CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION...

29
62 CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION TECHNIQUES 5.1 INTRODUCTION Artificial Intelligence (AI) is one of the widely preferred automated techniques for image processing applications. The availability of an in-built memory has made these techniques much superior to the conventional image processing algorithms. The presence of in-built memory has improved the accuracy of the results which is one of the necessary characteristic features of any automated image classification techniques. Thus, these techniques have been used significantly in the medical field where accuracy plays a major role in applications such as image classification. Some of the applications of AI in medical image processing are tumor detection in scan images, anatomical segmentation in scan images, etc. Even though AI is highly advantageous, these techniques have been seldom used in the field of ophthalmology. Most of the earlier researches have been focused only on the conventional image processing algorithms for automated retinal image classification applications. In this research work, emphasis has been given on exploring the usage of various AI techniques for abnormal retinal image classification. ANN and fuzzy theory are the two significant AI techniques for imaging applications. The exploration of the improvement in the accuracy of the results of these techniques over the conventional image processing algorithms has been one of the objectives of this work. In this work, Back Propagation Neural Network (BPN), Kohonen Self- Organizing Maps (SOM), Radial Basis Function Neural Networks (RBF) and Counter Propagation Networks (CPN) are used as representatives of the ANN in the context of retinal image classification. The fuzzy nearest neighbor classifier is used

Transcript of CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION...

Page 1: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

62

CHAPTER 5

CONVENTIONAL AI BASED IMAGE CLASSIFICATION TECHNIQUES

5.1 INTRODUCTION

Artificial Intelligence (AI) is one of the widely preferred automated

techniques for image processing applications. The availability of an in-built memory

has made these techniques much superior to the conventional image processing

algorithms. The presence of in-built memory has improved the accuracy of the

results which is one of the necessary characteristic features of any automated image

classification techniques. Thus, these techniques have been used significantly in the

medical field where accuracy plays a major role in applications such as image

classification. Some of the applications of AI in medical image processing are tumor

detection in scan images, anatomical segmentation in scan images, etc. Even though

AI is highly advantageous, these techniques have been seldom used in the field of

ophthalmology. Most of the earlier researches have been focused only on the

conventional image processing algorithms for automated retinal image classification

applications. In this research work, emphasis has been given on exploring the usage

of various AI techniques for abnormal retinal image classification. ANN and fuzzy

theory are the two significant AI techniques for imaging applications. The

exploration of the improvement in the accuracy of the results of these techniques

over the conventional image processing algorithms has been one of the objectives of

this work. In this work, Back Propagation Neural Network (BPN), Kohonen Self-

Organizing Maps (SOM), Radial Basis Function Neural Networks (RBF) and

Counter Propagation Networks (CPN) are used as representatives of the ANN in the

context of retinal image classification. The fuzzy nearest neighbor classifier is used

Page 2: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

63

as the representative of fuzzy techniques. An extensive analysis on the performance

measures of these AI approaches are also presented in this chapter.

5.2 BLOCK DIAGRAM OF ANN BASED IMAGE CLASSIFICATION

The work carried out with the ANN based techniques are shown in Figure 5.1.

The four neural classifiers are experimented with retinal images. A comparative

analysis has been also performed on these classifiers based on various quality

measures.

Figure 5.1 Framework of the proposed methodology

The techniques of pre-processing and feature extraction have been performed

before the classification process to enhance the quality of the results. The image

database, pre-processing techniques and feature extraction used for the Conventional

techniques are used for ANN based classification also. A description on these

techniques has been detailed in section 3.3 and 3.4 . The neural classifiers are

initially experimented followed by the experiments on fuzzy classifier.

5.3 NEURAL TECHNIQUES FOR RETINAL IMAGE CLASSIFICATION

In this work, four neural networks are used for retinal image classification.

The complete image database is divided into training set and testing set. Initially, the

Retinal Image Database

Image Pre-processing

Feature Extraction

BPN based

image

classification

RBF based

image

classification

SOM based

image

classification

CPN based

image

classification

Fuzzy based

image

classification

Comparative Analysis

Page 3: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

64

features are extracted from all the training images of the four categories and are used

as input to these neural classifiers. After training, the testing of the stabilized

networks is done with the testing image set. The training and the testing process are

carried out with the corresponding mathematical algorithms. Finally, the

performance of these classifiers in successful pattern classification is estimated

through various quality measures.

5.3.1 BPN based Image Classification

Back Propagation Neural networks are the primarily used supervised neural

network for imaging applications. BPN belongs to the category of feed forward

networks with gradient descent algorithm for the training methodology. A properly

trained BPN tends to give reasonable response to inputs that it has never been

subjected earlier.

5.3.1.1 Architecture of BPN

In this work, a three layer network is developed. The number of neurons in the

input layer is ‘n’, where ‘n’ corresponds to the number of input features. The number

of neurons in the output layer is based on the number of output classes. When

designing a neural network, one crucial and difficult step is determining the number

of neurons in the hidden layers. The hidden layer is responsible for internal

representation of the data and the information transformation input and output layers.

If there are too few neurons in the hidden layer, the network may not contain

sufficient degrees of freedom to form a representation. If too many neurons are

defined, the network might become over trained. In this work, one hidden layer with

20 neurons is used for the classification problem. The number of hidden layer

neurons is selected based on trial and error method.The architecture is shown in

Figure 5.2.

