VISVESVARAYA TECHNOLOGICAL UNIVERSITY

JNANA SANGAMA, BELAGAVI- 590 018

A Seminar Report
On

“The skin cancer classification using deep convolutional neural

network”

Submitted in partial fulfillment of the requirements for the VIII Semester degree of Bachelor
of Engineering in Computer Science and Engineering
Of Visvesvaraya Technological University, Belagavi

Submitted by

KARTHIK D 1RN15CS049

Under the Guidance of:

Mr. Anjan Kumar K N
Asst. Professor
Dept. of CSE

Department of Computer Science and Engineering

(NBA Accredited for academic years 2018-19, 2019-20, 2020-21)
RNS Institute of Technology
Channasandra, Dr. Vishnuvardhan Road, Bengaluru -560 098

2018-2019
 RNS Institute of Technology
Channasandra, Dr. Vishnuvardhan Road, Bengaluru-98

DEPARTMENT OF COMPUTER SCIENCE & ENGINEERING

(NBA Accredited for academic years 2018-19, 2019-20, 2020-21)

CERTIFICATE

Certified that the Seminar topic on “The skin cancer classification using deep
convolutional neural network ” has been successfully presented at RNS Institute of
Technology by Karthik D , bearing USN 1RN15CS049 , in partial fulfillment of the
requirements for the VIII Semester degree of Bachelor of Engineering in Computer Science
and Engineering of Visvesvaraya Technological University, Belgaum during academic year
2018-2019. It is certified that all corrections/suggestions indicated for Internal Assessment
have been incorporated in the report deposited in the departmental library. The Seminar report
has been approved as it satisfies the academic requirements in respect of Seminar work for
the said degree.

Mr. Anjan Kumar K N Mr. Devaraju B M

Asst. Professor, Guide Asst. Prof., Coordinator

Dr. G T Raju Dr. M K Venkatesha

Vice Principal, Principal
Professor & HOD
 ABSTRACT

Dermatology is one of the most unpredictable and difficult terrains to diagnose due its
complexity. In the field of dermatology, many a times extensive tests are to be carried out so
as to decide upon the skin condition the patient may be facing. The time may vary from
practitioner to practitioner. This is also based on the experience of that person too. So, there
is a need of a system which can diagnose the skin diseases without any of these constraints.

The System addresses the demand for an intelligent and rapid classification system of
skin cancer using contemporary highly-efficient deep convolutional neural network. The
Main focus is on the task of classifying the skin cancer using ECOC SVM, and deep
convolutional neural network. RGB images of the skin cancers are collected from the
Internet. Some collected images have noises such as other organs, and tools. These images
are cropped to reduce the noise for better results. In this paper, an existing, and pre-trained
AlexNet convolutional neural network model is used in extracting features. A ECOC SVM
classifier is utilized in classification the skin cancer. The results are obtained by executing a
proposed algorithm with a total of 3753 images, which include four kinds of skin cancers
images. The implementation result shows that maximum values of the average accuracy,
sensitivity, and specificity are 95.1 (squamous cell carcinoma), 98.9 (actinic keratosis), 94.17
(squamous cell carcinoma), respectively. Minimum values of the average in these measures
are 91.8 (basal cell carcinoma), 96.9 (Squamous cell carcinoma), and 90.74 (melanoma).

I
 ACKNOWLEDGMENT

At the very onset, I would like to place on record our gratefulness to all those people who
have helped us in making this Technical seminar work a reality. Our Institution has played a
paramount role in guiding us in the right direction.

I would like to profoundly thank the Management of RNS Institute of Technology for
providing such a healthy environment for the successful completion of this seminar work.

I would like to express my sincere thanks to our respected Director, Dr. H N Shivashankar
for his constant encouragement that motivated me for the successful completion of this work.

I would also like to thank our beloved Principal, Dr. M K Venkatesha, for providing the
necessary facilities to carry out this work.

I am extremely grateful to our beloved Vice–Principal and HoD-CSE, Dr. G T Raju, for
having accepted to patronize me in the right direction with all his wisdom.

