A BRAIN TUMOR IMAGE SEGMENTATION TECHNIQUE IN IMAGE PROCESSING USING ICA – LDA ALGORITHM WITH ARHE MODEL

A BRAIN TUMOR IMAGE SEGMENTATION TECHNIQUE IN IMAGE PROCESSING USING ICA – LDA ALGORITHM WITH ARHE MODEL

 

4.1 OBJECTIVE

            In digital image processing, image segmentation is the principal methodology that is to be used frequently.  In digital image processing, noise reduction and enhancement techniques played a vital role.  The brain is the significant and significant organ of the human body, which is to be controlled by the nervous system.  The brain tumor will occur when some brain cells presented in an unusual case.  The brain tumor detection is the crucial task for the analysis of diagnosis and the curing the tumor.  The brain tumor will identified in the correct manner, which can give a fast and effective way of determination of the tumor.  In this chapter, we proposed a brain tumor image enhancement technique with the help of the ICA – LDA (independent component analysis – linear Discriminant analysis) algorithm with ARHE (adaptive region-based histogram enhancement) model.  The image fusion technique is to apply for a combination of the two or more input images.  In this chapter, the weighted average method is to used for image fusion techniques.  The noise reduction and enhancement techniques are to apply in the preprocessing stage.  The adaptive median filter is to use for the preprocessing stage. The ARHE (adaptive region-based histogram enhancement) model is to use for improvement present in the preprocessing scene.  Feature extraction and feature optimization have utilized with the ICA (independent component analysis).  The LDA (Linear Discriminant analysis) used for the classification techniques.  Using this classifier, which is to separate the abnormal and normal stages.  When the brain tumor is denoted as the strange case, then the morphological based segmentation is to be done.  The simulation and result show the analysis of various parameters such as specificity, sensitivity, positive predictive value, negative predictive value, accuracy, precision, and recall.

 

4.2  INTRODUCTION

            Each year over 190,000 people in the United States and 10,000 people in Canada are diagnosed with various types of brain tumor. Brain tumours are the second leading cause of cancer-based deaths among children in the age group between 0 and 19 years. And also, they are the second leading cause of cancer-based deaths in young men up to age 39 years and the fifth leading cause of cancer-based deaths among young women between ages 20 and 39 years. There are over 120 different types of brain tumors, including malignant and benign, making the effective treatment highly complicated. In India, a total of 80,271 people are affected by various types of brain tumor. The first step to minimize the death rate of a brain tumor is to develop an effective diagnosing method.

Magnetic Resonance Imaging (MRI) is one of the best imaging systems to diagnose a brain tumor, even though it is crucial for clinicians to diagnose the early stage of brain tumor. So, image segmentation is the only possible tool for early detection of brain tumor. Manual segmentation by an expert will consume more time, and it is challenging to do proper segmentation. Hence, automatic segmentation algorithms are preferred in diagnosing brain tumor. Brain tumor segmentation is one of the most critical and challenging tasks in many medical-image applications because it usually involves a considerable amount of data. Artifacts due to patient’s motion, limited acquisition time, and soft tissue boundaries are generally not well defined. There is a large class of tumor types that have a variety of shapes and sizes. They may appear in different sizes and types with varying intensities of the image. Some of them may also affect the surrounding structures that change the image intensities around the tumor. Before the treatment of chemotherapy, radiotherapy, or brain surgeries, there is a need for medical practitioners to confirm the boundaries and regions of the brain tumor and determine where exactly it is located and the exact affected area. For reviewing the adverse effects of cancer, the tool can be automatic or semi-automatic for brain tumor segmentation can help and also acts as a pre-requisite stage for doctors to identify the brain tumor before performing surgeries. Image segmentation is the most major process in the image processing system, and also it has a vital role in the level of the computer vision.  Image segmentation defined as the process of the essential characteristics that are to be extracted, which are performed in the regions of the image.  The symptoms are to be in the original performance of the images, which are gray pixel value, colour, reflection features and textures etc.  It can also be the histogram features.  Image segmentation has a particular purpose, which is to separate the image into a number of the non-intersecting regions.  This region is consistent and the properties of the nearest areas, which have apparent differences.  This image segmentation method has two different classes, such as automatic segmentation and semi-automatic segmentation.  The segmentation of the brain tumor is a critical task, and the detection may be used in diagnosis and treatment planning.  The brain tumor will be detected accurately into sub-region segmentation with the help of the state of the art segmentation method, which helps to utilize the information from the multimodel MRI.

