enhanced gaussian bat algorithm and doppler effect (egba
TRANSCRIPT
Enhanced Gaussian Bat Algorithm and
Doppler Effect (EGBA-DE) and Supervised
HFNN Approach for Effective Classification of
Alzheimer‟s Disease 1C. Geetha and
2D. Ramyachitra
1Department of Computer Science,
Bharathiar University.
Sri Kayaka Parameswari Arts & Science College for Women,
Chennai, India. 2Department of Computer Science,
Bharathiar University,
Coimbatore, India.
Abstract Dementia is a growing health problem and Alzheimer’s disease (AD) is
the leading cause of dementia in the elderly accounting for 50–60% of all
cases. Magnetic Resonance Imaging (MRI) has emerged as a significant tool
to recognize in-between biomarkers of AD appropriate to its capability to
calculate regional changes in the brain with the purpose of are thought to
reflect disease severity and progression. This paper develops a novel
methodology for classification of AD from MRI images designed for
medical support. In this paper, Enhanced Gaussian Bat Algorithm and
Doppler Effect (EGBA-DE) is proposed for preprocessing step for removing
the noise features. An interesting extension of the BA will be to use
different schemes of wavelength or frequency variations by using Gaussian
distribution function. Then the feature extraction is performed by using
Haar Wavelet Transform method. In this work, feature selection is
performed by Gaussian Membership Based Fuzzy C-Means (GMFCM). The
similarity between two features are computed using Euclidean distance
function, then irrelevant features in the feature set is removed from original
feature subset. Finally Supervised Hybrid Fuzzy Neural Network (SHFNN)
International Journal of Pure and Applied MathematicsVolume 119 No. 10 2018, 273-299ISSN: 1311-8080 (printed version); ISSN: 1314-3395 (on-line version)url: http://www.ijpam.euSpecial Issue ijpam.eu
273
is used for classifying the features from the feature selection phase. The
features consequently derived are employed for the purpose of training a
SHFNN to classify features into three classes like Alzheimer’s, Mild
Alzheimer's and Huntington's disease. The experimental result shows that
the proposed SHFNN classifier provides better classification accuracy than
other classification approaches.
Index Terms:Dementia, Alzheimer’s disease (AD), Magnetic Resonance
Imaging (MRI), Feature Selection, Feature Extraction, Classification,
Enhanced Gaussian bat Algorithm and Doppler Effect (EGBA-DE,
Gaussian Membership based Fuzzy C-means (GMFCM), and Supervised
Hybrid Fuzzy Neural Network (SHFNN).
International Journal of Pure and Applied Mathematics Special Issue
274
1. Introduction
Alzheimer‟s disease (AD) is an advancing neurodegenerative condition, which
leads to the loss of cognitive capabilities and memory, along with corresponding
high morbidity rate and expense to society [1]. Accurately diagnosing AD, in
addition to detecting the prodromal stage, Mild Cognitive Impairment (MCI)
comprise to be an significant primary step towards a remedy and have remained
in the focus of several neuroimaging studies. Magnetic resonance imaging
(MRI) is typically a neuroimaging technique, which can be utilized for
visualizing the brain anatomy with a greater degree of spatial resolution and
contrast between the types of brain tissue. Structural MRI techniques have been
utilized for identifying the regional volumetric variations in brain areas that are
observed to be related with AD and MCI, illustrating the use of such techniques
for examining this disease [1], [2]. Particularly, structural MRI has detected AD
and MCI-associated cross-sectional differences and longitudinal variations in
volume and size of certain brain areas, like the hippocampus and entorhinal
cortex, in addition to the regional variations in gray matter, white matter and
Cerebrospinal Fluid (CSF) on the basis of a voxel-after-voxel [2]. In the recent
times, MRI data have been highlighted in machine learning experiments that are
targeted at the classification of subjects, which are AD vs. Cognitively Normal
(CN) or MCI vs. CN [3]. The current techniques use network analysis [1], [4] or
employ machine learning directly over the voxels [2], [5]. However, these
techniques, just take a two-way classification paradigm, AD vs. CN, into
consideration, where an exact decision boundary between these groups can be
easily acquired. Practically, this advancing neurodegenerative disease is a
continuum, having the subjects moving through several stages ranging from
MCI to AD, rendering the classification to be much more tedious.
Currently, machine learning and pattern classification techniques have been
extensively used in designing a computer-aided brain disease diagnosis system
with the help of neuro images like Magnetic Resonance Imaging (MRI) ,
Positron Emission Tomography (PET) [6], functional MRI (fMRI) [7], and
Diffusion Tensor Imaging (DTI) [8]. Research works have indicated that
structural MRI is the most standardized form of imaging modality in clinical
practice [9] and it is also helpful in tracking various clinical stages of AD [10].
Hence, the technique is assessed on structural MR images. Multiple kind of
features can be acquired from the structural MRI of the entire brain, like
intensities or gray-matter densities [11], group comparison made of cortical
thickness [12], morphometry [13], and texture measures [14]. The integration of
various kinds of features can enhance the accuracy of the diagnosis made of AD
compared to techniques that utilize only one feature [15]. The important issue
that is encountered in AD classification includes preprocessing and feature
extraction that leads to inaccurate classification on MRI images. The already
available techniques render results that are inaccurate in segmentation and also
feature selection. Moreover, it faces challenges with computational complexity
owing to the number of iterations. This way, the overall performance of the
International Journal of Pure and Applied Mathematics Special Issue
275
MRI classification is decreased substantially. Therefore, the newly introduced
research work, MRI classification has gained focused in these new models.
Texture analysis could analyze the subtle changes of body; therefore, it has been
widely used in AD diagnosis for extracting texture features. Moreover, the
number of feature dimensions in neuroimaging is commonly higher than the
number of samples. In order to solve the problem of overfitting, it is necessary
to select features. In the classification based on neuroimaging, several feature
selection techniques have been proposed, for example, univariate methods,
multivariate approaches, perturbation method, and Support Vector Machine
Recursive Feature Elimination (SVM-RFE). Firstly, the optimization algorithm
is employed for preprocessing of noise removal. Then the feature extraction is
performed by using wavelet transformation methods for better classification
results. Thirdly improve the process of feature selection by introducing
clustering algorithm. Moreover, we also realized some proposed Supervised
Hybrid Fuzzy Neural Network (SHFNN). It shows good performance in
Alzheimer's disease diagnosis. Because of this, feature selection is taken as a
comparison method in this work.