Page 4: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

65

Figure 5.2 Architecture of BPN

The proposed architecture consists of 16 neurons in the input layer, 20 neurons in the

hidden layer and 4 neurons in the output layer. The input layer and the hidden layer

neurons are interconnected by the set of weight vectors U and the hidden layer and

the output layer neurons are interconnected by the weight matrixV . In addition to the

input vector X and output vector Y, the target vector T is given to the output layer

neurons. Since BPN operates in the supervised mode, the target vector is mandatory.

During the training process, the difference between the output vector and the target

vector is calculated and the weight values are adjusted based on the difference value.

5.3.1.2 Training Algorithm

In this work, the gradient descent algorithm is used for training the BPN. The

gradient of the performance function is used in these algorithms to determine how to

adjust the weights in order to minimize the error. The weight vectors are randomly

initialized to trigger the training process. During training, the weights of the network

are iteratively adjusted to minimize the sum of squared error.

2

otE (5.1)

where ‘t’ is the target vector and ‘o’ is the output vector.

.

.

.

.

.

.

.

.

.

. 16

n

1

2

1

2

20

1

2

4

X

Inp

ut

X

U

Ou

tpu

t

Y

V

t

t

t

Page 5: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

66

Equation (5.1) can be expressed in terms of the input training vector X, the weight

vectors and the activation function. The gradient is determined using a technique

called back propagation where computational operation is performed in the backward

direction. The weights are adjusted in the direction where the error decreases most

rapidly with the gradient being negative. Such an iterative process can be expressed

as

qqq gWW 1 (5.2)

where qW is the weight vector (includes U and V), is the learning rate and qg is the

current gradient.

The derivative of the error value with respect to the weights is the gradient vector.

Hence, the weight adjustment criterion of the BPN network is given by

qqq

w

EWW

1 (5.3)

where ‘q’ is iteration counter and ‘E’ is the difference between the target and

output of the network.

The network is said to be stabilized when the weight vectors (U and V) of the

network remain constant for successive iterations. These weight vectors are the

finalized vectors which represent the trained network. The testing images are then

given as input to the trained network and the performance measures are analyzed.

5.3.2 RBF based Image Classification

One of the emerging neural networks from the supervised category is the

Radial Basis Function neural networks. Though supervised in nature, there are some

differences between RBF and other supervised neural networks in terms of

architecture and training algorithm. These differences have made RBF networks

much significant for usage in practical applications.

Page 6: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

67

5.3.2.1 Architecture of RBF

A feed-forward structure is used in RBF with a single input layer, a hidden

layer and a summation layer. The number of neurons in the input layer is based on

the number of input features. The number of hidden layer neurons is selected

randomly. The summation layer is the third layer which is equivalent to the output

layer of any other neural networks but the operation performed by this layer is only

summation. Hence, it is named as summation layer. Unlike other 2-layer neural

networks, the number of weight matrices used in RBF is only one which forms the

interconnection between the hidden layer and the output layer. The architecture of

RBF is shown in Figure 5.3.

t

Figure 5.3 Framework of RBF

An architectural size of 16 input neurons, 20 hidden neurons and 4output neurons are

used in this work for RBF neural network. The only weight matrix connecting the

input layer and the hidden layer is denoted byW . Each neuron in the hidden layer is

considered as a radial basis function. The radial basis function used in this work is

the Gaussian function. A target vector is also supplied to the summation layer of the

RBF neural network.

5.3.2.2 Training Algorithm of RBF

The training algorithm of RBF neural network is slightly different from the

other neural networks. Normally, the weight matrices alone are adjusted in the neural

.

.

.

.

.

.

.

.

.

. 16

n

1

2

1

2

20

1

2

4

X

Inp

ut

Ou

tpu

t

W

t

t

Page 7: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

68

networks but in RBF, the parameters of the basis function are also adjusted in

addition to the weight adjustment. Also, the hidden layer neurons’ outputs are not

calculated using the weighted-sum mechanism and sigmoid activation. Instead, the

Gaussian function is used to estimate the outputs of hidden layer neurons. In this

work, the training algorithm is carried out in two stages. Initially, the parameters of

the Gaussian function are adjusted using the distance measure and then the weight

matrices are adjusted in the second stage using the least square method. The detailed

algorithm is given below:

Step 1: The mean j , standard deviation j , target vector ‘o’ and the weight

matrix W are randomly initialized.

Step 2: The values of ‘ ’ is estimated for each hidden layer neuron ‘j’ iteratively

using the formula

q

jqj

qj X 1 (5.4)

This equation is executed for ‘q’ iterations till the second term in Equation

(5.4) falls below the specified threshold value. Thus, the closeness of the

input with the mean values is used as the measure to estimate the stabilized

set of ‘ ’ values. In the above equation, ‘ ’ is the learning rate.

