ADVANCED TECHNIQUE FOR MELANOMA

AND NONMELANOMA SKIN CANCER

DETECTION USING NEURAL NETWORK

A Thesis Submitted
in Partial Fulfillment of the Requirements
for the Degree of

MASTER OF TECHNOLOGY
in
COMPUTER SCIENCE AND ENGINEERING
by
SHANU GAURA
(Roll No. 1516510517)
Under the Supervision of

Ms. Farah Shan Khan

KANPUR INSTITUTE OF TECHNOLOGY,
KANPUR (U.P.), INDIA

to the

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

Dr. A.P.J. ABDUL KALAM TECHNICAL UNIVERSITY
LUCKNOW
MAY 2018
 CERTIFICATE

This is to certify that Ms. SHANU GAURA (University Roll No.- 1516510517) has carried
out the research work presented in this thesis entitled “ADVANCED TECHNIQUE FOR
MELANOMA AND NONMELANOMA SKIN CANCER DETECTION USING NEURAL
NETWORK”, for the award of Master of Technology from Dr. A. P. J. Abdul Kalam
Technical University, Lucknow under my supervision of Ms. FARAH SHAN KHAN. The
thesis embodies results of original work, and studies are carried out by the student herself and
the contents of the thesis do not form the basis for the award of any other degree to the
candidate and anybody else from this or any other University/ Institution.

Date: Ms.Farah Shan Khan

Place: Assistant Professor
Department of CSE
Kanpur Institute of Technology,
Kanpur, India

ii
 Annexure IV

AKTU-PG-FORM 02

Dr. A.P.J. ABDUL KALAM TECHNICAL UNIVERSITY , LUCKNOW

(Formerly Uttar Pradesh Technical University, Lucknow)

CERTIFICATE OF THESIS SUBMISSION FOR EVALUATION

(Submit in Duplicate)

1. Name: ……………………………………………………………………...…………..

2. Enrollment No. : ……………………………………………………………………….

3. Thesis title:…………………………………………………………………….………..

4. Degree for which the thesis is submitted: ……………………………………………

5. Faculty of the University to which the thesis is submitted

6. Thesis Preparation Guide was referred to for preparing the thesis. YES NO

7. Specifications regarding thesis format have beenclosely followed. YES NO

8. The contents of the thesis have been organized based on the guidelines. YES NO

9. The thesis has been prepared without r es or t in g to plagiarism. YES NO

10. All sources used have been cited appropriately. YES NO

11. The thesis has not been submitted elsewhere for a degree. YES NO

12. Submitted 4 spirals bound copies plus one CD. YES NO

(Signature of the Candidate)

Name…...……………..

iii
 Annexure V

AKTU-PG-FORM 03

DR. A.P.J. ABDUL KALAM TECHNICAL UNIVERSITY,

LUCKNOW
(Formerly Uttar Pradesh Technical University, Lucknow)

CERTIFICATE OF FINALTHESIS SUBMISSION

(To be submitted in duplicate)

1. Name: ……………………………………………………………………...…………..

2. Enrollment No. : ……………………………………………………………………….

3. Thesis title:…………………………………………………………………….………..

4. Degree for which the thesis is submitted: ……………………………………………

5. Faculty (of the University to which the thesis is submitted)…………………………

6. Thesis Preparation Guide was referred to for preparing the thesis. YES NO

7. Specifications regarding thesis format have been closely followed. YES NO

8. The contents of the thesis have been organized based on the guidelines. YES NO

9. The thesis has been prepared without resorting to plagiarism. YES NO

10. All sources used have been cited appropriately. YES NO

11. The thesis has not been submitted elsewhere for a degree. YES NO

12. All the corrections have been incorporated. YES NO

13. Submitted 2 hard bound copies plus one CD.

(Signature(s) of the Supervisor(s)) (Signature of the Candidate)

Name(s)……………...…… Enrollment No:………………

iv
 ABSTRACT

Skin cancers are the most common form of cancers found in humans. The persistent raise of this
cancer in the worldwide, the high medical cost and death rate have prioritized the early
diagnosis of this cancer. Most of the skin cancers are curable at initial stages. So an early
detection of skin cancer can save the patients, if it can be detected early, the survival rate would
be increased. Although lots of effort has been made to advance the detection of skin cancers, the
challenging concerns still about it. Many researchers have been developed in automated
detection of melanoma. With the advancement of technology, early detection of skin cancer is
possible. One such technology is the early detection of skin cancer using Artificial Neural
Network.