2. Literature Review
Recently, dementia has emerged to be a critical threat to global health and
society. There was estimation that in the whole world, 35.6 million people were
affected by dementia in the year 2010.
Westman et al [16] utilized the hippocampal volume obtained on MRI data to
act as the features and the Orthogonal Partial Least Squares to Latent Structures
(OPLS) analysis to be the classifier to distinguish AD and MCI from elderly
normal individuals. When it is about the shape features, every segmented
hippocampus is defined by a series of Spherical Harmonics (SPHARM), whose
coefficients were calculated with the SPHARM-PDM software designed by the
University of North Carolina and the National Alliance for Medical Imaging
Computing.
George and Karnan [17] introduced a new technique for the enhancement of
MRI image that, in turn, is dependent on the Modified Tracking Algorithm,
Histogram Equalization and Center Weighted Median (CWM) filter. This
technique comprises of two schemes. The first scheme is the application of the
modified tracking algorithm to eliminate the film artifacts, labels and skull
region and therefore using the Histogram Equalization and CWM filter methods
individually for the images enhancement. Priyanka and Balwinder [18]
proposed a Median Filter approach for the removal of salt and pepper noise and
Poisson noise away from the images. The value of every pixel is fixed to
median value of the pixels close to the respective input pixels. Then this filter is
used for removing these noises and afterwards the bounding box approach is
used for detecting the tumor position.
International Journal of Pure and Applied Mathematics Special Issue
276
Ramalakshmi and Chandran [19] have developed an anisotropic filter for the
elimination of background noise and thereby protecting the edge points present
in the image. This technique is involved with concurrent filtering and contrast
stitching. The selection of a Diffusion constant is related to the noise gradient
and it smoothens by eliminating noises present in the background with the
filtering done with a right value of threshold. Gerardin et al [20] used two sets,
one for every hippocampus, of 3D SPHARM coefficients to be the features and
adopted a univariate feature selection technique integrated with a bagging
approach to extract the most distinguishing features. Atrophy in the earlier
stages of AD is not restricted to the hippocampus or the entorhinal cortex. Other
regions are also affected in AD patients and Mild Cognitive Impairment (MCI)
patients. Hence, multi-Region Of Interest (ROI)-based feature extraction has
gained a lot of research interest.
Liu and Shen [21] introduced a multi-channel pattern analysis technique to
evaluate the hypo-metabolism patterns of AD and MCI on Fluorodeoxyglucose
(FDG)-Positron Emission Tomography (PET) data and recognized 21 brain
areas to be the most distinguishing biomarkers. Sakthivel et al [22] focused on
including various types of information, along with textual data, image visual
features obtained from scans along with direct input from the doctor. Generally,
features can have coefficients corresponding to a spectral transform of image
signal, e.g. Fourier or Discrete Cosine Transform coefficients (DCT), statistics
on image gradients etc. Features utilized for describing brain images include
Local Binary Patterns (LBPs) and DCT.
Xiao et al [23] performed the extraction of combination of three diverse features
from structural MR images which are: gray-matter volume, Gray-Level Co-
Occurrence Matrix (GLCM), and Gabor feature. These features can get both the
2D and 3D information about brains, and the experimental results reveal that
achieving a better performance is possible through the multi feature fusion.
Khedher et al [24] introduced a new classification technique dependent on
Independent Component Analysis (ICA) and supervised learning technique is
introduced to be used on segmented brain MRI from Alzheimer‟s disease Neuro
Imaging Initiative (ADNI) subjects for the purpose of automatic classification
task. At last, every brain image from the database gets projected onto the space
that is spanned by this individual components basis for undergoing feature
extraction, and a Support Vector Machine (SVM) is adopted to deal with the
task of classification. An 87.5% accuracy in recognizing AD from NC, along
with 90.4% specificity and 84.6% sensitivity is received. Based on the
experimental results, this novel technique can successfully distinguish AD, MCI
and NC individuals.
With the features being estimated on training cases, a classifier could be trained
and used for predicting the diagnosis related to a testing case, whose features
are obtained in the same manner. Horn et al [25] adapted the image features that
are compressed by the Partial Least Squares (PLS) to LDA for discriminating
International Journal of Pure and Applied Mathematics Special Issue
277
AD from FTD and attained an accuracy of 84 %, a sensitivity of 83 %, and a
specificity of 86 % over perfusion Single-Photon Emission Computed
Tomography (SPECT) images. Lopez et al [26] used the multivariate
techniques, like PCA and LDA, for feature extraction, and thereafter used the
Bayesian framework for automated diagnosis of AD and NC employing PET
and SPECT. Sarraf et al [27] provided an overview of deep learning-based
pipelines used for distinguishing the Alzheimer‟s MRI and fMRI obtained from
normal healthy control data for a particular age group. It nearly exactly
discriminated the Alzheimer‟s patients from healthy normal brains. Experiments
carried out on the MRI dataset used with no skull-stripping preprocessing had
revealed that it performed better compared to various traditional classifiers in
terms of accuracy.
Liu et al [28] presented the multifold Bayesian Kernelization technique that can
distinguish AD from NC with a greater accuracy, but yielded poorer results in
the diagnosis of MCI-converter (MCIc) and MCI-non-converter (MCInc). Zhao
et al [29] introduced an enhanced Iterative Trace Ratio (iITR) algorithm for
resolving the Trace Ratio Linear Discriminant Analysis (TR-LDA) problem for
the diagnosis of dementia and accomplished better performance compared to the
Principal Component Analysis (PCA), Locality Preserving Projections (LPP),
and Maximum Margin Criterion (MMC).
Siddiqui et al [30] performed the extraction of the Master Features of the
images by making use of a fast Discrete Wavelet Transform (DWT), and
thereafter these distinguishing features are analyzed further by means of
Principal Component Analysis (PCA). Various subset sizes of the principal
feature vectors are given to five diverse decision models. The classification
models are inclusive of the J48 decision tree, K-Nearest Neighbour (KNN),
Random Forest (RF), and Least-Squares Support Vector Machine (LS-SVM)
with polynomial and radial basis kernels.
Lama et al [31] studied and compared AD diagnosis techniques employing
structural Magnetic Resonance (sMR) images for the discrimination AD, Mild
Cognitive Impairment (MCI), and Healthy Control (HC) subjects employing a
Support Vector Machine (SVM), an Import Vector Machine (IVM), and a
Regularized Extreme Learning Machine (RELM). The performance of these
classifiers for volumetric MR image data is compared with the help of ADNI
datasets. The experiments conducted on the ADNI datasets revealed that RELM
with the feature selection technique can considerably enhance the classification
accuracy of AD from MCI and HC individuals.