 

4.2.1 challenges in the brain tumor segmentation process

The following section depicts the issues and problems in localizing the tumor region(s) and the area of interest in the MRI brain images. Some of the challenging issues are summarized as follows:

(1) The noise in the input MRI brain image.

(2) Different intensity levels in the input brain image.

(3) The contrast level of the input MRI brain image.

(4) Difficulty in identifying the tumors in the early stages due to the absence of a clear distinction between the healthy brain tissues and the tumor regions.

(5) If the tumor size is minimal, then the localization of such tiny tumors becomes a tedious process, or sometimes there are possibilities of misinterpretation (by human experts) as healthy brain tissues.

(6) To classify the brain tumor correctly and categories its nature, numerous MRI sequence brain images have to be captured and analyzed.

(7) Tumor identification is a challenging effort and time-consuming process, and

(8) Further, feasibilities of human errors due to inter and intra-variations of brain tissues are considered as tumors and vice-versa.

 

4.2.2 ARHE (ADAPTIVE REGION BASED HISTOGRAM ENHANCEMENT) MODEL

This is a fabulous technique for contrast enhancement method, which is to be applied both the original images and medical and other initially nonvisual images.  This is the standard contrast enhancement method, which is an interactive intensity windowing method is to be processed.  This is the primary method that involves the application of each pixel histogram equalization, which mapped depends on the pixels in the region surrounding that pixel.  This helps to increase the transforming function, which is derived from the area.  It is the simplest method when each pixel is to transform, which is based on the histogram of the square region of the pixels.  In the image boundary, the pixels near the region will be treated mainly because the neighbourhood pixels do not lie on the complete within the image.  It has various properties such as neighbourhood region is the parameter of the method, which consists of the length scale, large scale contrast is to be reduced, and small scale contrast is to be enhanced.  Figure (a) shows an original image of a brain tumor and  (b) denotes the region based histogram enhanced image.

 

Figure. 4.1 (a) original image of brain tumor and (b) region based histogram enhanced image

 

4.2.3 Feature extraction and classification

In image processing and pattern recognition, feature extraction is the unique form of the decreasing of dimensionality.  In particular feature extraction algorithm, the input data is to be processed with significance, and it is removed to the redundant. The data will be processed and transformed, which helps to reduce the set of features or feature vector representation.  This helps to determine the brain tumor, which is precisely located and used to predict the next stage.  The set of features of input data is to be transformed, which is called feature extraction.  This is the process of gathering the higher-level information of an image such as shape, texture, colour, and contrast.  This method is to be used effectively to increase the accuracy of the diagnosis system with the help of the prominent selecting features.  The cellular origin is to be classified in the correct manner, which is necessary to use of immunochemistry with the antibodies with high specificity and apply those methods with high sensitivity.  Image classification helps to analyze the numerical property of different image features and which is to organize the data into different categories.  Various classification algorithms are to be presented.  In this chapter, we proposed a Linear Discriminant Analysis (LDA) classification technique is to be used.  The classification algorithm has two stages of processing, such as a training stage and testing stage.  The training class description is the major equipment of the classification process.  The classification method has two different types, such as supervised and unsupervised classification.  The supervised classification is the statistical processes that depend on the prior knowledge of the probability distribution functions.  The unsupervised classification is to segment the training stage information into prototype classes automatically.  The motivating of the criteria is to be constructed, such as independent, discriminatory, and reliable.  Figure 3.2 shows the analysis of the brain tumor, which is processing with feature extraction and classification techniques.

 

Figure.4.2 sample brain tumor image of feature extraction and classification process

 

 

4.3 PROBLEM STATEMENT

Various research proposals and methods are designed in the study of image processing applications. Although many techniques are there, the main issues are clarity and classification, which remain a massive problem for the researchers. Hence there is a need for identifying a suitable technique for the process of medical image classification to detect the abnormality. Usually, the MRI image was not clear due to noises and illuminations. Though research findings offer several methods to overcome and face the classification issue, these methods prove to be economically expensive. Hence computationally efficient method was required for identifying brain tumor was required.