Step 3: The values of ‘ ’ is estimated using the formula

2

2

1

1

jj

j

(5.5)

Step 4: The stabilized output for each neuron in the hidden layer ‘j’ is determined

using the formula given by

2

2

2exp

j

j

j

XZ

(5.6)

Step 5: In the second stage, the output of the summation layer neurons ‘k’ is

estimated with the randomly initialized weights using the following formula

j

j

jkk Zwo

(5.7)

Page 8: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

69

Step 6: The error values are further estimated using the following formula

2

k

kk otE (5.8)

Step 7: If the error value is large, the following weight adjustment equation is used to

determine the stabilized set of weights

jkk

qjk

qjk Zotww 1 (5.9)

Step 8: The steps 5-7 are repeated till the error value falls below a specified threshold

level to obtain the stabilized weights.

Using these stabilized set of weights, mean and standard deviation values, the

network is tested to determine the outputs of the summation layer for every

individual input. The unknown input is allotted to the class for which the output of

the neuron is maximum.

5.3.3 SOM based Image Classification

SOM belongs to the category of unsupervised neural networks where there is

no requirement for the target vector. SOM is also called as Kohonen neural networks

where the ‘winner take-all’ training algorithm is used in the training algorithm.

Similar to statistical clustering algorithms, these Kohonen networks are able to find

the natural groupings from the training data set. As the training algorithm follows the

‘winner take-all’ principle, these networks are also called as competitive learning

networks.

5.3.3.1 Architecture of SOM

The topology of the Kohonen self-organizing map is represented as a 2-

Dimensional, single-layered output neural network. Each input neuron is connected

to each output neuron. The number of input nodes is determined by the number of

training patterns. There is no particular geometrical relationship between the output

neurons in the competitive learning networks. During the process of training, the

Page 9: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

70

input patterns are fed into the network sequentially. Output neurons represent the

‘trained’ classes and the center of each class is stored in the connection weights

between input and output neurons. The topology of SOM is shown in Figure 5.4.

Figure 5.4 Architecture of SOM

The architecture used in this work is 16-4 where 16 corresponds to the

number of input layer neurons and 4 corresponds to the number of output layer

neurons. The weight matrix is denoted byW .

5.3.3.2 Training Algorithm of SOM

One of the successive applications of SOM is categorization where the data is

grouped into any one of the categories based on some similarity measures. Mapping

of input vectors to the output vectors based on the characteristic features is

performed by the SOM training algorithm. The competitive learning rule is adopted

for training the network. The “winner take-all” concept is a methodology in which a

winner neuron is selected based on the performance metrics. The weight adjustment

is performed for the winner neurons and also the neighboring neurons of the winner

neuron. The weights of all other neurons remain unchanged. The neighboring

neurons are determined using a radius around the winner neuron. In this work, unit

.

.

.

.

.

.

. 16

1

2

1

2

4

X

Inp

ut

W

Ou

tpu

t

Page 10: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

71

radius is selected which shows the weights of the winner neuron alone is adjusted

during the process. A detailed training algorithm has been given below:

Step 1: Initialize weights ijw ; ‘i’ corresponds to the input layer and ‘j’ corresponds to

output layer.

Step 2: While stopping condition is false, do steps 3 to 6.

Step 3: The distance measure for each j (output layer neurons) is computed using the

formula given by

jD 2 i

iij xw (5.10)

Step 4: The index j with minimum jD is selected.

Step 5: The winner neuron’s weight is determined using the rule given by

oldwxoldwneww ijiijij (5.11)

xi denotes the feature values of input data set.

α denotes the learning rate.

Step 6: The training is stopped when the maximum number of iterations is reached.

The training process is carried out with the training image set. The entire process

is repeated for the specified number of iterations in the algorithm. The weights

yielded by the network in the last iteration are stored as the stabilized weights.

Further, the testing images are used to estimate the performance of the neural

network.

5.3.4 CPN based Image Classification

Counter Propagation Neural networks are one of the widely used neural

networks. It is named as hybrid neural network since the concept of both supervised

and unsupervised methodologies are involved in the training algorithm.

Theoretically, the characteristic features of both the training methodologies are

available in CPN.

Page 11: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

72

5.3.4.1 Architecture of CPN

CPN is a 2-layered network with a single hidden layer apart from the input

and the output layers. The hidden layer is also named as Kohonen layer which uses

the unsupervised methodology for training the network. The output layer is also

named as Grossberg layer which uses the supervised training methodology. Two set

of weight matrices are involved in the architecture. The target is usually supplied to

the Grossberg layer of the CPN. The architecture of CPN is same as that of BPN

which is given in Figure 5.2. The difference between these two networks is found

only in the mode of stabilizing the weights.

5.3.4.2 Training Algorithm of CPN

In the training algorithm of CPN, both the supervised and unsupervised

training methodologies are employed for adjusting the two set of weights. The

“winner take-all” strategy is used to adjust the weights of the Kohonen layer and the

output Grossberg layer’s weight adjustment is based on the error signal which is the

difference between the target and the output vector. The error signal is used to update

only the output layer weights unlike back propagation network where error is used to

update weights of both the layer. Thus, this network is named as Counter

Propagation Neural Network to show that it is contrary to the conventional BPN.

The weight adjustment between the input layer and the competition layer is given by

jq

ijq

ijq

ij ZUxUU 1

(5.12)

In the above expression, x denotes the input vector, ‘q’ is the iteration number,

‘i’ denotes the input layer neuron, ‘j’ corresponds to the hidden layer neuron and ‘ ’

is the learning coefficient. A value of ‘1’ is given to jZ if ‘j’ is the winner neuron or

‘0’ if the neuron ‘j’ fails in the competition. The fact that unsupervised training

methodology is used in the Kohonen layer is verified by Equation (5.12) which

Page 12: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

73

shows that the weights are adjusted only for the winner neuron and the weights of the

remaining neurons remain the same.