Luo et al [32] explained about an automated AD recognition algorithm, which is
on the basis of deep learning on 3D brain MRI. The algorithm makes use of a
Convolutional Neural Network (CNN) to complete AD recognition. It is distinct
in the fact that the three dimensional topology of brain is regarded to be
complete in AD recognition, leading to an accurate identification. The CNN
utilized in this work comprises of three sequential sets of processing layers, two
International Journal of Pure and Applied Mathematics Special Issue
278
completely connected layers and a classification layer. In the structure, each one
of the three groups is built with three layers, inclusive of a convolutional layer,
a pooling layer and a normalization layer.
Long et al [33] introduced a machine learning technique for discriminating the
patients with AD or Mild Cognitive Impairment (MCI) from the healthy elderly
subjects and for predicting the AD conversion in MCI patients by calculating
and evaluating the regional morphological variations of brain between different
groups. The new technique provided an accuracy of 96.5% in distinguishing
mild AD from healthy elderly providing the whole-brain gray matter or
temporal lobe to be ROI, 91.74% in discriminating progressive MCI from the
healthy elderly people and 88.99% in categorizing progressive MCI against the
stable MCI with amygdala or hippocampus as ROI.
3. Proposed Methodology
The major contribution of the work is to improve the MRI classification result
by selecting the optimal features from the given images. The Enhanced
Gaussian Bat Algorithm and Doppler Effect (EGBA-DE) is used for
preprocessing step for removing the noise features. Then the feature extraction
is performed by using Haar Wavelet Transform (HWT) method. Then Gaussian
Membership Based Fuzzy C-Means (GMFCM) is proposed for feature
selection; it computes the cluster center using Gaussian weights with fuzzy
membership function. The similarity between two features are computed using
Euclidean distance function, then irrelevant features in the feature set is
removed from original feature subset. Then Supervised Hybrid Fuzzy Neural
Network (SHFNN) is used for classifying the features from the feature selection
phase. It gives improved classification outcome while matched up with the other
techniques.
Image Denoising - Enhanced Gaussian Bat Algorithm and Doppler Effect (EGBA-DE)
For medical image processing, noise is one among the major issues that un
enviably disgraces medical images. Image denoising eliminates noise from
image, when preserving its quality. Noise removal is used to several medical
images improvements. In digital images, there are diverse kinds of noise.
Certain categories contain salt and pepper noise, speckle noise, Gaussian noise,
Rician noise, and so on. Speckle noise is detected in ultrasound images,
Gaussian noise frequently seems in natural images, Rician noise have an
emotional impact on MRI when random noise could seem in any kind of
images. Pattern of the noise based on its source [34]. In case of digital images, it
is very hard to eliminate the noise, which contains a low frequency as it is hard
to differentiate low frequency noise from the real signal [35]. Owing to bad
instruments in image processing or interface, Generation of noise could arise.
Noise on digital images might be gotten by compression, error in transmission
or some other aspects. By numerous kinds of noise, MRI images are disgraced
[36].
International Journal of Pure and Applied Mathematics Special Issue
279
Gaussian noise is a general kind of noise in MRI images. Noise is distributed
evenly over the image, at every pixel of the image, random value from Gaussian
distribution was included. Gaussian distribution of noise is stated here:
(1)
Here g is gray level, m is known as average or mean of the function and is
called the standard deviation. Gaussian noise is a statistical noise that
comprises the probability density function of the normal or Gaussian
distribution. Noise values were gotten from the Gaussian distribution. Gaussian
noise is the noise with a Gaussian amplitude distribution [37]. From complex
Gaussian noise, Rician noise is gotten, which is one more noise that is capable
of disgracing magnetic resonance images. This noise had probability density
function for intensity x provided by the subsequent equation:
(2)
Rician noise is based upon the signal in MRI that is not zero-mean. In bright
areas, distribution of Rician noise is nearer to Gaussian. Speckle noise called
Multiplicative noise seems in diverse imaging systems and as well in MRI
images. It seems in images, which is capable of giving valuable diagnostic info
regarding the disease in the human body. This noise is produced by errors in
data transmission and it is dependent upon gamma distribution such that it is
well-defined by the subsequent equation:
(3)
Denoising method remove the noise by utilizing a filter. Techniques of filtering
could contain a defect, with the intension that high-frequency signals are
removed from the component producing blurry edges in MRIs. Eliminating the
random noise in the MRIs by means of an EGBA-DE is elucidated in the sub
sequent‟s.
Bat Algorithm (BA)
Bat algorithm was carried out under motivation of echolocation conduct of bats
[38]. Usually the bats are carried out dependent upon the three rules:
1. Each and every bat utilizes echolocation to intellect distance, and they as
well „know‟ the variance amid food/prey and background barriers in
certain magical manner;
2. Bats fly arbitrarily with velocity vi at position xi with a frequency fmin,
changing wavelength and loudness A0 to look for prey. They could
mechanically regulate the wavelength (or frequency) of their produced
pulses and fine-tune the rate of pulse emission , based upon the
proximity of their target;
International Journal of Pure and Applied Mathematics Special Issue
280
3. Though the loudness could change in numerous means, presume that the
loudness changes from a large (positive) A0 to a least constant value
Amin. In Figure 4.2., the flowchart of the Bat Algorithm is depicted.
In BA variation of the frequency as well as wavelength mainly increase the
outcomes. Rather than carrying out linear implementation, interesting extension
would be to utilize Gaussian distribution for wavelength as well as frequency
variations. An additional main significant problem of the present BA is the time
delay amid pulse emission as well as the echo bounced happens. With the aim
of resolving this issue additional natural expansion to the present BA will be to
utilize the directional echolocation and Doppler Effect that might bring about
improving image enhancement outcomes.
Enhanced Gaussian Bat Algorithm and Doppler Effect (EGBA-DE)
Gaussian distribution is the well-known distribution.
(4)
The mean is depicted as µ and the variance is denoted as ζ2 for frequency f.