 

4.4 PROPOSED METHODOLOGY

 

We proposed a brain tumor image enhancement technique with the help of the ICA – LDA (independent component analysis – linear Discriminant analysis) algorithm with ARHE (adaptive region-based histogram enhancement) model.  The image fusion technique is to apply for a combination of the two or more input images.

 

The histogram is to be defined as the probability distribution of every gray level present in the digital image.  There is various histogram equalization, and enhancement techniques are to be presented, such as recursive mean separate histogram equalization, dynamic histogram equalization, recursive sub-image histogram equalization, dualistic sub-image histogram equalization, and brightness preserving bi – histogram technique.  So we proposed a novel and accurate technique like adaptive region-based histogram equalization.  This method helps to analyze and predict the tumour in a precise manner based on the regions.

 

In this chapter, the weighted average technique is to be used for image fusion techniques.  The noise reduction and enhancement techniques are to be applied in the preprocessing stage.  The adaptive median filter is to be used for the preprocessing stage. The ARHE (adaptive region-based histogram enhancement) model is to be used for improvement present in the preprocessing stage.  The feature extraction and feature optimization have to be utilized with the ICA (independent component analysis).  The LDA (Linear Discriminant analysis) is to be used for the classification techniques. They are using this classifier, which is to separate the abnormal and normal stages.  When the brain tumor is denoted as an unusual case, then the morphological based segmentation is to be done.  Figure 4.3 shows the proposed flow of our proposed methodology.

Figure 4.3 Proposed flow of our methodology

4.4.1   preprocessing stage

The brain tumor image will be taken from the MR images.  The tumor’s tissue is to be required for the final diagnosis.  An MRI is to be used in magnetic fields, not like x – rays, which is to produce detail information about the images.  MRI is to be used to measure the tumor’s size.  To create a clearer picture, the unique dye will be used for the contrast medium.  The result of the MR scanned image is passed through the input stage.  The magnetic resonance imaging (MRI) has the artifacts and intensity variation in the image.  Artifacts affect and intensity variation affects the analysis quality.  So the preprocessing stage is used to produce a suitable image for further processing.  This stage helps to increase the image quality without changing the original image information.

 

4.4.2   Image Denoising

It is the retrieve of the digital image which has been contaminated with full of noise.  In an image, the noise will be presented. This is an unavoidable one.  This noise will have occurred during the image formation, recording or the transmission phase.  The image will be further processed after the completion of the noise removal process.  When the peak accuracy is to be required, the little bit of noise will present an image that is harmful.  The noise will be removed using the adaptive medium filter, which is proposed in this chapter.

 

4.4.2.1 Adaptive median filter

It is to be applied widely in an advance methodology on image processing.  It performs spatial processing, which helps to estimate pixels in an image is to be affected by the impulse noise.  It separates the pixels as noise with the help of every pixel in an image and its adjacent pixels comparisons.  The adjacent pixel size is the most adjustable one.  In this method, the noise pixels are to be replaced by the median pixel value of the pixels in the neighbourhood, which has been passed the noise labelling test.  The adaptive median filter has the following purpose, such as remove impulse noise, smoothing of other noise, and reduce distortion.  The performance of the medium filter is to be represented on the below algorithm.  Figure 3.4 shows the block diagram of the adaptive median filter.

Algorithm

Size – The adaptive median filter helps to change the size of the neighbourhood during operation.