After the weight vectors ijU have stabilized, as the second step, the output

layer begins learning the desired output. The weight adjustments for the output layer

is given by

qjk

qjk

qjk VtVV

1 (5.13)

where ‘t’ is the target vector, ‘k’ corresponds to the output layer neurons. The

involvement of the target vector has justified the usage of supervised training

methodology in the output layer. After the training process, the CPN is experimented

with the testing images using the stabilized set of weights. All the four categories of

the training inputs are now represented by the weight matrices and hence the

unknown input can be associated to the corresponding category during the testing

process.

5.4 FUZZY TECHNIQUE FOR RETINAL IMAGE CLASSIFICATION

The classifier in which the fuzzy sets or fuzzy logic is used in course of its operation

is called as fuzzy classifier. A degree of membership is assigned to the four classes

and the unknown input is assigned to the class for which the membership value is

maximum. Fuzzy classifiers are often designed to be transparent, i.e., steps and logic

statements leading to the class prediction are traceable and comprehensible. The

fuzzy systems are accurate only if sufficient expert opinion is available about the

corresponding application. In this work, the fuzzy nearest neighbor classifier is used

to test the applicability of fuzzy systems for retinal image classification.

5.4.1 Fuzzy Nearest Neighbor Classifier for image classification

The operation of this algorithm is carried out in two phases. In the first phase,

each of the training data set is clustered into three clusters (background, blood

Page 13: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

74

vessels and the defective region) using the fuzzy C-means (FCM) algorithm. The

centroid value of the abnormal region is stored. This process is repeated for all the

images from the four categories and the centroid values of all the four categories are

stored. In the second phase, for an unknown testing data, the corresponding cluster

center is observed using FCM algorithm and the Euclidean distance is calculated

with the trained four categories. The testing data is allotted to the category for which

the Euclidean distance is minimum. The algorithm is given below:

Fuzzy C-means algorithm is based on minimization of the following objective

function:

c

i

n

j

ij

m

ij

c

i

ic duJcccUJ1 1

2

1

21 ),...,,,( (5.14)

uij is between 0 and 1;

ci is the centroid of cluster i;

dij is the Euclidian distance between ith centroid (ci) and jth

data point.

m є [1,∞] is a weighting exponent.

Step 1: Fuzzy partitioning of the known data sample is carried out through an

iterative optimization of the objective function shown in Equation (5.14), with the

update of membership iju and the cluster centers ic by:

C

h

m

hj

ij

ij

d

du

1

)1/(2

1 ;

n

j

m

ij

n

j j

m

ij

i

u

xuc

1

1 (5.15)

where jx = optimal feature set.

At the (q+1)th

iteration, if ququ 1 < , then the classifier is assumed to have

reached the stabilized condition.

Page 14: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

75

Step 2: The cluster centre of the defective region for all the training samples is

observed. The cluster centres A1, A2, A3 and A4 which corresponds to the four

abnormal categories namely CNVM, CRVO, CSR and NPDR is stored for further

processing.

Step 3: For a new data, the Euclidean distance between the data and all the cluster

centres of the training samples is calculated.

Step 4: The data is assigned to the class with the cluster centre whose Euclidean

distance is minimum.

ff

i AXdAXd ,min,41

(5.16)

where iA corresponds to the cluster centre (defective region) of the testing data and

fA corresponds to the cluster centre of the defective region of the training data. The

fuzzy classifier is mainly implemented to check the possibility of any performance

enhancement over the neural classifiers.

5.5 IMPLEMENTATION

The implementation of the above mentioned neural networks are carried out

using the MATLAB software. The procedural flow and the design values of various

parameters involved in the implementation are discussed in this section.

5.5.1 IMPLEMENTATION OF BPN

The step-by-step procedure of the implementation is as follows:

1) The entire image dataset is divided into training dataset and testing dataset.

2) The features extracted from the images from each category are given as input

to the neural network.

Page 15: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

76

3) The hidden layer outputs are calculated using the randomly initialized weights

and the input values.

4) Further, the output values of the output layer are calculated using the second

set of randomly initialized weights and the output of the hidden layer neurons.

5) The error value is estimated between the output vector and the target vector

which is supplied to the output layer.

6) Based on the error values, the weights of the output layer neurons are

estimated initially followed by the weight adjustment of the hidden layer

neurons using Equation (5.3).

7) The entire training process is repeated till the error reaches the specified

threshold level and finally the stabilized set of weights are observed.

8) Next, an unknown image from the testing dataset is given as input and the

output is calculated using the stabilized set of weights. The unknown input is

categorized to the class represented by the neuron for which the output is

maximum.

9) The process is repeated for all the testing images and the performance

measures are calculated.

The sigmoid activation function is used throughout the implementation of BPN. The

error threshold specified for BPN is 0.01. The value of the learning rate is 0.7. The

binary representation is used for the target vector and the number of iterations

required to reach near the specified error value is roughly 2500.