The potential advantages of such studies are significant and incalculable. Moreover, the
difficulties entangle are a lot, and the new contributions in the area are highly appreciated.
However, it is extensively acknowledged that the more trustful and reliable detection systems
require higher accuracy. The different components in an automated diagnosis of skin cancer
include: an automatically skin cancer classification system is developed and the relationship of
skin cancer image across different type of neural network are studied with different types of
preprocessing. The collected images are feed into the system, and across different image
processing procedure to enhance the image properties. Statistical region merging (SRM)
algorithm is based on region growing and merging. Then the normal skin is removed from the
skin affected area and the cancer cell is left in the image. Useful information can be extracted
from these images and pass to the classification system for training and testing. Two neural
networks are used as classifier, Back-propagation neural network (BNN) and Auto- associative
neural network (AANN). Recognition accuracy of the 3- layers back-propagation neural
network classifier is 91% and auto-associative neural network is 82.6% in the image database
that include dermoscopy photo and digital photo. The analysis of work based on MATLAB.

Keywords: Skin Cancer, Neural Network, Image Detection, Image Processing, Statistical
region merging (SRM)

v
 ACKNOWLEDGEMENT

With deep gratitude I express my earnest thanks to my esteemed supervisor Ms. Farah Shan
Khan (Assistant Professor), Department of Computer Science and Engineering for their
constant involvement, energetic efforts and proficient guidance, which gave me direction and
body to work, reported here. Without their wise counsel and encouragement, it would have been
impossible to complete the thesis work in this manner.
I am thankful to all the faculty of the Computer Science and Engineering department especially
for their intellectual support during my research work.
I also want to thank to my friends for their valuable support whenever I needed. I would like to
thank all those people who have helped me some way or the other in my thesis work.
Lastly, and most importantly, I thank my parents for their moral support and encouragement
towards completing my thesis successfully.

Shanu Gaura
Date: M.Tech (Final Year)
Place: Roll No: 1516510517
Computer Science and Engineering
Kanpur Institute of Technology,
Kanpur, India

vi
 TABLE OF CONTENTS

CERTIFICATE ........................................................................................................................... ii

ANNEXTURE IV ....................................................................................................................... iii

ANNEXTURE V ........................................................................................................................ iv

ABSTRACT ................................................................................................................................. v

AKNOWLEDGEMENT ............................................................................................................ vi

TABLE OF CONTENTS .......................................................................................................... vii

LIST OF TABLES ................................................................................................................... viii

LIST OF FIGURES ................................................................................................................... ix

LIST OF SYMBOLS AND ABBREVIATION ........................................................................ x

CHAPTER 1: INTRODUCTION ......................................................................................... 1-12

1.1. General Introduction ...................................................................................................... 2

1.2. Motivation of study......................................................................................................... 3

1.3. Overview of human skin cancer ..................................................................................... 4

1.3.1 Human skin ........................................................................................................... 4

1.3.2 Skin cancer.............................................................................................................6

1.4. Skin lesion imagining methods ..................................................................................... 10

1.5. Objective of the work .................................................................................................... 11

1.6. Organization of the thesis .............................................................................................. 11

1.7. Conclusion .................................................................................................................... 12

CHAPTER 2: LITERATURE REVIEW ......................................................................... 13-17

2.1 Introduction .................................................................................................................. 14

2.2 Review .......................................................................................................................... 14

2.3 Conclusion .................................................................................................................... 17

CHAPTER 3: NEURAL NETWORK AND CLASSIFIER ............................................ 18-25

vii
 3.1 Introduction ...................................................................................................................... 19

3.2 Biological neural network................................................................................................ 19

3.3 Overview of ANN ...........................................................................................................................21

3.4 ANN classifier..................................................................................................................25

3.4.1 Back propagation neural network......................................................................... 25

3.4.2 Auto associative neural network ............................................................................26

3.5 Conclusion .................................................................................................................... 27

CHAPTER 4: METHODOLOGY AND P R O P O SE D ALGORITHMS.................... 29-39

4.1 Introduction ..................................................................................................................... 29

4.2 Pre-processing ................................................................................................................. 30

4.3 Segmentation process .......................................................................................................34

4.4 Feature extraction and selection .......................................................................................37

4.5 Classification....................................................................................................................38

4.6 Conclusion ....................................................................................................................... 39

CHAPTER 5: RESULTS AND DISCUSSIONS .............................................................. 40-48

5.1 Introduction .................................................................................................................... 41

