Into the result analysis first image is a input test CT scan image in figure 3, after that preprocessed filter image in figure 4, then we plot histogram of before and after equalization of test image in figure 5, after that grey level distribution of test image and filter image in figure 6 & 7 then total 12 texture features are extracted in figure 8, then shape features are cal-culated in figure 9 and then classification of test is obtained in figure 10.

Figure 2. CT Scan Machine- A Philips 64 slice 'Brilliance' Scanner

B. Discussions

. Into the result analysis first image is a input test CT scan image in figure 3 which has a many cells overlapped on each other need to reduce noise and to enhance the contrast, after that pre-processed filter image in figure 4 in which we reduce the noise, then we plot histogram of before and after equalization of test image in figure 5 by using The Contrast Limited Adaptive Histogram Equalization (CLAHE), after that grey level distribution of test image and filter image in figure 6 & 7 calculated then total 12 texture features are ex-tracted in figure 8, then by comparing 12 features with each other we select the best one by using SFS algorithm then shape features are calculated in figure 9 and classification of test is obtained in figure 10.

Figure 3. Test image

Figure 4. Filtered Image

Figure 5. Histogram Equalisation

Figure 6. Gray level distribution for test image

Figure 7. Gray level distribution for filtered image

Figure 8. Texture Features

Figure 9. Shape Fiture

Figure 10. Classification dialogue box

VI. CONCLUSION

The work in this research involves using kernelised Bayesian to classify the input which is CT lung image into normal and abnormal conditions. We intend to prove that this kennel technique will help to get more accurate result. Thus we have achieved high accuracy. From the previous research work it is found that the HNN segmentation results are more accurate and reliable than FCM clustering in all cases. The HNN succeeded in detecting and segmenting the nuclei and cytoplasm regions. However FCM failed in detecting the nuclei, instead it detected only part of it. In addition to that, the FCM is not sensitive to intensity variations as the segmentation error at convergence is larger with FCM compared to that with HNN. Further we have to use modified FCM algorithm.

REFERENCES[1] Jason J. Corso, Eitan Sharon, Shishir Dube, Suzie El-Saden,Usha Sinha, and Alan Yuille, “Efficient Multilevel Brain TumorSegmentation with Integrated Bayesian Model Classification”,IEEE Transactions on medical imaging, vol. 27, no. 5, May2008.[2] S.Shah, “Automatic Cell Images segmentation using a Shape-Classification Model”, Proceedings of IAPR Conference on Machine vision Applications, pp. 428-432,Tokyo, Japan, 2007.[3] F.Taher and R. Sammouda, “Lung Cancer Detection based on the Analysis of Sputum Color Images”, Proceeding in the International Conference on Image Processing, Computer Vision, & Pattern Re-cognition (IPCV'10:WORLDCOMP 2010), pp. 168-173, Las Vegas, USA, 12-15 July, 2010.[4] Balaji Ganeshan, Sandra Abaleke, Rupert C.D. Young, Chris-topher R. Chatwin, Kenneth A. Miles, “Texture analysis of non-small cell lung cancer on unenhanced computed tomography: initial evid-

ence for a relationship with tumour glucose metabolism and stage,” Cancer Imaging, v.10 (1): 137–143, 2010 July 6. [5] Ali RAfiee, Hassan Masoumi, and Alireza Roosta, “Using neural network for liver detection in abdominal MRI images” International Conference on Signal and Image Processing Applications, pp. 21-26. 2009.[6] S. K Vijai Anand, “Segmantation coupled textural feature classi-fication for lung tumor prediction”, ICCCCT’10, pp. 518-524, 2010.

[7] S. Anna Durai and E. Anna Saro, Image compression with back-propagation neural network using cumulative distribu-tion function, World Academy of Science, Engineering and Technology, pp. 60-64, 2006.[8] Michela Antonelli, Beatrice Lazzerini, and Francesco Marcelloni, Segmentation and reconstruction of the lung volume in CT images, ACM Symposium on Applied Computing, pp.255-259, 2005.