The Standard Deviation is calculated in this manner:
(5)
Mean is an arithmetic average of the frequencies, computed by toting all the
frequencies and dividing by the total amount of frequencies
(6)
Movement of Virtual Bats
In the EGBA-DE, virtual bats are utilized. Describe the rules how their pixel
position and velocities vi in a d-dimensional image search space are brought-
up-to-date. The novel quality analysis solution and velocities at iteration
t are provided by
(7)
(8)
(9)
Here is known as a random vector derived from a uniform
distribution. is known as the present global best quality outcome that is
placed subsequently comparing all the solutions among teach and every n bats
at every iteration t. Since the product is the velocity increment, utilize (or
λi ) to regulate the velocity change when fixing the other factor λi (or ), based
upon the kind of the problem of interest. For the local quality analysis part, as
soon as a solution is chosen amongst the present best solutions, a novel quality
assessed solution for every bat is produced locally by means of random walk.
(10)
International Journal of Pure and Applied Mathematics Special Issue
281
Here ε is known as a random number that is derived from a uniform distribution
in [−1,1], when At = is the average loudness of all the bats at iteration t.
Loudness and Pulse Emission
Likewise, the loudness Ai and the rate ri of pulse emission must be brought up to
date consequently since the iterations go on. Since the loudness reduces as soon
as a bat has identified its ideal quality outcomes, when the pulse emission rate
rises, the loudness is selected as any value of convenience. For easiness, utilize
A0 = 1 and Amin = 0, presuming Amin = 0 denotes that a bat has identified the
prey and provisionally stop producing any sound.
(11)
(12)
Here α and γ are constants. For any 0 < α < 1 and γ > 0, have
as (13)
In the modest case, we could utilize α = γ, α = γ = 0.9 in the simulations. The
selection of parameters needs certain testing. Firstly, every bat must contain
diverse values of loudness as well as pulse emission rate, and this is attained by
randomization. Perhaps, the preliminary loudness is archetypally be around
[39], when the primary emission rate isabout zero, or any value ∈ [0,1] if
utilizing (13). Their loudness as well as emission rates are brought up to date
only when the novel solutions are enhanced and these bats are moving in the
direction of the ideal solution [40]. Dependent upon the Peak Signal To Noise
Ratio (PSNR), the fitness value of the Bat Algorithm is calculated. The
fundamental steps of the EGBA-DE are précised as the pseudo code depicted in
algorithm 1., dependent upon the above guesstimates and idealization.
Algorithm1. EGBA-DE Algorithm for image denoising in AD prediction
1. Input : Pi (i = 1,2,...,n) with number of pixels
2. Begin
3. Initialize a population of n bats Pi (i = 1,2,...,n) with number of pixels in the
image and velocity vi
4. Initialize frequencies , pulse rates and the loudness .
5. While (t <Max number of iterations)
1.1. Generate new optimal PSNR results by adjusting frequency new
frequency from equation (7), and updating velocities and
locations/solutions [(8) to (9)]
6. If (rand > ri)
6.1.Select a best PSNR quality results among the best PSNR values
6.2.Generate a local PSNR values around the selected best PSNR values
7. End if //step 6
8. Else
8.1. Generate a new PSNR values by flying randomly
9. if (rand < Ai & f(Pi) < f(Pi*))
9.1. Accept the new solutions
International Journal of Pure and Applied Mathematics Special Issue
282
9.2. Increase ri and reduce Ai
10. End if// step 9
11. Rank the bats and find the current best Pi∗
12. End while// step 5
13. End begin
Feature extraction- Haar Wavelet Transformation (HWT)
Haar Transform is the modest and fundamental transformation from the space
domain to a local frequency domain and could act as a sample for orthonormal
wavelet transforms [41]. When an image encompasses with N
elements; there is N/2 averages and N/2 wavelet coefficient values. In the first
half of the N element array, the averages are kept, and in the second half of the
N element, array coefficients are kept. The averages turn out to be the input for
the subsequent step in the wavelet computation. The Haar equations to compute
an average and a wavelet coefficient from an odd and even element in the
image are defined in this manner:
(14)
(15)
Steps for a 1D Haar transform of an array of N elements are in this way:
1. Identify the average of every pair of elements by utilizing Equation 1.
(N/2 averages)
2. Identify the difference amid every pair of elements and divide it by 2.
(N/2 coefficients)
3. Fill-up the first half of the array with averages.
4. Fill-up the second half of the array with coefficients.
5. Do again the process on an average part of the array till a single average
and a single coefficient are computed.
Feature selection- Gaussian Membership Based Fuzzy C-Means (GMFCM)
The Fuzzy C-Means (FCM) algorithm [42-43] is useful when the required
number of clusters is pre-determined; consequently, the FCM algorithm tries to
put each of the features to one of the clusters. What makes FCM algorithm is
different that it doesn‟t decide the absolute membership of features to a given
cluster; instead, it calculates the degree of membership that features will belong
to that cluster.
Hence, depending on the accuracy of the clustering that is required in practice,
appropriate tolerance measures be able to be put in place. Since the absolute
membership is not calculated, FCM be able to be extremely fast because the
number of iterations required to achieve a specific clustering exercise
corresponds to the required accuracy. In each iteration of the FCM algorithm,
the following objective function J [43] is minimized then the features are
selected or else the features are not selected:
International Journal of Pure and Applied Mathematics Special Issue
283
(16)
Here, N is the number of features, C is the number of clusters required, cj is the
centre vector for cluster j, and δij is the degree of membership for the ith
feature
point in cluster j. The norm, measures the similarity of the
feature point to the centre vector cj of cluster j. Note that in each iteration,
the algorithm maintains a centre vector for every clusters. These features are
computed as the weighted average of the features, in which the weights are
provided by the degrees of membership [44].
Degree of membership: For provided features , the amount of its
membership to cluster j is computed in this manner:
(17)
here, m is known as the fuzziness coefficient and the centre vector cj is
computed in this manner:
(18)
In equation (18), δij is known as the value of the amount of membership
computed in the prior iteration. The amount of membership for feature point i to
cluster j is set with a random value θij, 0 ≤ θij ≤ 1, then .
Fuzziness coefficient: In equations (17) and (18) the fuzziness coefficient m,
here , gauges the tolerance of the needed clustering. This value
identifies the amount of clusters could overlay with each other. The greater the
value of m denotes the superior of overlap amid clusters. Alternatively, the
greater the fuzziness coefficient the algorithm utilizes, a greater amount of data
points would fall ith
in a „fuzzy‟ band in which the amount of membership is
neither 0 nor 1, on the other hand somewhere amidst them.