5.2 Simulation tool............................................................................................................... 41

5.3 Result and analysis ..........................................................................................................41

5.3.1 Results of pre-processing .......................................................................................41

5.3.2 Results of Segmentation process ........................................................................... 43

5.3.3 Results of Feature extraction and selection........................................................... 43

5.3.4 Results of Classification ......................................................................................... 44

5.3.4.1 Results of BNN classifier ............................................................................. 45

5.3.4.2 Results of AANN classifier ......................................................................... 46

5.4 Conclusion ................................................................................................................. 48

viii
CHAPTER 6: CONCLUSION AND FUTURE SCOPE .........................................................49-50

6.1 Introduction ...................................................................................................................... 50

6.2 Future Scope of work ....................................................................................................... 51

REFERENCES ..................................................................................................................... 53-55

APPENDIX ...........................................................................................................................56-68

LIST OF PUBLICATIONS ...................................................................................................... 69

C.V. OF RESEARCH SCHOLAR........................................................................................... 70

ix
 LIST OF TABLES

Table.1.1. BNN classification results with different layer .................................................. 46

Table.1.2. AANN classification results with size of neurons ..............................................47

x
 LIST OF FIGURES

Figure1.1. Epidermis and dermis layer in human skin ........................................................... 5

Figure1.2. Malignant melanoma ........................................................................................................... 7

Figure1.3. Nodulocystic based cell carcinoma ...................................................................... 8

Figure1.4. Superficial based cell carcinoma .......................................................................... 8

Figure1.5. Sclerosing based cell carcinoma ............................................................................9

Figure1.6. Squamous cell carcinoma ..................................................................................... 9

Figure3.1. Two connected biological neurons ..................................................................... 20

Figure3.2. Model of an artificial neurons .............................................................................. 21

Figure3.3. Feed forward ANN ............................................................................................................ 23

Figure3.4. Selection of the number of layer .......................................................................... 24

Figure3.5. Multilayer feed forward ANN ........................................................................................ 25

Figure3.6. Structure of AANN ........................................................................................................... 26

Figure4.1. Skin cancer system detection system methodology ............................................ 30

Figure4.2. True image identification of skin cancer ............................................................ 31

Figure4.3. Image de-noised by wavelet ............................................................................... 31

Figure4.4. Image from median filter .....................................................................................32

Figure4.5. Computation of mean value ................................................................................. 32

Figure4.6. An ideal histogram.............................................................................................. 36

Figure4.7. Segmented image ............................................................................................... 36

Figure4.8. Original image and first, second level decomposition ....................................... 38

Figure5.1. True image identification of skin cancer ............................................................ 42

Figure5.2. False image identification of skin cancer ............................................................ 42

xi
Figure5.3. Segmentation from SRM .................................................................................... 43

Figure5.4. Gray and BW image composition........................................................................ 43

Figure5.5. DWT Gray and DWT BW image composition ................................................... 44

Figure5.6. Training image results of detected skin cancer .................................................... 45

Figure5.7. Neural network training tool simulation.............................................................. 48

xii
 LIST OF SYMBOLS AND ABBREVIATIONS

ANN Artificial Neural Network

AANN Auto-Associative Neural Network
BPNN Back propagation Neural Network
GHE Global Histogram Equalization
LHE Local Histogram Equalization
OCT Optical Coherence Tomography
RCM Reflectance Confocal Microscopy
SRM Statistical region merging
SVM Support Vector Machine

xiii
 CHAPTER-1

INTRODUCTION

1
 CHAPTER I

INTRODUCTION

1.1 GENERAL INTRODUCTION

Cancer is one of the biggest threats to human beings and is the second leading reason regarding
demise the whole ball human [24]. According in accordance with the statistical statistics out of
WHO(World Health Organization) , cancer precipitated respecting 7.6 bags of people death
international in 2008, or that is estimated so the quantity over deaths prompted via cancer is
animal improved yet the number choice perchance expand according to 13.1 million in 2030 .
Based concerning related research, most cancers intention becomes the lead loss of life within
next 20 years [18]. There are three types of pores and skin cancer: Basal mobile most cancers yet
squamous mobile most cancers are called non-melanoma pores and skin most cancers whilst
melanoma is concerning melanocytes cells, Basal cell cancer then squamous cell most cancers
are frequent but much less dangerous. Melanoma is not common as non-melanoma skin cancers;
however, such is greater probably in accordance with extent or emerge as deadly [19]. Although
skin cancer is a threat in accordance with human life, thankfully also melanoma perform remain
cured if it is detected early. According to research, the curable dimensions pleasure keep more
than 90% excessive condition most cancers do stay identified within its before long stages while
the curable dimensions pleasure keep less than 50% into its late stage [4]. Thus, the detection
over skin cancer between its promptly flooring has attracted the interest beyond distinctive fields
[21]. With observance to the raises between statistical basis of pores and skin cancer, the worth
of it’s quickly discovery has been considered so a critical problem and the laptop primarily based
analysis is an necessary tool for this purpose.