I place my heartfelt thanks to all the Coordinators of seminar work. I would like to thank the
internal guide Mr. Anjan Kumar K N, Asst. Professor, for his continuous guidance and
constructive suggestions for this work.

Last but not the least, I am thankful to all the staff members of Computer Science and
Engineering Department for their encouragement and support throughout this work.

KARTHI
KD
1RN15CS049

II
 CONTENTS

Sl. No. Chapters Page No.

Abstract I

Acknowledgment II

Contents III

List of Figures IV

List of Tables V

1 Introduction 01

2 Literature Survey 03

3 Approach 05

3.1 Convolution neural networks 05

3.2 AlexNet Architecture 09

4 Methodology 12

4.1 Acquire Image Data 12

4.2 Classification in groups using convolution neural network 13

5 Results and Discussion 14

6 Conclusion and Future works 18

References

III
III
 LIST OF FIGURES

Fig. No Description Page No.

3.1 Convolution layer 05

3.2 Output image depth defined by filters 06

3.3 Running filters with size 2 x 2 x d 07

3.4 2 x 2 Max-pooling with stride 2 08

3.5 AlexNet Architecture 09

5.1 Skin cancer sample images which are used for classification 14

IV
IV
Table No. Description Page No.

5.1 Number of training and testing samples of skin cancer 15

5.2 Implementation results of the skin cancer classification 16

5.3 Accuracy, sensitivity and specificity of skin cancer classification 17

Implementation result comparison between previous research

5.4 17
works and the proposed study

LIST OF TABLES

V
V
CHAPTER 1

INTRODUCTION

Recently, cancer patients are increasing due to the factors such as lifestyle, smoking
tobacco, alcohol usage, diet, physical activity, environmental change, sun and other types of
radiation, viruses and so on. The most common type cancer includes skin cancer. Unusual
swellings of cells for skin can be the skin cancer. There are four kinds of skin cancers: Actinic
Keratosis (AK), Basal cell carcinoma (BCC), Squamous cell carcinoma (SCC), and
Melanoma. Early diagnosis of the cancer helps to treat it successfully. The late diagnosis of
the cancer causes the cancer spread to other nearby organs, and cannot be treat. The
classification of pigmented skin lesions with unaided eye is challenging, even for the highly
experienced dermatologists. So, dermoscopy is used for visual inspection of the skin lesions
in a better way. This device can magnify the inspected regions as well as eliminates the
surface reflection of the skin which leads to improve the diagnostic accuracy. But in several
occasions using dermatoscope a trained experts also fail to make correct prediction. Hence,
several computer assisted automated approaches are proposed to analysis dermoscopy images
.The identification of skin disease from dermoscopy images are treated as an image
classification problem. The tradition approach of image classification needs robust feature
representation which are feed to the classifier for training. So, inspired by the medical
diagnostic procedure several color, texture and shape features are used to characterize the
skin lesion. However, it is very much difficult to develop robust feature representation to deal
with the dermoscopy images obtained from different acquisition devices and captured in
diverse illumination conditions.

There are numerous publications are composed in detecting, segmenting, and classifying
skin cancers employing different computer vision, machine learning, image processing,
neural network, and classification techniques. Esteva et al. developed the skin cancer
classification using convolutional neural network. Hitoshi et al. presented a quite automatic
system for the classification of the melanomas. While Anas et al. composed the melanoma
classification by four kinds classification, Almansour et al. demonstrated a classification
method for melanoma using the k-means clustering, and Support Vector Machine (SVM).
Abbas et al., and Capdehourat et al. composed skin lesions, cancers, and dermoscopy image
classification methods using AdaBoost MC, separately. Giotis et al., and Ruiz et al. developed

Dept. of CSE, RNSIT 2018-19 1

CHAPTER 1 INTRODUCTION

decision support systems utilizing image processing and neural network algorithms by lesion
texture, color, visual diagnostic attributes, and affected area, degree of damage for the
melanoma, respectively. Also, Isasi et al. presented an automatic melanomas diagnosis
system. The cutaneous melanoma is diagnosed by Blum et al. Kiran et al. developed an
Android application for the melanoma classification.