Termination condition: The needed accurateness of the amount of
membership identifies the amount of iterations accomplished by the FCM
algorithm. The accurateness is computed by means of the amount of
membership from one iteration to the subsequent, considering the biggest of
these values crossways all feature points are taking all of the clusters. When
denote the measure of accuracy amid iteration k and k + 1 with compute its
value in this manner
(19)
International Journal of Pure and Applied Mathematics Special Issue
284
Here, and are correspondingly the amount of membership at iteration k
and k + 1, and the operator ∆, while provided a vector of values, yields the
biggest value in that vector. A novel clustering algorithm for a data objects
with diverse feature-weights that denote that data with features of diverse
weights must be clustered. In equation (16) the weight is calculated by
summation of all data points. In this research, we alter the amount of its
membership to cluster j dependent upon the weight value of the features i (wei)
and Gaussian membership function is computed in this manner:
(20)
Here ci and ζi are the centre and width of the ith
feature point vector,
correspondingly. Lastly most significant features are chosen by minimization of
the objective function (16).
Classification- Supervised HFNN approach
The Hybrid Fuzzy Neural Network (HFNN) model unites the anticipated
properties of Neural Network (NN) as well as fuzzy logic in order to create a
system that resolves the borders of NN and fuzzy logic. By means of the NN,
the HFNN model system learns inductively from the data. Fuzzy rules are taken
out from the trained NN and with the extracted rules, beforehand unseen
Alzheimer‟s disease is categorized as four classes for instance Alzheimer‟s
disease, Mild Alzheimer's disease, Huntington's disease and normal. The HFNN
model experiences two phases, called: training and implementation testing.
Figure 1 depicts the learning phase of the HFNN model. The classification
outcome is dependent upon the four classes for instance Alzheimer‟s disease,
Mild Alzheimer's disease, Huntington's disease and normal.
Figure 1: Architecture of the Learning Stage of the HFNN Model
Fuzzy Distribute Data
Configure HFNN
Start
Train NN
Evaluate NN
Optimal
NN?
Select Best Epoch
Tuning
Optimized
Re-Fuzzify/Re-
distribute data
Tune Input Fuzzy
Sets
Yes
No
No Yes
Input Fuzzy
Sets
Tuned?
Determine
Redundant Inputs
Prune Redundant
Inputs & Links
Yes
No
Extract Fuzzy
Rules
Retain Best k
rules
Predict Test Data
Using be best K
Rules
Convert Rules to
Human Readable
Form
End
International Journal of Pure and Applied Mathematics Special Issue
285
The diverse stages in the learning phase are: Fuzzification of the Data, NN
Learning, Fuzzy Sets Tuning, Pruning Input Variables, and Fuzzy Rule
Extraction & Evaluation.
a) Fuzzification of Data
The fuzzification of Alzheimer‟s samples are carried out and configured
physically. Each and every fuzzy sets are set before hand the training
Alzheimer‟s samples are put in into the neural net. The input variables are
divided by overlapping fuzzy sets and the membership functions are set
dependent upon the Alzheimer‟s samples. The requirement of membership
function is subjective and might change from diverse experts. On the other
hand, membership functions could not be allotted randomly. The amount of
membership functions is selected with the intension that the ensuing fuzzy rules
are simply readable and exact for categorizing alzheimer‟s samples.
b) Fuzzy Input Set Tuning
The tuning process of fuzzy input sets is carried out mechanically. Merely
inputs, which are continuous variables are tuned. Alteration on the limits of the
fuzzy members of a provided fuzzy set follows particular constraints. The
subsequent limitations are amended:
1. The fuzzy sets are kept overlapped with the adjacent fuzzy sets.
2. While bringing up to date the parameters, the parameters a, b, and c
must stay valid; that is to say l ≤ a ≤ b ≤ c ≤ u should all the time hold,
here [l, u] is known as the domain of the equivalent variable.
3. While keeping up to date the position parameters of the present fuzzy
member, the parameters of present member must not turn out to be
lesser than the equivalent parameters of the left neighbor or greater than
the equivalent parameters of the right neighbor.
4. The parameter „a‟ of the leftmost member and the parameter c of the
rightmost member stay static.
5. The parameter b of the present fuzzy member being brought up to date
must not turn out to be superior to the parameter c of the right member
or smaller than the parameter c of the left member. For each error back
propagated from the output to input fuzzy sets, the input fuzzy sets,
which are continuous variables, are attuned with the intention of
decreasing the error. Subsequent to the reasonable amounts of change
were done with the input fuzzy sets and up till now the error doesn‟t go
down, the tuning had attained its optimum. As soon as the fuzzy input
sets are tuned, the NN is re-trained to an optimum accurateness for the
freshly tuned fuzzy input sets.
c) Pruning the Input Variables or Fuzzy Sets
Pruning or elimination of redundant input variables would increase the
readability of a fuzzy rule base taken out for the period of the learning phase.
Elimination of the redundant input fuzzy sets would make the model simpler.
Pruning methods are amended from NN in which, examinations are made for its
International Journal of Pure and Applied Mathematics Special Issue
286
parameters that is to say weights or neurons for identifying how the error will
modify when the parameter is eliminated [45]. The fuzzy sets, which signify
numerous alzheimer‟s samples variables in all probable combinations, are
detached and the network classification performance is assessed. The process in
finding redundant alzheimer‟s samples variables is carried out by systematically
facilitating or inactivating the inputs. While an alzheimer‟s samples are
facilitated, its contribution is agreed into the NN. On the other hand while an
input is inactivated, its value is clogged in the NN. It is as if the alzheimer‟s
samples don‟t be present. With all the probable grouping of enabled and
inactivated alzheimer‟s samples, the prediction accurateness of every
combination is recorded.
d) Fuzzy Rule Extraction and Evaluation
All likely fuzzy rules are taken out from the pruned system. Each potential
mixture of the fuzzy inputs is taken as a rule. This rule is being assessed by NN
whether the combination outcomes as four classes for instance Alzheimer‟s
disease, Mild Alzheimer's disease, Huntington's disease and normal. The rules
are drawn from distinctive combinations of the enabled fuzzy inputs. There are
no two rules, which have accurately the identical mixture of the enabled fuzzy
inputs. Therefore, there are no conflicting rules that come from these groupings.