Currently the prognosis on pores and skin most cancers are done frequently by a human,
whichever is called dermatologist. During diagnosis, dermatologist assessments the skin
carefully by means of his eyes or using a gadget referred to as dermascope. The skin cancer cruel
melanoma execute keep sound salvo diagnosed yet deal with such among shortly stages.
Therefore before long diagnosing is a quintessential difficulty because patients. However, solely
skilled doctor is able to marshal the skin cancer from sordid pores and skin diseases. Thus the pc

2
based totally pores and skin most cancers discovery is vital in accordance with grant advice
because non-specialized user. The development about that diagnosis system over 20 years, the
precision regarding diagnosis is around 73% in accordance with 98%. The variants of prognosis
are efficiency vast or there are lacks of element of the check methods. In this work, a developing
an routinely skin most cancers detection dictation after group the most cancers photos between
both prevailing yet malignant melanoma is then mentioned through the use of exclusive
approaches. The pictures in the databases aged are contained each digital Image then
dermoscopy images. Dermoscopy, also calls Dermatoscopy yet Epiluminescence light
microscopy (ELM). It used to be advance announced concerning 1987, that is a form over
imaging method makes use of in imitation of exanimate lesions with a dermatoscope. The
technique is performed with the aid of setting an oil immersion within the pores and skin and the
optics. The lighting fixtures execute enlarge the pores and skin that improve regarding divulge
most of the pigmented structure, one of a kind color colors that is no longer seen in accordance
with naked eye; or approves advise viewing then evaluation about the epidermis (the outer strata
regarding the skin) and papillary dermis (the sound vascular intestinal tier about the skin).
Physician use it technique because diagnosis concerning pores and skin most cancers extra
efficiency. The historical researches have measured as ELM be able improve the diagnostic
precision by way of 5 – 30% evaluate in conformity with usual imaging [25]. Furthermore, ELM
has growth or digitalize so much be able be used according to alignment including computer. It
attached countless benefits between diagnosing as considering younger dubious hit then goal
contrast concerning parameters: geometry, coloration then texture; yet storage concerning Image
or related because after development [17]. ELM might also improve the rigor about clinical
diagnosing however such requires the skilled dermatology physician to exanimate the image.
This treatise factory of evaluate distinctive strategy along their performance and accuracy.

1.2 MOTIVATION FOR THE STUDY

The tremendous interest has been developed among computer-based detection concerning skin
most cancers throughout latest decades. This performance includes ethnic deportment prediction,
rate prediction, potential after fulfilled a task, place of job design, yet dense vile ergonomics
studies so situation such as is done via ethnic beings. The cause on such systems is after supply
2nd opinion over analysis including much less calamity yet higher exactness and reliability than

3
the effects attain normally with the aid of a human specialist [11]. Many investigations have been
flourished among automatic detection concerning skin cancer. The main blessings on such
studies are sizeable yet incalculable. Moreover, the difficulties overturn are a lot, then the new
contributions in the place are surprisingly appreciated. Generally, an automatic analysis about
skin most cancers consist of quite a few components: Preprocessing, segmentation, feature
extraction then selection, yet classification.

1.3 OVERVIEW OF HUMAN SKIN CANCER

Skin Cancer as much one on the leading purpose of demise is the threat in conformity with
ethnical beings into complete world. This most cancers may keep cured postulate such is
diagnoses of shortly degrees .With observance in imitation of the raises among statistical basis
on skin cancer, the respect over its early detection has been considered as like a essential
difficulty then the computer based analysis is an vital tool because this purpose. The promptly
detection regarding skin cancer has attracted much concern out of one of a kind fields. Since the
assignment introduced between those dissertations lies on the automatic detection dictation for
pores and skin cancer, the knowledge in relation to human skin alongside including the unique
available strategies because of discovery structures is the essential information between this area.