5.5.2 IMPLEMENTATION OF RBF

The step-by-step procedure of the implementation is as follows:

1) The entire image dataset is divided into training dataset and testing dataset.

2) The features extracted from the images from each category are given as input

to the neural network.

Page 16: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

77

3) Initially, the stabilized set of mean and standard deviation values are

determined using Equation (5.4).

4) The outputs of the hidden layer neurons are calculated using these stabilized

values with the Gaussian basis function.

5) The error value is estimated and the weights are adjusted using Equation (5.9)

till the error is minimized.

6) Further, the stabilized set of weights, mean values and standard deviation

values are used for each testing image to determine the output. The testing

image is categorized to the class for which the corresponding neuron yields

the maximum output value.

The error value specified for RBF is 0.01 and the Gaussian function is used as the

basis function. The value of the learning rate is 0.7 and the binary representation

is used for the target vector. The number of iterations used to reach near the

specified error value is roughly 2000.

5.5.3 IMPLEMENTATION OF SOM

The step-by-step procedure of the implementation is as follows:

1) The entire image dataset is divided into training dataset and testing dataset.

2) The features extracted from the images from each category are given as input

to the neural network.

3) The winner neuron among the output layer neurons is determined using the

distance measure with the input values and the randomly initialized weight

vectors.

4) The weights of the winner neuron is adjusted using Equation (5.11)

5) The same procedure is repeated for specified number of iterations and the

stabilized set of weights is observed.

6) The testing process is further carried out using these weight vectors.

Page 17: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

78

The Euclidean distance is used as the measure to determine the winner neurons. Unit

radius is used to determine the neighbor neurons which show that the weights of the

winner neuron alone are adjusted.

The learning rate used for SOM is 0.7 and the number of iterations used is roughly

1600. No significant changes in the weight values are observed beyond 1600

iterations.

5.5.4 IMPLEMENTATION OF CPN

The step-by-step procedure of the implementation is as follows:

1) The entire image dataset is divided into training dataset and testing dataset.

2) The features extracted from the images from each category are given as input

to the neural network.

3) Initially, the stabilization of the Kohonen layer weights is performed using

Equation (5.12) with the ‘winner take-all’ training rule.

4) Further, the weights of the Grossberg layer are stabilized using Equation

(5.13) with the modified delta rule.

5) In CPN, weights are adjusted in the forward direction unlike BPN where the

weights are adjusted in the reverse direction.

The learning rate used for CPN is 0.7 and the unit radius methodology is followed

for training the Kohonen layer. The error threshold value for Grossberg layer is 0.01.

The sigmoid activation function is used for estimating the output of Grossberg layer

and the number of iterations required is roughly 2300.

5.5.5 IMPLEMENTATION OF FUZZY NEAREST NEIGHBOR CLASSIFIER

The step-by-step procedure of the implementation is as follows:

Page 18: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

79

1) The distance metrics between the input data and the randomly initialized

centroid values are estimated.

2) These distance metrics are used to calculate the membership values.

3) These membership values are further used to calculate the cluster centres.

4) Thus, the simultaneous adjustment of membership values and cluster centres

are performed for the specified number of iterations.

5) The final cluster centre values are stored for further investigation.

6) The same process is repeated for the testing input data and the distance

between the testing data centres and already stored centres are calculated.

7) The data is assigned to the class for which the distance is minimum.

In all the above networks, the various parameters are suitably initialized before

the training process. Normalization is also done to squash the range of the input and

output values for reducing the complexity of the implementation. The inputs from all

the four categories are sequentially applied to the input layer and hence the

individual stabilized weight matrices of these neural networks represent all the four

abnormal categories.

5.6 EXPERIMENTAL RESULTS and DISCUSSIONS

The above mentioned neural networks are experimented with the same input

dataset. The performance measures used to analyze these networks are classification

accuracy, sensitivity, and specificity, Positive Likelihood Ratio (PLR), Negative

Likelihood Ratio (NLR) and Convergence rate.

5.6.1 Performance Analysis of BPN

Page 19: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

80

One of the main difficulties with the BPN is to determine the number of

neurons in the hidden layer. Usually, the number of neurons is selected randomly but

in this work the number of neurons is chosen based on an experiment. The neural

network is trained with different number of hidden neurons and the training error is

estimated. The number of hidden neurons for which the training error is minimum is

fixed for further analysis of the network. The results are shown in Table 5.1.

Table 5.1 Analysis of BPN network with different ‘n’ hidden neurons

Class Training

data

Error (%)

(a) (b) (c) (d)

n=10 n=15 n=20 n=25

CNVM 40 8.28 9.12 2.05 7.17

CRVO 40 0.25 0.37 0.07 1.15

CSR 40 12.4 9.48 3.1 9.89

NPDR 40 1.34 1.37 0.04 0.98

Average 5.56 5.05 1.31 4.8

In this experiment, 40 images from each class are used during the training

process. Initially, 10 neurons are used in the hidden layer. The training error for each

class is determined and the average error is estimated. The same experiment is

repeated for the same BPN with 15 neurons, 20 neurons and 25 neurons in the

hidden layer. From Table 5.1, it is evident that the error gradually decreases for

increasing number of hidden neurons. The minimum error is obtained for the BPN

with 20 hidden neurons and once again an increase in the error rate is obtained for

higher number of neurons. Hence, the optimal value of 20 is chosen and used in this

work for further investigation.