As a subfield of the deep neural network, convolutional neural networks has recently
achieved great success in machine learning and computer vision. In 2012, Krizhevsky et al.
presented a novel method using deep convolutional neural network to classify large (1.2
million high-resolution) images in the ImageNet LSVRC-2010 at the Neural Information
Processing Systems Conference. It is a first-mover advantage of the convolutional neural
network in computer vision. The AlexNet model has some benefits, i.e., preeminent
implementation, few training parameters and extreme validity. Also, the major superiority is
the network pre-trained on the ImageNet dataset. Additionally, the input images are clearly
recognizable from anything in the ImageNet network. A majority of the image is black
because the image is being viewed through an endoscope. The actual content of the image is
generally tissue, with the tool being the contrasting object from the background. This poses a
trickier problem given that the majority of the images contain a repetitive background with
some type of surgical tool which tends to bear the same color and relative shape. Cireşan et
al. applied multi-column deep neural networks for image classification. Several publications
are demonstrated in multiclass classification. Amornsamankul et al. presented combining
hybrid approaches, i.e., Complementary Neural Networks (CMTNN) and Error-Correcting
Output Codes (ECOC) in solution multiclass classification problem. While Wiharto et al.
demonstrated performance analysis in classification of multiclass Support Vector Macine
(SVM) for coronary heart disease diagnosis, Bagheri et al. composed ECOC in multiclass
classification. Yan et al. analyzed a performance in ECOC SVMs, and also applied ECOC
SVM in remote sensing image classification. Liu et al. developed a method in speech
recognition using SVM and ECOC.

Dept. of CSE, RNSIT 2018-19 2

CHAPTER 2

LITERATURE SURVEY

a) An automated system for recognizing disease conditions of human skin .In this model, the
condition of the skin disease is identified by evaluating skin disease images by using grey
normalized symmetrical simultaneous occurrence stencils (GLCM) method. The proposed
system is used in an efficient and economical for the automatic recognition of skin diseases.
This system is useful for the skin to reduce the error with medical diagnosis. Another is the
first test for patients in rural areas, where the good doctors are missing. The system works
with relational databases to the storage of implying the need for textual skin images. This
system can also work for same type of images directly over feature vector.

b) Image-based diagnosis method: This system mainly focuses on diagnosing diseases of skin
that are occurred by viruses and bacteria. This system used image of the diverse area and
those images are taken and then machine learning techniques and image processing applied to
train the computer to diagnose the skin disease. This is an optional diagnosis method for these
skin diseases and it is safe and no risks, side effects or inconveniences from the patient
perspective. It also gives advantageous to doctors because it is fast and can be implemented in
various ways (mobile phones, computers and digital cameras). And also it can be safely used
by non-Specialized medical personnel. First, the patients were clinically analyzed by a
professional (dermatologist/medical doctor), then laboratory tests were conducted to foresee
and confirm the skin disorder. The doctor then apprehended some images from the patients
whose results showed that they had a viral or bacterial infection.

c) Expert System for Diagnosis of Skin Diseases: This system is developed for diagnosing
skin diseases which allow user to identify diseases of the human skin to provide advises or
medical treatments in a very short time period. The system uses technologies such as image
processing and data mining for the diagnosis of the disease of the skin. The image of skin
disease is taken and it must be subjected to various processing for noise eliminating and
enhancement of image. This image is immediately segmentation of images using threshold

Dept. of CSE, RNSIT 2018-19 3

CHAPTER 2 LITERATURE SURVEY

values. Finally data mining techniques are used to identify the skin disease and to suggest
medical treatments or advice for users.

d) The Development of Online Children Skin Diseases Diagnosis System: A system enables
the user recognize skin diseases confronted of children through the Internet and make user for
advice or Treatments in the shortest period of time. This is based on law and the ahead was
used a sequence heuristics engine for development from the system. With this system, to
assist and allows the user to Recognition of Pediatric. Dermatology through the Internet and
offer helpful Proposal the user.