1.3.1 Human Skin

The skin keeps ethnical physique safe beside heat, injury, infection yet damages occur with the
aid of ultraviolet (UV) radiation. Also that is able produce nutrition D, preserve cloud and fat.

4
 There are exclusive layers of human skin.

Figure 1.1 A squamous cell, basal cell, and melanocyte and epidermis and dermis layers in
Human skin

Epidermis and Dermis are the two main layers in human skin which are described as follows :

Epidermis: This tier namely the pinnacle seam among ethnical skin is formed on squamous cells
which are even cells into skin. The spherical cells beneath the squamous cells are called basal
cells. The cells among a deepest portion concerning dermis are known as melanocytes which
have been placed between the basal cells. The stain (color) within pores and skin is appeared by
Melanocytes.

Dermis: The 2d predominant strata over skin are dermis as is located below the epidermis. It
includes extraordinary types concerning cells such namely lymph vessels, gore vessels then
glands. Some glands help the pores and skin after dry out, partial others help in accordance with
peaceful the body then perform sweetening. The discern 1.1 suggests the layers and cells
permanency of skin.

5
1.3.2 Skin Cancer

Trillions about dwelling cells consists the ethnical body. These cells grow, and division of
current cells within everyday our bodies yet die orderly. In adults, the cells divide is after
alternative worn-out, damaged, and additionally death cells. When the rising regarding unnatural
cells within a part concerning body extend out about limit motives in imitation of most cancers
[11]. The increase among cancer cells desire perform recent cancer cells and able in conformity
with invade mean tissues as much well [10]. Mostly cancer cells construct a tumor, however
between partial cancers certain as leukemia, the tumors are formed rarely. The cells of these
kinds on cancers are discovered in gore and skeleton marrow. Not entire the tumors are cancers,
those are known as forcible who execute develop and edit problems then pressure on healthful
organs. They are not in a position according to fall on within sordid plantain [11].

Skin cancer as much the just common most cancers into human begins within the pores and skin
[25]. Some cancers also may start of mean organs then extent concerning the skin, however it
cancers are not regarded as like pores and skin ones [26]. The extraordinary types about pores
and skin cancers generally do lie categorized so cruel melanoma yet non-melanoma pores and
skin cancer (NMSC), the current which includes Basal Cell Carcinoma or Squamous Cell
Carcinoma as the essential subtypes.

Malignant melanoma

Malignant melanoma, namely some over the kinds on pores and skin cancer, is increasing global
and leads to loss of life regarding 65% of its victims.

Between 1991 and 2000 into UK, the sufferers concerning melanoma grew by using 59% and
41% in men and women, respectively. In 2010, Australia, the mortality concerning Skin most
cancers and melanoma with projected occurrence life 11,500 and 1500 respectively. The
excessive occurrence concerning each melanoma then non-melanoma skin cancer between
Australia grant that U.S.A as a vicinity about lookup among the location [13, 27] It is spinoff
beyond epidermal melanocytes or perform arise into any skill which carries it cells, but in many
instances such is seemed concerning the decrease limbs of ladies yet of the again in males. As it
happens on the skin surface; therefore such is detectable through visual inspection. The clinical

6
appearance is special according to the kind then site over the tumour. Figure 1.2 shows the
pattern picture on cruel melanoma [27].

Figure 1.2 Malignant melanoma [27]

Malignant melanoma perform show up by way of the phenotypic elements such so solar
exposure habits, intermittent and ultraviolet radiation. The lousy gamble factors are the tidy skin
type, lowlife the history over unkind melanoma of private or first-degree blood relation [13, 27].
The exclusive kinds on unkind melanomas are [13]:

Superficial spreading and nodular melanomas: The lesions are regularly asymmetrical
including unmethodical border. It has more than one colour or the diameter is greater than 0.6
cm. It may also stand washing yet ulcerate.

Lentigo maligna and lentigo maligna melanoma: It normally happens about the surface in
elderly patients. It looks like a enormous then ultra vires knot who tends in accordance with
develop lightly.

Acral lentiginous melanoma: It usually happens over the pores and skin over fingers yet soles
as doesn’t have some hair. It almost identified late, for that reason bear a poorest forecast
amongst vile sorts over unkind melanoma.

Amelanotic melanoma: It typically foreboding false. The correct diagnosis is decided since
biopsy.

7
  Non-melanoma skin cancer

The main two types of non-melanoma skin cancer are [13, 27]:

Basal Cell Carcinoma: It is the near frequent malignancy between different countries. It occurs
within one-of-a-kind components of shoulders, ears, face, back, and scalp. Its medical
appearance is exceptional according after the kind or web site of tumour.