The classification accuracy, sensitivity and specificity results are estimated with the

help of True Positives (TP), True Negatives (TN), False Positives (FP) and False

Negatives (FN). These TP, TN, FP, FN values correspond to the number of wrongly

Page 20: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

81

diagnosed images and correctly diagnosed images. The values of these parameters

are determined using the confusion matrix given in Table 5.2.

Table 5.2 Confusion Matrix for BPN

Category Class1 Class 2 Class 3 Class 4

CNVM 58 1 3 2

CRVO 1 55 2 2

CSR 3 2 65 3

NPDR 1 3 2 77 Class 1=CNVM; Class 2=CRVO; Class 3 = CSR; Class 4=NPDR

The testing images alone are used to measure the performance of the BPN

classifier. Among the 64 testing images of CNVM category, 58 images are correctly

classified images and the rest are the wrongly classified images. Thus, the correlation

between the category and the class which it belongs to is extremely high which is

evident from the diagonal elements of the matrix. From this matrix, the values of TP,

TN, FP, FN are calculated which is essential for performance measure analysis. The

performance measure analysis of the BPN is shown in Table 5.3.

Table 5.3 Performance Measures of the BPN

TP TN FP FN Sensitivity Specificity Accuracy (%) PLR NLR

CNVM 58 211 5 6 0.90 0.98 96 45 0.1

CRVO 55 214 6 5 0.92 0.97 96 30.66 0.08

CSR 65 200 7 8 0.89 0.96 94 22.25 0.11

NPDR 77 200 7 6 0.93 0.96 95 23.25 0.07

Average Value

0.91 0.97 95 30.29 0.09

From the above results, it is evident that the classification accuracy of BPN is

sufficiently high for practical applications. Also, the ability of the BPN to distinguish

between positive and negative results is also high which is evident from the

sensitivity and specificity values. The PLR value is lesser than the ideal value (100)

and the NLR is highly closer to the ideal value (0). But, any value of PLR> 10 and

NLR<1 is acceptable for practical applications. Thus, it is evident that this network is

able to respond much better to the negative results than the positive results. For

Page 21: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

82

example, the BPN is better in categorizing the non-CNVM image to any of the non-

CNVM categories than the CNVM image to the CNVM category. Nevertheless, the

results obtained from the BPN are of high quality for possible application in the

medical field.

5.6.2 Performance Analysis of RBF

The number of hidden neurons used in RBF is also 20 to maintain uniformity

among all the neural networks. Initially, the confusion matrix for this neural network

is determined to estimate the values of TP, TN, FP and FN. The matrix between the

different categories and their corresponding classes is named as confusion matrix.

The number of correctly classified images for each category is estimated from this

matrix. The confusion matrix of SOM is shown in Table 5.4.

Table 5.4 Confusion Matrix of RBF

Class1 Class 2 Class 3 Class 4

CNVM 53 3 4 4

CRVO 3 50 3 4

CSR 4 5 59 5

NPDR 2 5 5 71

From Table 5.4, it is clearly evident that the level of misclassification rate has

been increased over the results of the BPN. Using this table, the values of TP, FP,

TN and FN are determined using the procedure detailed in section 6.1. The values of

these parameters and the performance measures which are calculated using these

values are displayed in Table 5.5.

Table 5.5 Performance Analysis of RBF

TP TN FP FN Sensitivity Specificity Accuracy

(%)

PLR NLR

CNVM 53 207 9 11 0.83 0.95 93 17 0.18

CRVO 50 207 13 10 0.84 0.94 92 14 0.17

CSR 59 195 12 14 0.81 0.94 91 14 0.20

NPDR 71 184 13 12 0.86 0.93 91 12 0.15

Average Value 0.83 0.94 92 14 0.18

Page 22: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

83

The ability of the RBF for pattern recognition is proved from the experimental

results shown in Table 5.5. The classification accuracy is sufficiently high because of

the supervised nature of the training methodology. The capability of this neural

network in yielding accurate positive results (sensitivity) and negative results

(specificity) is also shown in the above table. The PLR and the NLR values are also

sufficiently effective to show the superior nature of the proposed method. The time

taken for convergence of RBF is also similar to the BPN which is calculated from the

number of iterations used during the training process.

5.6.3 Performance Analysis of SOM

The difficulty of designing the architecture of SOM is completely minimized

since there are no hidden layers for this neural network. The performance measures

of this neural network are determined using the confusion matrix. The false

classification rate and the correct classification rate are determined using this matrix.

The confusion matrix of SOM is shown in Table 5.6.

Table 5.6 Confusion Matrix of SOM

Class1 Class 2 Class 3 Class 4

CNVM 40 5 8 11

CRVO 8 39 7 6

CSR 7 11 47 8

NPDR 6 8 14 55

The number of correctly classified images has been significantly reduced for

SOM when compared with the RBF and BPN networks. From Table 5.6, the values

of TP, FP, TN and FN are determined using the procedure detailed in section 6.1.

The values of these parameters and the performance measures which are calculated

using these values are displayed in Table 5.7.