e) High Performance Convolutional Neural Networks for Image Classification: A system

which offers a fast, fully parameterizable GPU implementation of Convolutional Neural
Network variants. The feature extractors are neither carefully designed nor pre-wired, but
rather learned in a supervised way. The deep hierarchical architectures achieve the best
published results on benchmarks for object classification (NORB, CIFAR10) and handwritten
digit recognition (MNIST), with error rates of 2.53%, 19.51%, and 0.35%, respectively. Deep
nets trained by simple back-propagation perform better than more shallow ones. Learning is
surprisingly rapid. NORB is completely trained within five epochs. Test error rates on MNIST
drop to 2.42%, 0.97% and 0.48% after 1, 3 and 17 epochs, respectively.

f) Classification of Dermoscopic Skin Cancer Images Using Color and Hybrid Texture
Features: The proposed system is a new feature extraction technique for classification of
dermoscopic images into melanoma and non-melanoma. Two types of features have been
used, color and texture. For texture features, GLCM and LBP have been used. Combining
these features improves the accuracy of the classification results. In this way, our proposed
technique has been able to better classify dermoscopic images into Melanoma and Non-
Melanoma groups. In order to evaluate the usefulness and performance of proposed model,
experimentation is performed on standard dataset of dermIS.

Dept. of CSE, RNSIT 2018-19 4

CHAPTER 3
APPROACH

3.1 Convolution neural networks

Convolutional neural network is an effective machine learning technique from the deep
learning and it is similar to ordinary Neural Networks. Convolutional neural network is a
network with convolutional layers. Convolutional neural network is consists of three steps of
neural layers to build its architectures: Convolutional, Pooling, and Fully-Connected.

Convolutional layers
The convolutional layer is the main part of the convolutional neural network. In the
convolutional layer the output result is derived from the input by filtering in certain
conditions. Convolutional layers consist of a rectangular grid or cubic block of neurons. It
means input, output layers with filters can be a rectangular grid or cubic block of neurons
(See Fig. 3.1). The filter is supplied from uppermost left to downside right. In every location
of the filter, the weighted volume of the pixels is determined by W TX + b and a new neuron is
acquired. In the convolutional layer three kind of hyper parameters determine the volume of
the output neurons: depth, stride, and zero-padding.

Fig. 3.1 - Convolutional layer. Input size is 4x5xd pixels (height, width, depth). Filter (kernel) size is
2 × 2 pixels.

A number of the filters with certain strides determines the depth as shown in Fig. 3.2. For
original RGB image depth is equal to three.
Dept. of CSE, RNSIT 2018-19 5
CHAPTER 3 APPROACH

Fig. 3.2 - Output Image depth defined by filters

The stride specifies the step of running filters by pixels in the input neuron. For better
understanding Fig. 3.3 illustrates running of the filters with size 2x2xd in input neuron with
size 5x7xd by stride 1, 2 and 3.

Here, whereas the stride equal to 1, the filters runs a single pixel. In this case output
neuron is 5x5xd. Whereas the stride equal to 2, the filters are skipped two pixels and output
neuron size is 3x3xd. In the last case, when stride is 3 the filters are skipped 3 pixels. Output
neuron size is 2x2xd.

Zero padding is the filling process in the input neuron where zeros are around the edge.
Zero-padding is mostly consumed for adjustment the size of the input neuron for research
purposes. It is normally applied for when the input neuron size needs to maintain in the output
neuron. Figure 4 is presented examples in zero-padding of different sizes.
The output neuron size is calculated by Eq.3.1.
Dept. of CSE, RNSIT 2018-19 6
CHAPTER 3 APPROACH

OUTPUT = ( (W – K + 2P) / S ) + 1…………………………………. 3.1

Where W is the input neuron’s size, K is the filter (kernel) size, P is size of the padding, and S
is the stride.

In the convolutional neural network linear algebraic operations are also used. Suppose that
matrix dimensions are m and n (m rows, n columns). 2D convolutional cube calculates the
two-dimensional convolution with two input matrices. (MA, NA) is dimensions of the matrix
A, and (MB, NB) is dimensions for the matrix B. In case, the cube determines the complete
output size, convolution equation shown in Eq 3.2.