GL_min – maximum value of gray level in Size_pq

GL_max – maximum value of gray level in Size_pq

GL_med – median gray levels in Size_pq

GL_pq – the gray level at coordinates (p, q)

Size_max – the maximum allowed size of Size_pq

———————————————————–

Steps

Level X:  X1 = GL_med – GL_min

X2 = GL_med – GL_max

If X1 > 0 AND X2 < 0, go to the level B

 Else increase the window size

If window size < Size_max, repeat level X

Else output GL_pq

Level Y: Y1 = GL_pq – GL_min

Y2 = GL_pq – GL_max

If Y1 > 0 AND Y2 < 0, output GL_pq

   Else output GL_med

———————————————————–

Explanation

Level X: If GL_min < GL_med < GL_max, then

• GL_med is not an impulse à go to level Y to test if GL_pq is an impulse

  Else

• GL_med is an impulse à size of the window is increased, and level X is repeated until GL_med is not an impulse and go to level Y or Size_max reached then the output is GL_pq

Level Y: If GL_min < GL_pq < GL_max, then

• GL_pq is not an impulse à output is GL_pq (distortion reduced)

  Else

• Either GL_pq = GL_min or GL_pq = GL_max à output is GL_med (standard median filter), GL_med is not an impulse (from level X)

Figure 4.4 block diagram of the adaptive median filter

4.4.2.2 Adaptive Region-Based Histogram Enhancement 

The adaptive histogram equalization is the most effective and efficient technique for the image contrast enhancement.  There are various experiments that give the results that the adaptive region-based histogram enhancement technique is effective in the contrast enhancement.  The computational complexity is expensive, which helps in the real-time occasion.  The computation will be reduced based on the following technique.

• Figure 4.5 represents the sliding window for adaptive histogram enhancement here. The window size is eight. The window center is denoted as A and B, which moves from A to B in order to obtain the histogram of the next block, we need not re scan the entire contextual region.  When the window is the slide from left to right, then it can eliminate the left column pixels of the last block from the current histogram and add the right column of the current block to it.

 

 

Figure 4.5 sliding window of the adaptive region based on histogram enhancement

 

In the histogram enhancement, the 1st position of each row is to be obtained with the help of the 1st position of the last row by subtracting the testing row and adding the new leading row.  The boundary conditions will be required, which is to be checked and resolved during the practical implementation method.

• In the adaptive region-based histogram enhancement, the grey level is to depend on the cumulative histogram function value in the original grey level. The total pixel in the particular region is to be fixed to W² (the cumulative histogram to M is fixed to W²).  The origin grey level is higher than the M / 2. The cumulative histogram is lower than the origin grey level.  Mathematically, it is represented given below,

 

Cumul_Histogram_m =   (4.1)

• The conceptual block size is to be calculated by the product of the M and the integral power of 2, and it can eliminate the multiplication and division operations, which is to be replaced by the fast bitwise shift operations.

 

(4.2)

The mathematical grey level by bitwise shift cumulative histogram x bits rightwards,

M^’ = Cumul_Histogram_m * M /  = Cumul_Histogram_m >> x        (4.3)

Where’>>’ denotes the rightward shift operations.  The adaptive region-based histogram enhancement algorithm is to be defined, which is represented given below.

Algorithm

For every pixel k in the image do

            For every pixel l in the last left column do

                        Hist [g (l)] = Hist [g (l) – 1];

End

            For every pixel l in current right column do

                        Hist [g (l)] = Hist [g (l)] + 1;

End

If m <= M / 2

Sum: Cumul_Histogram_m =

Else

                        Sum: Cumul_Histogram_m =

End

                        M^’ = Cumul_Histogram_m >> x       // (here

End

 

4.4.2   FEATURE EXTRACTION (INDEPENDENT COMPONENT ANALYSIS)

Feature selection plays a vital role, and it is a simple technique in image processing. In this chapter, the independent component analysis is to be applied for feature extraction purposes.  The independent component analysis is to be separated as the unsupervised learning because it outputs the set of highest independent component vectors.  This method by nature is related to the input distribution, but it cannot guarantee the best performance in the classification problems.  Consider the two features f₁ and f₂ which is uniformly distributed on [-1, 1] of the binary classification and the output class O is determined and it is denoted given below,

O =                                                                  (4.4)

The data points which are presented in the shaded areas present in this problem.  The problem is separable with linear, and it can easily pick out the predicted features of f1 + f2. To perform the ICA on the set of the data providers with the N-dimensional vectors while denoting the direction of independent components on feature space.  Figure 4.6 shows the apparent discontinuity in the data distribution over the extended feature space.  This vector gives the best deal of the details about the problem, which projecting it onto the (f1, f2) features space, which gives the new relevant feature to the output class.