Page 23: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

84

Table 5.7 Performance Measures of the SOM

TP TN FP FN Sensitivity Specificity Accuracy

(%)

PLR NLR

CNVM 40 195 21 24 0.63 0.90 84 6.3 0.40

CRVO 39 196 24 21 0.65 0.89 84 5.9 0.39

CSR 47 178 29 26 0.64 0.85 80 4.2 0.42

NPDR 55 172 25 28 0.66 0.87 81 5.0 0.39

Average Value 0.65 0.88 82 5.3 0.40

The inferior nature of the SOM over the BPN is verified from the

experimental results shown in Table 5.7. The classification accuracy is lesser than

BPN which shows the inability of the SOM in correctly categorizing the images. The

sensitivity value is also low which is mainly due to the increase in the FN. This

condition is undesirable which mainly limits the usage of SOM for practical

applications. The other performance measures are also inferior to the results of BPN

which is evident from the results. The main reason behind the inferior results of

SOM is the lack of target vectors which ultimately reduces the quality of the results.

But, the time taken for convergence is much better than the BPN which is estimated

from the number of iterations. Thus, these networks are mostly used for applications

where convergence time period is more significant than the accuracy.

5.6.4 Performance Analysis of CPN

The number of hidden layer neurons used in this work for CPN is also 20 to

maintain uniformity among all the approaches. The confusion matrix of the CPN is

given in Table 5.8.

Table 5.8 Confusion Matrix of CPN

Class1 Class 2 Class 3 Class 4

CNVM 47 4 6 7

CRVO 5 44 5 6

CSR 6 8 53 6

NPDR 4 7 8 64

Page 24: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

85

From Table 5.8, it is evident that the rate of misclassification is highly

reduced over the SOM but still inferior to BPN and RBF. The performance measure

analysis of CPN is shown in Table 5.9.

Table 5.9 Performance Measures of the CPN

TP TN FP FN Sensitivity Specificity Accuracy

(%)

PLR NLR

CNVM 47 201 15 17 0.74 0.93 88 11 0.28

CRVO 44 201 19 16 0.73 0.91 87 8 0.30

CSR 53 188 19 20 0.72 0.90 86 7.2 0.32

NPDR 64 178 19 19 0.77 0.90 86 7.7 0.26

Average Value 0.74 0.91 87 8.5 0.29

The classification accuracy results of CPN are better than the SOM which is

illustrated on Table 5.9. The inclusion of the supervised training methodology has

improved the accuracy to high extent. Also, the sensitivity and specificity values are

comparatively better than SOM. There is also scope for improvement since the

response to the positive results (sensitivity) is not highly encouraging. The likelihood

ratios are also better than the unsupervised neural network. The convergence rate is

also approximately equivalent to the supervised neural networks. Thus, CPN has

yielded results which are better than the SOM but not highly sufficient for practical

applications.

5.6.5 Performance Analysis of fuzzy nearest neighbor classifier

The confusion matrix of conventional fuzzy classifier is shown in Table 5.10.

Table 5.10 Confusion Matrix of Fuzzy nearest neighbor classifier

Class1 Class 2 Class 3 Class 4

CNVM 35 6 10 13

CRVO 9 35 10 6

CSR 11 14 40 8

NPDR 7 11 17 48

Page 25: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

86

The level of correct classification rate has been declined sharply in

comparison to the conventional neural classifiers. The performance measures are

shown in Table 5.11.

Table 5.11 Performance Measures of the fuzzy nearest neighbor classifier

TP TN FP FN Sensitivity Specificity Accuracy

(%)

PLR NLR

CNVM 35 189 27 29 0.55 0.87 80 4.4 0.52

CRVO 35 189 31 25 0.58 0.86 80 4.1 0.49

CSR 40 170 37 33 0.55 0.82 75 3.1 0.55

NPDR 48 170 27 35 0.58 0.86 78 4.2 0.49

Average Value 0.57 0.85 78 3.8 0.51

The inferior results of the fuzzy classifier are evident from Table 5.11. The

low quality results are due to two important factors. This approach is purely based on

clustering and since the abnormality is wide spread in the image, the probability of

achieving accurate centroid values is low. The incorrect centroid values have resulted

in the inferior results for the fuzzy classifier.

5.6.6 Comparative Analysis of the Classifiers

A comparative analysis between the neural classifiers in terms of the performance

measures is shown in Table 5.12.

Table 5.12 Comparative analysis of the classifiers

Classifiers Average CA

(%)

Average

Sensitivity

Average

Specificity

PLR NLR No. of iterations

used

BPN 95 0.91 0.97 30 0.09 2500

RBF 92 0.83 0.94 14 0.18 2000

SOM 82 0.65 0.88 5.3 0.40 1600

CPN 87 0.74 0.91 8.5 0.29 2300

Fuzzy classifier 78 0.57 0.85 3.8 0.51 2500

The exploration of application of conventional neural classifiers and fuzzy

classifier for pattern recognition is performed in this work. Two supervised neural

classifiers, a single unsupervised neural classifier and a single hybrid classifier are

Page 26: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

87

used for the experiments. A fuzzy classifier is also used in the experiments. Thus,

this work has explored the ability of different types of available AI techniques.

The main performance measure of these experiments is the classification accuracy.