C (i, j )  m0A 
( M 1) ( N A 1)
n 0
A(m, n) * B (i  m, j  n) ………………...3.2

where 0 ≤ I < MA+ MB − 1 and 0 ≤ j < NA + NB − 1.

Dept. of CSE, RNSIT 2018-19 7

CHAPTER 3 APPROACH

Fig. 3.3 – Example in the running filters with size 2x2xd.

Max-pooling layers
After each convolutional layer, a pooling layer executes next operation. The pooling layers

are utilized to decrease size of the input neuron. This layer acquires small rectangular blocks
from the convolutional layer and samples it to provide a single output from that block. The
max pooling method is common used. A formulation for a single type of the pooling
layer, maxpooling is presented in Eq. 3.3.

Dept. of CSE, RNSIT 2018-19 8

CHAPTER 3 APPROACH

……..……………………… 3.3

The pooling is created by sampling filter in each layer applying the filters. Figure 5 is
shown an example in max-
pooling for better understanding.
Here, input layer size is 4 × 4,
filter size is 2 × 2, and stride
equal to 2.

Fig. 3.4 - 2 × 2 Filter

Max-pooling, stride is
2

Fully Connected Layers

Fully connected layer is the last layer, which is constructed from connection of all previous
neurons. The fully connected layer normally promotes reduction of spatial information
because it “fully connected” from all input neurons until all output neurons.

3.2 AlexNet Architecture

Dept. of CSE, RNSIT 2018-19 9

CHAPTER 3 APPROACH

AlexNet is the winning solution of IMAGENET Challenge 2012. This is one of the
most reputed computer vision challenge and 2012 was the first time that a deep learning
network was used for solving this problem. This convolutional neural network model include
five convolutional and three fully connected layers as shown in Fig.3.5. Three various filter
sizes composed in various convolutional layers, 11 × 11, 5 × 5, and 3 × 3. In this study
pooling layers will always be maxpooling layers.

Fig. 3.5 - AlexNet architecture

A total of eleven layers are present from input until output in this convolutional neural
network. Each layer’s output nodes could be the next layer’s input. Detailed explanation are
as below:
Dept. of CSE, RNSIT 2018-19 10
CHAPTER 3 APPROACH

Input layer (IL): The size of the skin cancer’s raw RGB image is 227 × 227 × 3. The raw
RGB image is extracted into three channels, i.e. red, green, and blue. The feature extraction is
composed randomly 227 × 227 × 3 parts in the whole area of the input image for each RGB
channel respectively.

Layer 1 (CL): This layer is the first feature extraction layer of our network. 96 filters with
size 11 × 11 include in this layer, and padding 0 are generated the feature maps by stride 4.
Accordingly eq. 1, output of this convolutional layer is (227–11) / 4 + 1 = 55. Then output
size is 55 × 55 × 96. It means 96 feature maps with size 55 × 55 are input of the next layer.

Layer 2 (MPL): The first Max-Pooling layer, which is used 3 × 3 filter, and stride 2. The
output size of this layer is calculated by (55–3) / 2 + 1 = 27. Depth is same as layer 1.
Then we can get 96 feature maps with size 27 × 27.

Layer 3 (CL): The second convolution layer with 256 filters to convolution. The filter size is
5 × 5, stride is 1, and zero padding size is 2. Because of zero padding of (5– 1)/2 = 2, the
original size is restored. Depth is 256 because of 256 filters. Thus size is 27 × 27 × 256.

Layer 4 (MPL): The second Max-Pooling with 3 × 3 filter, stride 2. Size of the output is
computed as (27–3) / 2 + 1 = 13. Depth is same as previous layer, i.e. 256 because pooling is
composed independently on each layer. Thus size of this pooling layer is 13 × 13 × 256. Next
three layers are convolution layers in this network. Even though these layers run by similar
principle to previous two layers, size and depth are revolved. The layer 5, and layer 6 are
same.

Layer 5 (CL): The 3rd convolution layer, which have 384 filters with size 3 × 3, stride 1, and
padding 1. The original size is restored by reason of zero padding of (3–1) / 2 = 1. Depth is
384 because of 384 filters. Thus 384 feature maps are produced with size 13 × 13 × 384 in
this layer.