Figure 4.6 ICA based feature extraction concept

Figure 4.7 shows the analysis and processing flow of the independent component analysis.

Figure 4.7 Processing flow analysis of the independent component analysis

The independent component analysis helps to feature extract, which leads to a tumor in the brain cells.  After completing the preprocessing level, the information will be extracted regarding with the tumor.  This method helps to process of extracting the tumor in an effective manner.

 

4.4.3   LINEAR DISCRIMINANT ANALYSIS CLASSIFICATION METHOD

The linear discriminant analysis (LDA) is one of the classification techniques in the image processing system.  It is the application of the use of the generation of matrices, which is to be represented in image processing operators acting on images.  The processing of the LDA analysis is given below.  The first stage of the linear Discriminant analysis has to generate the matrix based on training features of samples, which are derived from the LDA feature space.  The LDA has the CS classes (CS >3) and assumes P_a be the set of S_a samples of class W_a in DS dimensional space.  For each and every class, the scatter matrix will be derived between the class S_bc and within the class scatter matrix is S_{wi_c} which are defined as follows,

S_{wi_c} =  ;  =  (4.5)

S_bc =  (4.6)

The d x d is the matrix A, which is used for dimensionality reduction to make d dimensional features y = A^T x.  The covariance matrix of all samples is given by,

P =                                                 (4.7)

The linear Discriminant analysis helps to maximize the component axes for class separation.  Figure 4.8 shows the classification and projection diagram of the LDA.

Figure 4.8 projection diagram of LDA

LDA approaching steps

• To estimate the D – dimensional mean vectors for the various classes from the data set.
• To estimate the scatter matrices in two ways, such as between the class and within the class.
• To evaluate the Eigenvectors and corresponding Eigenvalues for the scatter matrices.
• To sort the eigenvectors when decreasing eigenvalues and choose the eigenvectors with the largest eigenvalues to form a D x DM dimensional matrix W (here every column represents an eigenvector)
• To use the D x DM Eigenvector matrix to transform the samples onto the new subspace. This is to be summarized with the help of the matrix multiplication.

 

The performance will be analyzed based on the various parameters which are represented in the simulation and result in the discussion stage and the analysis of various parameters such as specificity, sensitivity, positive predictive value, negative predictive value, accuracy, precision, and recall.

 

4.5 SIMULATION AND RESULT DISCUSSION

We proposed a brain tumor image enhancement technique with the help of the ICA – LDA (independent component analysis – linear Discriminant analysis) algorithm with ARHE (adaptive region-based histogram enhancement) model.  The image fusion technique is to apply for a combination of the two or more input images.  In this chapter, the weighted average technique is to be used for image fusion techniques.  The feature extraction and feature optimization have to be utilized with the ICA (independent component analysis).  The LDA (Linear Discriminant analysis) is to be used for the classification techniques.  Using this classifier, which is to separate the abnormal and normal stages.  The simulation and result show the analysis of various parameters such as specificity, sensitivity, positive predictive value, negative predictive value, accuracy, precision, and recall.  The intensity levels are to be processed with normalization got getting more predictive information from the input image.  After completing the histogram normalization process, the output image development equation is given below,

 

Out_image = δ * INT_img + (1 – δ) * In_img                       (4.8)

Where INT_img is an image obtained after the normalization process is to be applied, In_img is an input image, and Out_image is finally output image.  δ is a statistical parameter which has a range between 0 –  1.  The statistical parameter is a constant take as 0.1, and it is the optimum value.  The various parameters are to be analyzed, such as peak signal to noise ratio, discrete entropy, and absolute mean brightness error.  Table 4.1 shows that there is an analysis of the following parameters, such as PSNR, DE, and AMBE.  Figure 4.9, 4.10, 4.11 shows the analysis of the following parameters, such as peak signal to noise ratio, DE and AMBE.