The last rules for the rule base are chosen from the extracted rules by
calculating the performance of every rule. Rules, which are just accountable for
static amounts of classifications, might be removed. Just an amount of k best
rules are stored [45]. Here, k is known as a number identified by the no of rules
drawn from particular mixture of input variables. „k‟ is relative with the no of
input variable and fuzzy member for each input variable. Specified these
aspects, k is identified by the no of rules, which contain “hits” no less than
twice in the training set. Every mined fuzzy rule is rated by counting the times it
is utilized or “hit” by the training samples. Each and every rule, which were not
utilized were removed from the list of finest k rules that were utilized just as
soon as are removed from the list. The best k rules describe the prediction
model.
e) Fuzzy Logic Implementation Stage of HFNN Model
The optimized fuzzy sets attained for the period of the learning phase are
utilized in the implementation phase. The test samples are initially fuzzified via
these enhanced fuzzy sets beforehand these input are inferred with the fuzzy
rule base. As soon as the rules are assessed and compiled, every fuzzified input
is matched up with the compiled rules and is categorized, as depicted in Figure
2.
International Journal of Pure and Applied Mathematics Special Issue
287
Figure 2: Architecture of the Fuzzy Logic Implementation Stage of the
HFNN Model
In case corresponding rule is not identified, the alzheimer‟s samples are
categorized or uncategorized. Or else, alzheimer‟s samples are categorized as
appropriately categorized or misclassified.
4. Results and Discussion
In this section, performance of proposed SHFNN classifier is evaluated. The
experiments were carried out on the MATLAB platform. The datasets includes
T2-weighted MRI in axial plane and 256 256 in-plane resolution, which were
obtained from the website of Harvard Medical School (URL:
http://med.harvard.edu/AANLIB/), OASIS dataset (URL: http:// www.oasis-
brains.org/), and ADNI dataset (URL: http://adni.loni.uc- la.edu/). The
proposed SHFNN classifier is compared with the existing Discrete Cosine
Transform (DCT) with Artificial Neural Network (ANN) and Support Vector
Machine (SVM) classifier. In the classification result is based on four classes
such as Alzheimer‟s disease, Mild Alzheimer's disease, Huntington's disease
and normal. The confusion matrix results of the four classifiers such as the
SVM, DCT-ANN , DWT-FNN and SHFNN classifiers is discussed in table 2,
3, 4 and 5 respectively. Doing so, the following standard definitions are
obtained: 1) True Positives (TP): predicts class as class. 2) True Negatives
(TN): predicts non-class as class. 3) False Positives (FP): predicts non-class as
non-class. 4) False Negatives (FN): predicts class as non-class. As a
consequence, the following standard definitions are obtained:
Fuzzifer
Fuzzy Rules
Inference
Crisp Inputs Crisp Output
Fuzzification of the Data Fuzzy System Classification of the
DATA
International Journal of Pure and Applied Mathematics Special Issue
288
(21)
(22)
(23)
Peak Signal To Noise Ratio (PSNR)
It is the ratio between the utmost possible powers to the power of corrupting
noise is given as PSNR. It considerably has an effect on the fidelity of its
representation. It can be also observed that it is the logarithmic function of peak
value of image and Mean Square Error (MSE).
(24)
Mean Square Error (MSE)
Mean Square Error (MSE) of an estimator is to quantify the difference between
an estimator and the true value of the quantity being estimated.
(25)
The proposed DWT with FNN classifier is compared with the existing Discrete
Cosine Transform (DCT) with Artificial Neural Network (ANN) and Support
Vector Machine (SVM) classifier.
In the classification result is based on four classes such as Alzheimer‟s disease,
Mild Alzheimer's disease, Huntington's disease and normal. To assess the
classifier performance, count the number of True Positive, True Negative, False
Positive (actually negative, but classified as positive), False Negative (actually
positive but classified as negative) samples and form a confusion matrix (Table
1) as given below:
Table 1: Confusion Matrix
True Class
Classifier output
Conditions Positive Negative
Positive TP FP
Negative FN TN
Table 2: Confusion Matrix for SVM Classifier
True Class
Classifier output
Conditions Positive Negative
Positive 51 10
Negative 9 30
Column sum 60 40
International Journal of Pure and Applied Mathematics Special Issue
289
Table 3: Confusion Matrix for DCT-ANN Classifier
True Class
Classifier output
Conditions Positive Negative
Positive 51 8
Negative 9 32
Column sum 60 40
Table 4: Confusion Matrix for DWT-FNN Classifier
True Class
Classifier output
Conditions Positive Negative
Positive 53 6
Negative 7 34
Column sum 60 40
Table 5: Confusion Matrix for SHFNN Classifier
True Class
Classifier output
Conditions Positive Negative
Positive 55 4
Negative 5 36
Column sum 60 40
Some preliminary results, in terms of sensitivity, specificity, and accuracy are
reported in Table 6 for the MR brain images.
Table 6: Overall Comparison Results(Proposed SHFNN model)
Classification
Methods
Accuracy (%) Error Rate (%) Specificity (%) Sensitivity (%)
SVM 81 19 75 85
DCT - ANN 83 17 80 85
DWT - FNN 87 13 85 88.33
SHFNN 91 9 90 91.67
Figure 3: Accuracy Vs Methods (Proposed SHFNN Model)
Figure 3 shows the evaluation comparison of the accuracy in terms of accuracy
(%). The figure 3 shows that the proposed SHFNN classifier has provides
0
10
20
30
40
50
60
70
80
90
100
Classification results
Accura
cy (
%)
SVM
DCT-ANN
DWT-FNN
SHFNN
International Journal of Pure and Applied Mathematics Special Issue
290
higher average accuracy of 91% which is 4%, 8% and 10% higher when
compared to other DWT-FNN, DCT-ANN and SVM classifiers.
Figure 4: Specificity Vs Methods (Proposed SHFNN Model)
Figure 4 shows the comparison results in terms of specificity metric. SHFNN
classifier has provides higher average specificity of 90% which is 5%, 10%, and
15% higher when compared to other DWT-FNN, DCT-ANN and SVM
algorithms.
Figure 5: Sensitivity Vs Methods (Proposed SHFNN Model)
Figure 5 shows the comparison results in terms of sensitivity metric. SHFNN
classifier has provides higher average sensitivity of 91.67% which is 3.34%,
6.67%, and 6.67% higher when compared to other DWT-FNN, DCT-ANN and
SVM algorithms.
0
10
20
30
40
50
60
70
80
90
100
Classification results
Specific
ity (
%)
SVM
DCT-ANN
DWT-FNN
SHFNN
0
10
20
30
40
50
60
70
80
90
100
Classification results
Sensitiv
ity (
%)
SVM
DCT-ANN
DWT-FNN
SHFNN
International Journal of Pure and Applied Mathematics Special Issue
291
Figure 6: Error Rate Vs Methods (Proposed SHFNN Model)
Figure 6 shows the comparison results in terms of error rate metric. SHFNN
classifier has provides higher average error rate of 9% which is 4%, 6%, and
10% lesser when compared to other DWT-FNN, DCT-ANN and SVM
algorithms.