Nodulocystic basal cell carcinoma: It is small, pearly nodule, translucent or oft along surface
telangiectasia. As the hit is magnified, such normally ulcerates according to make a rolled area
then considerate crust. Figure 1.3 is a pattern on nodulocystic basal cell carcinoma.

Figure 1.3 Nodulocystic basal cell carcinoma [13]

Superficial basal cell carcinoma: It is scaly, plaque yet crimson that grows slowly. It is
normally show up over the trunk. The telangiectasia and rolled side are normally observable by
means of excellent light. Mass 1.4 is a pattern about superficial basal telephone carcinoma.

Figure 1.4 Superficial basal cell carcinoma [13]

8
Sclerosing (morphoeic) basal cell carcinoma: It is scar-like plaque as the side is unhygienic
specified. It is a pure lesion with a gently expanding. Figure 1.5 is a sample regarding sclerosing
(morphoeic) basal cellphone carcinoma.

Figure 1.5 Sclerosing (morphoeic) basal cell carcinoma [13]

Squamous Cell Carcinoma: It is usually seemed in continual photo voltaic harm encompass
scalp, dorsum on hand, lower lip, arm and ear. It starts off evolved out of tiny then crusted
plaque or will become indurate yet nodular. It is nearly together with ulceration. Formal 2.6 is a
sample of Squamous Cell Carcinoma.

Figure 1.6 Squamous Cell Carcinoma [27]

9
1.4 SKIN LESION IMAGING METHODS

Imaging is the methods and techniques after edit the picture and statistics out of organic
structures and purposes concerning the body. Different imaging methods over skin lesions are
chronic according to discover the skin cancer. The frequent imaging techniques are namely
follows [21]:

 Dermoscopy

Dermoscopy as much non-invasive imaging approach has been utilized because of detecting
skin cancers. It is additionally known as epilumence microscopy (ELM) yet pores and skin
floor microscopy. This approach renders the skin by using applying the floor reflectance
lordly light methods. It permits visualizing the coloration and subsurface constructions
certain so gore vessels yet paint in imitation of help for discovery about skin cancer between
promptly ranges .The shape on dermoscopy is in imitation of turn to advantage the cross
polarized mild then immersion mediocre certain so immersion salad oil or booze according to
reduce surface reflections.

 Ultrasound

In this technique, the ultrasound waves reflected from the banana is ancient in conformity
with anticipating the pores and skin morphology. Although this technique execute penetrate
in conformity with the pores and skin deep because proved the depth on tumor then
evaluating the lymph nodes, but the scientific application is no longer big because about the
paltry decision as is not in a position according to function histomorphologic distinction
between pores and skin lesions.

 Optical Coherence Tomography (OCT)

This is a non-invasive approach which is based totally of interferometer. It produces cross-

sectional and 2-dimensional images. Although OCT pictures may characterize the macro
morphology namely nicely as like gore vessels and indexical structures, is not able in
conformity with discern the sub mobile details yet basement membrane. Hence, that is not in
a position in accordance with reliably discover the shortly incursion concerning tumor.
10
  Reflectance confocal microscopy (RCM)

This technique is non-invasive and painless approaches who perform invent the skin’s cell
details within vivo besides artifacts’ processing. RCM perform discriminate the refraction
indices in imitation of discover the pores and skin chromospheres encompass water, hemoglobin
or melanin. A short spot within a tissue is irradiated via a point light source or then cogitation is
carried out in accordance with the detector.

1.5 OBJECTIVE OF THE WORK

The intention regarding dissertation is after study, increase then request strategies yet
methodologies based totally over neural community classification. The automatically notice
pores and skin cancer array law is promoted in conformity with relation about skin most cancers
image throughout concerning different sorts of neural network; it networks are studied along one-
of-a-kind sorts of preprocessing. Then pores and skin most cancers affected pictures statistics
skip in imitation of the alignment dictation because coaching yet testing. Then recognition
precision well-read with the again manufacture yet auto convivial neural networks.

1.6 ORGANIZATION OF THE THESIS

Present thesis is organized in seven chapters as discussed below:

Chapter 1: This chapter includes ordinary introduction, Overview over ethnical skin cancer and
types over skin cancer, skin coup imaging methods, dictate and goal regarding the work. Finally,
it book explains organization over the thesis.

Chapter 2: This chapter explains writing critiques over a number of journals then conference
publications.