 

PSNR

DE

AMBE

Statistical parameter (δ)	Existing method	Proposed method	Existing method	Proposed method	Existing method	Proposed method
0.1	25.5	25.7	7.1	7.15	2.5	2.9
0.2	26	26.3	7.13	7.18	2.4	2.8
0.3	26.2	26.5	7.15	7.19	2.3	2.7
0.4	26.7	26.9	7.16	7.20	2.2	2.5
0.5	27	27.5	7.2	7.3	2.1	2.3
0.6	27.5	27.9	7.25	7.32	2	2.2
0.7	26.8	27	7.21	7.35	2.15	2.19
0.8	26.5           26.9		7.19               7.25		2.24                    2.29
0.9	26              26.5		7.14               7.24		2.28                  2.299

Table.4.1 PSNR, DE, AMBE analysis

Figure 4.9 PSNR analysis

Figure 4.10 DE analysis

Figure 4.11 AMBE analysis

Table 4.2 shows that there is an analysis of the tumor error detection performance compare with various mechanisms that are taken from the brain data set.

 

Various methods	Existing Average tumor error detection	Proposed Average tumor error detection
LDA	1.2	1.5
PAM	0.5	0.7
SDDA	0.7	0.8
SCRDA	0.8	0.9

 

Table 4.2 analysis of the tumor error detection with the existing and proposed analysis

Figure 4.12 shows that there are the analysis and detection of tumor error with various existing methods.  In this figure, the various methods are to be analyzed for the detection of the tumor error.  The linear discriminant analysis gives better detection compare with other methods.  The PAM is the prediction analysis for microarrays, SDDA is shrinkage diagonal discriminant analysis, and SCRDA is the shrinkage centroid regularized discriminant analysis.  Our proposed method is the brain tumor image enhancement technique with the help of the ICA – LDA (independent component analysis – linear Discriminant analysis) algorithm with ARHE (adaptive region-based histogram enhancement) model and existing method is ICA in automated segmentation of brain tumours and various classification methods.

 

Figure 4.12 tumor error detection analysis

The performance analysis is to be evaluated of the various algorithms for feature extraction and classification technique with various parameters such as specificity, sensitivity and accuracy.  Table 4.3 shows that there is an analysis of the performance parameters with various methods.

 

Algorithm	Existing method sensitivity (%)	Proposed method sensitivity (%)	Existing method specificity (%)	Proposed method specificity (%)	Existing method accuracy (%)	Proposed method accuracy (%)
ICA + SVM	85.80	87.50	99.85	99.87	99.37	99.57
ICA + FCM	85.82	87.20	99.91	99.94	99.61	99.87
ICA + LDA	93.23	95.23	99.93	99.97	99.90	99.98

 

Table 4.3 performance parameter analysis

Figure 4.13 shows that there is an analysis of the sensitivity and figure 4.14 shows that there is an analysis of the specificity.  Figure 4.15 shows the analysis of accuracy.

 

Figure 4.13 sensitivity analysis

Figure 4.14specificity analysis

Figure 4.15 Accuracy analysis

There are various histogram techniques to be analyzed for entropy compared with our proposed method, which is to be tabulated in table 4.4.  Figure 4.16 shows that various histogram techniques compared to our proposed method.

Images	Original	HE	ARHE
Girl	5.59	6.25	6.5
Tank	5.99	5.88	6
Avg	6.22	6.06	6.09
Cameraman	6.86	6.8	6.9

 

Table 4.4 Entropy analysis

 

Figure 4.16 Entropy analysis

 

4.6 CONCLUSION

We proposed a brain tumor image enhancement technique with the help of the ICA – LDA (independent component analysis – linear Discriminant analysis) algorithm with ARHE (adaptive region-based histogram enhancement) model.  The image fusion technique is to apply for a combination of the two or more input images.  The weighted average technique is to be used for image fusion techniques.  The feature extraction and feature optimization have to be utilized with the ICA (independent component analysis).  The LDA (Linear Discriminant analysis) is to be used for the classification techniques.  Using this classifier, which is to separate the abnormal and normal stages.  In this chapter, the simulation stage gives the result as various parameters such as specificity, sensitivity, and accuracy and tumor detection rate. So our proposed method gives better results compared with various existing mechanisms.  Our proposed method has high specificity, sensitivity, and accuracy than the existing parameters and algorithms.