Figure 7: PSNR Vs Methods (Proposed SHFNN Model)
From Figure 7, it can be proved that the proposed classifier has provides higher
PSNR (dB) of 38.48 dB, higher than 4.38 dB, 3.03 dB and 2.02 dB when
compared to other DWT-FNN, DCT-ANN and SVM algorithms. The table
values of PSNR and MSE are discussed in table 7.
0
2
4
6
8
10
12
14
16
18
20
Classification results
Err
or
Rate
(%
)
SVM
DCT-ANN
DWT-FNN
SHFNN
0
5
10
15
20
25
30
35
40
Classification results
PS
NR
(dB
)
SVM
DCT-ANN
DWT-FNN
SHFNN
International Journal of Pure and Applied Mathematics Special Issue
292
Figure 8: MSE Vs Methods (Proposed SHFNN Model)
From Figure 8, it can be proved that the proposed SHFNN classifier has
provides lesser MSE of 0.58 lesser than 0.07, 0.12 and 0.16 when compared to
other DWT-FNN, DCT-ANN and SVM algorithms. The table values of MSE
are discussed in table 7.
Table 7: PSNR and MSE Vs methods (Proposed SHFNN model)
Methods PSNR(dB) MSE
SVM 34.06 0.74
DCT - ANN 35.41 0.7
DWT - FNN 36.42 0.65
SHFNN 38.44 0.58
5. Conclusion and Future Work
In this paper Gaussian Membership Based Fuzzy C-Means (GMFCM) is
proposed for features selection. To improve the MRI classification result by
selecting the optimal features from the given images. Initially Enhanced
Gaussian Bat Algorithm and Doppler Effect (EGBA-DE) is used for
preprocessing step for removing the noise features. Then the feature extraction
is performed by using Haar Wavelet Transform (HWT). After feature selection
is performed by GMFCM. GMFCM based clustering algorithm, it computes the
cluster center using Gaussian weights with fuzzy membership function and
classification by SHFNN framework. Finally Supervised Hybrid Fuzzy Neural
Network (SHFNN) is used for classifying the features from the feature selection
phase. It provides better classification result. In the future, the various Scale
Invariant Feature Transform (SIFT) images can be improved by using hybrid
optimization based FS and transformation with advanced classification
strategies. In the future work new ensemble based classification algorithms are
introduced for detection of AD which increases detection results.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Classification results
MS
E
SVM
DCT-ANN
DWT-FNN
SHFNN
International Journal of Pure and Applied Mathematics Special Issue
293
References
[1] Sun L., Patel R., Liu J., Chen K., Wu T., Li J., Ye J. Mining brain region connectivity for alzheimer's disease study via sparse inverse covariance estimation, In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining (2009), 1335-1344.
[2] Ye J., Chen K., Wu T., Li J., Zhao Z., Patel R., Reiman, E, data fusion for alzheimer's disease study, In Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining (2008), 1025-1033.
[3] Cuingnet R., Gerardin E., Tessieras J., Auzias G., Lehéricy S., Habert M.O., Alzheimer's Disease Neuroimaging Initiative, Automatic classification of patients with Alzheimer's disease from structural MRI: a comparison of ten methods using the ADNI database, neuro image 56(2) (2011), 766-781.
[4] Supekar K., Menon V., Rubin D., Musen M., Greicius M.D., Network analysis of intrinsic functional brain connectivity in Alzheimer's disease, PLoS computational biology 4(6) (2008).
[5] Chen K., Reiman E.M., Alexander G.E., Bandy D., Renaut R., Crum W.R., Rossor M.N., An automated algorithm for the computation of brain volume change from sequential MRIs using an iterative principal component analysis and its evaluation for the assessment of whole-brain atrophy rates in patients with probable Alzheimer's disease, Neuro image 22(1) (2004), 134-143.
[6] Nordberg A., Rinne J.O., Kadir A., Långström B., The use of PET in Alzheimer disease. Nature Reviews Neurology 6(2) (2009), 78-87
[7] Suk H.I., Wee C.Y., Shen D., Machine Learning in Medical Imaging, Springer, Discriminative group sparse representation for mild cognitive impairment classification (2013), 131–138.
[8] Keihaninejad S., Zhang H., Ryan N.S., An unbiased longitudinal analysis framework for tracking white matter changes using diffusion tensor imaging with application to Alzheimer's diseas, Neuro Image 72 (2013), 153–163.
[9] Tong T., Wolz R., Gao Q., Guerrero R., Hajnal J.V., Rueckert D., Multiple instance learning for classification of dementia in brain MRI, Medical Image Analysis18(5) (2014), 808–818.
[10] Goryawala M., Zhou Q., Barker W., Loewenstein D.A., Duara R., Adjouadi M. Inclusion of neuropsychological scores in atrophy models improves diagnostic classification of Alzheimer's disease
International Journal of Pure and Applied Mathematics Special Issue
294
and mild cognitive impairment, Computational Intelligence and Neuroscience 55 (2015), 856–867.
[11] Janousova E., Vounou M., Wolz R., Gray K.R., Rueckert D., Montana G., Biomarker discovery for sparse classification of brain images in Alzheimer's disease (2012).
[12] Eskildsen S.F., Coupé P., García-Lorenzo D., Fonov V., Pruessner J.C., Collins D.L. Prediction of Alzheimer's disease in subjects with mild cognitive impairment from the ADNI cohort using patterns of cortical thinning, Neuro Image 65 (2013), 511–521.
[13] Wang W.Y., Yu J.T., Liu Y., Voxel-based meta-analysis of grey matter changes in Alzheimer's disease, Translational Neuro degeneration 4(1) (2015), 1-6.
[14] Jing Y.T., Yong H.T., Phooi Y.L., One-dimensional grey-level co-occurrence matrices for texture classification, Proceedings of the MMU International Symposium on Information and Communications Technologies (2008), 197–202.
[15] Jiji G.W., Suji G.E., Rangini M., An intelligent technique for detecting Alzheimer's disease based on brain structural changes and hippocampal shape, Computer Methods in Biomechanics and Biomedical Engineering: Imaging and Visualization 2(2) (2014), 121–128.