Chapter 3: This chapter includes a concept on artificial neural networks, its classifier.

11
Chapter 4: This chapter mentioned about methodology used in this dissertation work. It is pores
and skin most cancers discovery provision or proposed image processing algorithms.

Chapter 5: This chapter consists of consistency and artificial effects regarding thesis work.

Chapter 6: This chapter consists of the finishing about the job along barriers then after scope.

1.7 CONCLUSION

In this chapter introduction of this thesis, overview concerning this arrangement. Also explains
so the overview on ethnical pores and skin cancer yet types of skin cancer. A primary pores and
skin lesion imaging methods discusses into this book then objective about the thesis.

12
 CHAPTER 2

LITERATURE REVIEW

13
 CHAPTER 2

LITERATURE REVIEW

2.1 INTRODUCTION

This chapter summarizes the pioneer work on discovery concerning skin cancer. The major goal
over this lookup is to study automatically detection concerning skin cancer by alignment
regarding neural community with image technology algorithm, as better as like presenting
acceptable computational complexity appropriate for sensible implementation.

2.2 REVIEW

One regarding the auspicious methods in accordance with take aforementioned challenges within
automating medical imaging diagnosis is in imitation of simplify the goal on the evaluation or to
exploit half form about hypothetical records respecting the imaged structures. The facts touching
the structures in accordance with stay analyzed do stay corporal abilities respecting theirs
standard look (such as like form and gray levels) and position; then statistical capabilities over
their residences (such as much mature degree concerning the tissues included of those
structures). The pictures execute below lie categorized the use of theirs morphological, color,
fractal, and earth properties. Permanency Laws, 1980 of his work converted digital images in
imitation of become aware of areas on hobby or supplied an input dataset because segmentation
or purposes detection operation.

Various kinds concerning strategies hold been proposed in imitation of improve the precision
concerning skin most cancers diagnosis. The dermoscope yet epiluminescence microscope
(ELM) was once preceding described concerning 1987 [25]; such enables non-invasive analysis
procedure primarily based concerning the usage of about obtaining light, oil immersion then a
magnifier. But its exactness is still on the whole depends of health practitioner experience. The

14
research into automatic discovery for skin cancer has been carried out within closing not many a
long time with several strategies then techniques. During the last few years, telemedicine with
remote image viewing and analysis has emerged as a highly valuable and versatile tool,
particularly suited to places where local medical expertise is limited. Granot et al., 2008 worked
on creating a medical imaging system consisting of physically separated components of medical
imaging system in order to produce a robust and less expensive system that can be used by
trained non-medical personnel. Adoption of simple method of micro image processing which
could significantly increase opportunities and quality diagnostics while lowering costs and
considerably increasing connectivity between most isolated laboratories and distant reference
center has been proposed by Aher & Kaore, 2010.

Dobrescu et al., 2010 described a technique on an algorithm because of automatic discovery on

power on pores and skin wound who is based on both provincial fractal purposes (local fractal
dimension) or texture services derived from moderate co-occurrence matrices (such as much
contrast, energy, and homogeneity). Tushabe et al., 2011 proposed an image-based analysis
method where images on pores and skin ailment have been old in accordance with classify pores
and skin ailments between huge class on either viral contaminated then bacterial infected.

Malignant melanoma presently money owed because of a 0.33 regarding most everyday kind
over pores and skin cancer yet 79% on pores and skin cancer death. The fall concerning unkind
melanoma between fair-skinned sufferers has improved histrionically among near parts about the
world atop the past not many many years (Rubegni et al., 2002; Stanganelli et al., 2005;
Dobrescu et al., 2010). In Europe, it’s been stated so malignant melanoma fall is growing
through 5% each and every year then such is accountable for 91% concerning skin cancer death
(Sboner et al., 2001; Ali & Deserno, 2012). In a command after improving early detection, a
variety on diagnostic checklists then rules have been proposed such as Seven Point Checklist
(Healsmith et al., 1994), and ABCDE: Asymmetry, Border, Color, Diameter, Evolution checklist
(Fitzpatrick et al., 1998). These policies or checklists specify visible applications associated
including unkind hit symptoms.