Layer 6 (CL): The 4th convolution layer. This layer similar to layer 5. It have 384 filters, size
3 × 3, stride 1, and padding 1. Size: 13 × 13 × 384.
Layer 7 (CL): This is the last convolution layer of this study. It has 256 filters with size 3 ×
3, stride 1, and padding 1. The original size is also restored by reason of zero padding of (3–
1)/2 = 1. Depth is 256. Here, 256 feature maps are generated with size 13 × 13 × 256.

Dept. of CSE, RNSIT 2018-19 11

CHAPTER 3 APPROACH

Layer 8 (MPL): This is the last Max-Pooling layer of this network. The Max-Pooling layer
has 3 × 3 filter with stride 2. (13–3) / 2 + 1 = 6 is size of the output. Depth is same as before,
i.e. 256 because pooling is composed independently on each layer. Therefore 256 feature
maps are composed with size 6 × 6 × 256. The last three layers are fully connected layers in
this network.

Layer 9 (FCL): Fully connected with 4096 neuron. In this layer, each of the 6 x 6 x 256 =
9216 pixels are fed into each of the 4096 neurons and weights determined by back
propagation.

Layer 10 (FCL): Fully connected with 4096 neuron. Similar to layer #9.

Layer 11 (FCL): At the end we can obtain 1000 fully connected neurons.

Dept. of CSE, RNSIT 2018-19 12

CHAPTER 4
METHODOLOGY

The Main focus on the task of classifying skin cancers using deep convolutional
neural network. The hardware used includes: Intel Core i7 with 2.20GHz, 8.00GB Dual-
Channel DDR3 @ 665 MHz (9–9–9-24), 2047 MB NVIDIA GeForce GT 540 M Graphics
card. Software: Mat lab R2016b program is used to analyse images in this research.

An algorithm to obtain the overall methodology used is as follows:

 Acquire Image Data:

 Collect Image Data
 Crop Images
 Prepare Training and Test Image Sets

 Classification in groups utilizing convolutional neural network features:

 Extract Training features using convolutional neural network
 Train, and test a ECOC-SVM classifier using convolutional neural network
features

4.1 Acquire Image Data

Data set in this study was originated from related Internet sites. In this study we used
skin cancer RGB images with 500–1000 pixels, and jpg, tiff types. Some collected images
have noises, such as other organs, or tools. Then those images are cropped to reduce noises
for better result. Images are grouped by four kinds of skin cancers (Actinic Keratosis, Basal
cell carcinoma, Squamous cell carcinoma, and Melanoma), and saved in different folders for
training and testing process.

Dept. of CSE, RNSIT 2018-19 12

CHAPTER 4 METHODOLOGY

4.2 Classification in groups utilizing convolution neural

network features

The acquired images are extracted by deep convolutional neural network, and obtained
convolutional neural network features are classified into four groups, i.e., Actinic Keratosis,
Basal cell carcinoma, Squamous cell carcinoma, and Melanoma, using ECOC SVM. In this
an existing, and pre-trained AlexNet convolutional neural network model is used in extract
training features.

Next step is training and testing process in the classification for the skin cancers.
ECOC SVM classifier is utilized in the classification skin cancers. Combination of the SVMs,
and ECOC has produced the ECOC SVM which is used to classify multiple classes. Error
Correcting Output Codes or ECOC method transforms a multi-class classification problem
into two-class classification problems.

Dietterich et al. at first presented the ECOC in 1995, and Allwein et al. employed it
into the SVM in 2000. Since then several works demonstrated to minimize the multiclass
problem, and apply the ECOC SVM in any data classification. In addition, results of the
previous research work, are shown that ECOC SVM error is lower than other classification
method’s error, and accuracy is higher than other methods. Because of this reason the ECOC
SVM is used in this study.