Higher classification accuracy is always desirable and the experimental results have

shown that the ability of BPN and RBF to classify the images is better than the SOM

and CPN. This increase in the accuracy is mainly due to the presence of target

vectors in BPN and RBF. Among RBF and BPN, the accuracy of BPN is slightly

higher than RBF since the number of adjustable parameters is more for RBF than the

BPN. Nevertheless, the accuracy of both these networks is almost equivalent. Even

though, CPN includes the target vectors, the presence of unsupervised training has

reduced the quality of the results. One of the weight matrices is not accurate which

has lead to inferior results. On the other hand, SOM has yielded the lowest quality

results among the four neural classifiers. The lack of standard convergence condition

is the main reason for the inferior results of SOM.

Sensitivity is the ability of the classifier to correctly identify the ‘abnormal’

objects (positive results) and specificity is the capability of the classifier to correctly

identify the ‘normal’ objects (negative results). It may be noted that the category in

question is treated as ‘abnormal’ and all other three categories are treated as

‘normal’. Higher value of sensitivity and specificity are desirable and the ideal value

for both the cases is 1. The experimental results are interesting due to the fact that all

the four networks responded better in identifying the negative results than the

positive results. Among the four neural classifiers, the sensitivity and specificity

results of BPN are encouraging since it responded efficiently to both the positive and

negative results. Since the number of FN images is low for BPN, the sensitivity and

specificity results are much higher than other classifiers. The results of SOM are the

most inferior among all the neural classifiers since the accuracy is very low.

PLR is defined as the probability of a person who has the disease testing

positive divided by the probability of a person who does not have the disease testing

Page 27: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

88

positive. NLR is defined as the probability of a person who has the disease testing

negative divided by the probability of a person who does not have the disease testing

negative. Higher value of PLR and lower value of NLR is always desirable for any

diagnostic applications. The experimental results have suggested favorable results for

BPN and RBF in terms of PLR and NLR whereas the PLR and NLR values of SOM

and CPN are not optimal for applications. Thus, these PLR and NLR values are

indicators of the ‘reassurance’ about the performance of the classifiers. These values

also indirectly have provided some ideology about the robustness of the system.

On the other hand, the performance of the fuzzy classifier is highly inferior to

the neural classifiers. The difficulty in clustering the abnormal retinal image is the

main reason for the inferior results. Since, the abnormality is wide spread in the

image, clustering the defective region into single group is practically non-feasible.

Since the classification process is based on the clustered centroid values, the

accuracy is much lower than the other neural classifiers. The number of parameters

to be initialized is also higher for fuzzy classifiers which are another reason for the

inferior results. Incorrect initialization of these parameters has yielded low accuracy

results.

Finally, an analysis on the convergence rate of the classifiers is done in this

work. This analysis is based on the number of iteration used since the iterations are

directly proportional to the convergence time period. Experimental results have

suggested that SOM is much superior to other networks since the time requirement

for convergence is very low. Even though the number of iterations is increased

beyond 1600 for SOM, there is no significant improvement in the accuracy. On the

other hand, time period requirement of BPN is high because the number of iterations

to achieve the minimum error value is high. Another reason for the increased time

period of BPN is that it is a multilayer network whereas SOM and RBF are single

layer networks. The time period of CPN is higher than SOM but lesser than BPN

since the training methodology lies in between these two extremes. The time period

Page 28: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

89

requirement for fuzzy classifier is also much higher for practical applications. Since

accuracy is more important than convergence rate, BPN is preferred for diagnostic

applications. These networks are executed on Pentium processor with system speed

of 1.66 GHz and 2 GB RAM.

5.7 Conclusion

In this work, four different neural networks and a single fuzzy classifier are

used for abnormal retinal image classification. The networks are tested on 420 real-

time images collected from ophthalmologists. The performance measures are

analyzed in terms of classification accuracy, sensitivity, specificity, PLR and NLR.

A brief convergence rate analysis is also performed in this work. Experimental

results suggested promising results for the supervised neural networks such as BPN

and RBF in terms of the accuracy measures. The increase in the accuracy is mainly

due to the availability of standard convergence condition such as minimization of the

error value. On the other hand, the accuracy of SOM is very less since the

unsupervised training methodology is adopted in this network.

A hybrid neural network with both training methodologies is also tested in

this work with an objective to enhance the performance of the automated system.

Even though the accuracy of CPN is better than the SOM, it is inferior to supervised

neural systems in terms of accuracy. The fuzzy classifier is not capable of

recognizing the patterns unlike the neural classifiers. But, contrasting results has

been obtained when the classifiers are analyzed in terms of convergence time period.

The time period requirement of SOM is very much lesser than the other neural

networks and the fuzzy classifier. The main reason behind this superior convergence

rate is twofold: (a) Number of iterations is less and (b) single layer network which

involves one set of weight matrix adjustments.

Page 29: CHAPTER 5 CONVENTIONAL AI BASED IMAGE CLASSIFICATION ...shodhganga.inflibnet.ac.in/bitstream/10603/10111/11... · of various AI techniques for abnormal retinal image classification.

90

Thus, this work has suggested the usage of supervised neural networks for

applications where accuracy is important and unsupervised neural networks where

the convergence rate is significant. Also, there is scope for performance

improvement of these AI techniques. The accuracy of all the AI classifiers can be

improved by performing some changes in the input feature set since the training data

set is significant for the performance of these methodologies.