[16] Westman E., Simmons A., Zhang Y., Muehlboeck J.S., Tunnard C., Liu Y., Tsolaki, M., Multivariate analysis of MRI data for Alzheimer's disease, mild cognitive impairment and healthy controls, Neuro image 54(2) (2011), 1178-1187.
[17] George E.B., Karnan, M., MRI Brain Image enhancement using filtering techniques, International Journal of Computer Science & Engineering Technology (IJCSET) 3 (2012), 399-403.
[18] Priyanka, Balwinder S., An Improvement In Brain Tumor Detection Using Segmentation And Bounding Box, International Journal of Computer Science and Mobile Computing (IJCSMC) 2 (2013), 239 – 246.
[19] Ramalakshmi C., Chandran A.J., Automatic brain tumor detection in MR images using neural network based classification, Biometrics and Bioinformatics 5(6) (2013), 221-225.
[20] Gerardin E, Chételat G, Chupin M, Cuingnet R, Desgranges B, Kim H.S., Eustache F., Multidimensional classification of hippo campal shape features discriminates Alzheimer's disease and
International Journal of Pure and Applied Mathematics Special Issue
295
mild cognitive impairment from normal aging, Neuro image 47(4) (2009), 1476-1486.
[21] Liu F., Shen, C., Learning Deep Convolutional Features for MRI Based Alzheimer's Disease Classification, (2014).
[22] Sakthivel K., Jayanthiladevi A., Kavitha C., Automatic detection of lung cancer nodules by employing intelligent fuzzy c means and support vector machine, Biomedical Research 27 (2016), 123-127.
[23] Xiao Z., Ding Y., Lan T., Zhang C., Luo C., Qin Z., Brain MR Image Classification for Alzheimer’s Disease Diagnosis Based on Multi feature Fusion. Computational and Mathematical Methods in Medicine (2017), 1-14.
[24] Khedher L., Ramírez J., Górriz J.M., Brahim A., Illán I.A., Independent component analysis-based classification of Alzheimer’s disease from segmented MRI data, In International Work-Conference on the Interplay between Natural and Artificial Computation. Springer, Cham (2015), 78-87.
[25] Horn J.F., Habert M.O., Kas A., Malek Z., Maksud P., Lacomblez L., Fertil B., Differential automatic diagnosis between Alzheimer's disease and front temporal dementia based on perfusion SPECT images, Artificial Intelligence in Medicine 47(2) (2009), 147-158.
[26] López M., Ramírez J., Górriz J.M., Salas-Gonzalez D., Alvarez I., Segovia F., Puntonet C.G., Automatic tool for Alzheimer's disease diagnosis using PCA and Bayesian classification rules, Electronics Letters 45(8) (2009), 389-391.
[27] Sarraf S., Anderson J., Tofighi G., Deep AD: Alzheimer’s Disease Classification via Deep Convolutional Neural Networks Using MRI and fMRI, Bio Rxiv., (2016), 1-32.
[28] Liu S., Song Y., Cai W., Pujol S., Kikinis R., Wang X., Feng, D., Multifold Bayesian kernelization in Alzheimer’s diagnosis. In International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer, Berlin, Heidelberg (2013), 303-310.
[29] Zhao M., Chan R.H., Tang P., Chow T.W., Wong S.W., Trace ratio linear discriminant analysis for medical diagnosis: a case study of dementia, IEEE signal processing letters 20(5) (2013), 431-434.
[30] Siddiqui M.F., Mujtaba G., Reza A.W., Shuib L., Multi-Class Disease Classification in Brain MRIs Using a Computer-Aided Diagnostic System, Symmetry 9(3) (2017), 1-37.
International Journal of Pure and Applied Mathematics Special Issue
296
[31] Lama R.K. Gwak J., Park J.S., Lee S.W., Diagnosis of Alzheimer’s Disease Based on Structural MRI Images Using a Regularized Extreme Learning Machine and PCA Features, Journal of Healthcare Engineering (2017), 1-11.
[32] Luo S., Li X., Li J., Automatic Alzheimer’s Disease Recognition from MRI Data Using Deep Learning Method, Journal of Applied Mathematics and Physics 5(9) (2017), 1892-1898.
[33] Long X., Chen L., Jiang C., Zhang L., Alzheimer’s Disease Neuro Imaging Initiative, Prediction and classification of Alzheimer disease based on quantification of MRI deformation, PloS one 12(3) (2017).
[34] Yousuf M.A., Nobi, M.N., A new method to remove noise in magnetic resonance and ultrasound images, Journal of scientific research 3(1)(2010), 81.-89.
[35] Bhonsle D., Chandra V., Sinha G., Medical image denoising using bilateral filter, International Journal of Image, Graphics and Signal Processing 4 (6) (2012), 36–43.
[36] Rajeswaran N., Gokilavani C., Reduction of FBM noise in brain MRI images using wavelet thresholding techniques. Asian Journal of Information Technology 15(5) (2016) 855–861.
[37] Sarode M.V., Deshmukh D.P.R., Performance evaluation of rician noise reduction algorithm in magnetic resonance images, Journal of Emerging Trends in Computing and Information Sciences 2 (2010), 2010–11.
[38] Yang X.S., A new meta-heuristic bat-inspired algorithm. In Proceedings of Nature Inspired Cooperative Strategies for Optimization (NISCO 2010), Springer Berlin 284 (2010), 65–74.
[39] Barthelemy P., Bertolotti J., Wiersma D.S., A L´evy flight for light, Nature 453 (2008), 495–498.
[40] Yang X.S., Nature-Inspired Metaheuristic Algorithms, 2nd edn., Luniver Press, UK (2010b), 1-75.
[41] Anuj B., Rashid A., Image Compression using Modified Fast Haar Wavelet transform, Journal of World Applied Sciences 7(5) (2009), 647-653.
[42] Bezdek J.C., Ehrlich R., Full W., FCM: The fuzzy c-means clustering algorithm. Computers & Geosciences 10(2-3) (1984), 191-203.
[43] Chen W., Giger M.L., Bick U., A Fuzzy c-Means (FCM)-based approach for computerized segmentation of breast lesions in dynamic contrast-enhanced MR images, Academic radiology 13(1) (2006), 63-72.
International Journal of Pure and Applied Mathematics Special Issue
297
[44] Dembélé D., Kastner P., Fuzzy C-means method for clustering microarray data, Bioinformatics 19(8) (2003), 973-980.
[45] Haykin S., Neural Networks: A Comprehensive Foundation. Prentice Hall, Upper Saddle River (1999), 1-823.
International Journal of Pure and Applied Mathematics Special Issue
298
299
300