Dept. of CSE, RNSIT 2018-19 13

CHAPTER 5
RESULTS AND DISCUSSION

The main goal of this study is to classify the skin cancer images, which are extracted
by the deep convolutional neural network. A total of four different kinds of skin cancers are
classified using ECOC SVM. The results are obtained with a total of 3753 images, which are
collected from the internet (google.com, and naver.com, baidu.com, and bing.com). Figure 5.1
illustrates the some skin cancer sample images, which are used for the classification process.

Fig. 5.1 - Skin cancer sample images, which are used for the classification

Dept. of CSE, RNSIT 2018-19 14

CHAPTER 5 RESULTS
Numbers of training and testing images for the four kinds of skin cancers are presented in
Table 5.1. Total numbers of training and testing images for all kind skin cancers are 2985, and
768, respectively. Average value of the training process is 745. The numbers of the testing
images are between 185 and 200.

In this study the obtained convolutional neural network features are classified into four
groups, i.e., Actinic Keratosis, Basal cell carcinoma, Squamous cell carcinoma, and
Melanoma, using ECOC SVM.

Table 5.1 - Number of training and test samples of the skin cancers

No. Cancer Name No. of Training Images No. of Testing Images Total

1 Actinic Keratosis 712 185 897

2 Basal cell carcinoma 728 193 921
3 Squamous cell carcinoma 777 200 977
4 Melanoma 768 190 958
Total 2985 768 3753

The accuracy, sensitivity and specificity are calculated in this study by follow equations,
respectively.

Accuracy: The probability of the study.

TP  TN
Accuracy  …………………………. 5.1
TP  TN  FP  FN

Sensitivity (true positive rate): The result shows positive or cancer.

TP
Sensitivity  ….............................................................. 5.2
TP  FN

Specificity (true negative rate): The result presents negative or non-cancer.

Dept. of CSE, RNSIT 2018-19 15

CHAPTER 5 RESULTS
TN
Specificity  …………………………………………. 5.3
TN  FP

where TP – true positive or accurately classified positive conditions, TN – true negative or

accurately classified negative conditions, FP – false positive or inaccurately classified positive
conditions, FN – false negative or inaccurately classified negative conditions.

Table 5.2 presents implementation result for the skin cancer classification algorithm.
Maximum value of the average accuracy is 98.1% (for the cancer Squamous cell carcinoma),
minimum value is 82.17% (for the cancer actinic Keratosis). Also, maximum value for the
average accuracies in the four kinds of cancers is 95.1%, minimum value is 91.8%.

Table 5.2 - Implementation results for the skin cancer classification

Cancer Name Min Accuracy, % Max Accuracy, % Avg Accuracy, %

Actinic Keratosis 82.17 96.29 92.3
Basal cell carcinoma 85.23 95.6 91.8
Squamous cell carcinoma 85.95 98.1 95.1
Melanoma 86.46 97.2 94.2

The average values of the accuracy, sensitivity, and specificity from implementation
results are shown in Table 5.3 for each skin cancer. Maximum values of the average accuracy,
sensitivity, and specificity are 95.1 (squamous cell carcinoma), 98.9 (actinic keratosis), 94.17
(squamous cell carcinoma), respectively. Minimum values of the average in these measures
are 91.8 (basal cell carcinoma), 96.9 (Squamous cell carcinoma), and 90.74 (melanoma),
respectively.

Table 5.3 - Accuracy, sensitivity and specificity of skin cancer classification

Dept. of CSE, RNSIT 2018-19 16

CHAPTER 5 RESULTS

Cancer Name Accuracy, % Sensitivity, % Specificity, %

Actinic Keratosis 92.3 98.9 91.67
Basal cell carcinoma 91.8 97.7 86.73
Squamous cell carcinoma 95.1 96.9 94.17
Melanoma 94.2 97.83 90.74
A comparative study is performed with other existing previous works to inspect the eminence
of the proposed study is as shown in Table 5.4.

Table 5.4 - Implementation result comparison between previous research works and the
proposed study

Research Work Accuracy, % Sensitivity, % Specificity, %

Esteva et al. 72.1 96 –

Hitoshi et al. – 81.1 92.1
Anas et al. 83.33 – –
Almansour et al. 90.32 93.97 85.84
Abbas et al. – 89.28 93.75
Capdehourat et al. – 90 77
Giotis et al. 81 – –
Ruiz et al. 78.43 97.87
Isasi et al. 85 – –
Blum et al. 87.2 92.6 87.2
Kiran et al. 66.7 60.7 80.5
The proposed study 94.2 97.83 90.74

Dept. of CSE, RNSIT 2018-19 17

CHAPTER 6

CONCLUSION AND FUTURE WORKS

In this study, discussion of the classification of the skin cancers using ECOC SVM
with deep convolutional neural network. Whole four kinds of skin cancer images, i.e., actinic
keratosis, basal cell carcinoma, squamous cell carcinoma, and melanoma, are extracted by
deep convolutional neural network model. A total number of training images for all kinds of
skin cancers is 2985, and average value of the training process is 745. The numbers of the
testing images are between 185 and 200. A total number of the testing images is 768.

The implementation result shows as below: Maximum values of the average accuracy,
sensitivity, and specificity are 95.1 (squamous cell carcinoma), 98.9 (actinic keratosis), 94.17
(squamous cell carcinoma), respectively. Minimum values of the average in these measures
are 91.8 (basal cell carcinoma), 96.9 (Squamous cell carcinoma), and 90.74 (melanoma),
respectively. A comparative study is also performed with other existing previous works to
inspect the eminence of the proposed study. According to the comparison, the accuracy,
sensitivity, and specificity of our study is higher than previous research work’s results. Even
though the specificity of the previous work is higher than results of our study, their sensitivity
is very low.

Future work will focus on to extend the skin cancer classification based on the ABCD
(asymmetry, border, color, and diameter) rule for each cancer, and development a smartphone
application for the skin cancer classifications.

Dept. of CSE, RNSIT 2018-19 18

 REFERENCES

[1]. Aldebaro K, Nikola J, Alon O (2003) On nearest-neighbor error-correcting output codes

with application toall-pairs multiclass support vector machines. J Mach Learn Res 4:1–
15.
[2]. ali Bagheri M, Montazer GA, Escalera S (2012) Error Correcting Output Codes for
multiclass classification: Application to two image vision problems. The 16th CSI
International Symposium on Artificial Intelligence and Signal Processing (AISP 2012),
p 508–513.
[3]. Allwein EL, Schapire RE, Singer Y (2000) Reducing multiclass to binary: a unifying
approach for marginclassifiers. J Mach Learn Res 1:113–141.
[4]. Andre E, Brett K, Novoa Roberto A, Justin K, Swetter Susan M, Blau Helen M,
Sebastian T(2017) Dermatologist-level classification of skin cancer with deep neural
networks. Nature 542: 115–118.
[5]. Blum A, Luedtke H, Ellwanger U, Schwabe R, Rassner G, Garbe C (2004) Digital
image analysis fordiagnosis of cutaneous melanoma. Development of a highly effective
computer algorithm based on analysis of 837 melanocytic lesions. Br J Dermatol
151:1029–1038.
[6]. Cireşan DC, Meier U, Schmidhuber J (2012) Multi-column deep neural networks for
image classification. In: Proceedings of the CVPR, p 3642–3649.
[7]. Daniel R, Vicente B, Antonio S, Belén S (2011) A decision support system for the
diagnosis of melanoma: acomparative approach. Expert Syst Appl 38:15217–15223.
[8]. Dietterich TG, Bakiri G (1995) Solving multiclass learning problems via error-
correcting output codes. JArtif Intell Res 2:263–286Ebtihal A, Arfan JM (2016)
Classification of Dermoscopic skin cancer images using color and hybridtexture
features. Int J Comput Sci Netw Secur 16(4):135–139.
[9]. Germán C, Andrés C, Anabella B, Rodrigo A, Pablo M (2011) Toward a combined tool
to assistdermatologists in melanoma detection from dermoscopic images of pigmented
skin lesions. Pattern Recogn Lett 32:2187–2196.
[10]. Hu W, Huang Y, Wei L, Zhang F, Li H (2015) Deep convolutional neural networks for
Hyperspectral imageclassification. J Sensors 2015:258619.