Stolz et al., 1994 of their work, promoted a diagnosis plan because dermoscopic images,
receivingaccess to the Asymmetry (A), Border (B), Color (C), then Diameter (D) about specific

15
photo structures. This ABCD regimen grew to be the value into Dermoscopy for staging PSL of
benign, suspicious, or malignant moles (melanoma). However, dermoscopic analysis is
frequently complicated or subjective, as a consequence associated together with poor
reproducibility and vile truth specially among inexperience dermatologist, as much the precision
of experts is 65-84% (Argenziano et al., 2003; Lee, 2001, Stanganelli et al., 2005). Also, visual
interpretations on this capability by using dermatologist hold so some distance validated after lie
a hard task. Lee 2001 in his learning pointed out detection dimension based totally concerning
scientific visual taking care of in accordance with be about 65%. stability Melanoma is highly
curable if diagnosed early and treated properly as survival rate varies between 15% and 65%
from terminal to early stages respectively (Ali & Deserno, 2012). Depending on the observer’s
experience, Dermoscopy improves the diagnostic accuracy for melanoma detection up to 50% as
compared with traditional visual inspection (Kittler, 2004). In the last decade, usage of
Dermoscopy or Epiluminiscence Light Microscopy (ELM) changed the dermatologist’s
approach to suspicious PSL. However, the analysis made using ELM are extremely complex and
subjective (Rubegni et al., 2002). To prevent aforementioned challenge of quantitative
interpretation, methods based on Computer-Aided Diagnosis (CAD) have been introduced
towards automating the diagnosis procedures, such as in Rubegni et al., 2002; Stanganelli et al.,
2005; Mittra & Parekh, 2011. In 2011,Daniel Ruiz, Vicente Berengue, Antonio Sorianoand
Belén Sánchez proposed sorts of ANN classifiers, which location multilayered perceptron, a
Bayesian classifier or the algorithm concerning the K nearest neighbors. These strategies
assignment independently and additionally of combination erection a collaborative selection aid
system. The array rates near are around 87% [23].

Gilmore et al., 2009 ancient lacunarity (a measurement about transitional invariance on an

destination old within quantifying aspects concerning patterns so exhibit scale-dependent
modifications in structure) after provide a hopeful approach for automatic evaluation concerning
melanocytic nevi or melanoma. The fuzzy-based histogram analysis technique back via Stanley
et al., 2003 supplied a possibility because automatic skin lesion comparison of dermatology
clinical images. Rubegni et al, 2002 advanced an computerized manner using synthetic neural
network methods based regarding mathematical evaluation over pigmented pores and skin
lesions in imitation of avoid the trouble on characteristic signification made via the usage about

16
ELM by using Dermatologist. Kreutz et al., 2000 a aggregate regarding artificial neural network
approach with ground evaluation using digital picture processing yet mixture-of-experts after
strive automation of pores and skin cancer diagnosis. Ganster et al., 2001 flourished a dictation
so supplied for computerized computerized evaluation regarding images present out of ELM
according to enhance the express awareness about cruel melanoma. Sheha et al. 2012 ancient
Grey-Level Co-occurrence Matrix (GLCM) and Multilayer perceptron classifier (MLP) because
of automatic Detection regarding Melanoma Skin Cancer using Texture Analysis. One core
challenge then again along dense of aforementioned tactics is their incapability to integrate well
with thoroughgoing devices certain so mobile telephony, at last largely on hand according to
underserved areas (Mobithink, 2012). This thesis degenerated on section on our assignment
among [16] investigated one type regarding skin cancer detection based totally on laptop
learning.

In this thesis, a thriving an robotically skin most cancers discovery dictation to classify the
cancer snap shots in both efficacious or malignant melanoma is then mentioned through using
one of a kind approaches. The pictures of the databases ancient are contained each digital picture
or dermoscopy images.

2.3 CONCLUSION

This section concludes the pioneer manufactory regarding express discovery over pores and skin
cancer. This dissertation also discusses about the preliminary research as neural network
alignment techniques; center of attention of action about attention regarding truth together with
synthetic neural network.

17
 CHAPTER 3
NEURAL NETWORK AND
CLASSIFIER

18
 CHAPTER 3

NEURAL NETWORK AND CLASSIFIER

3.1 INTRODUCTION

There are many expert systems [29] and Neural network is applied here. The ANNs consist
regarding deep related neurons simulating a brain at work. A primary characteristic as
distinguishes an ANN beside an algorithmic program is the capability to generalize the abilities
over latter data that was once no longer presented in the course of the learning process. Usual
structures need in conformity with gather genuine abilities over its precise area. However, ANNs
solely desires one coaching and exhibit tolerance because discontinuity, accidental disturbances
yet also defects between the coaching records set. This approves because utilization on ANNs
within fixing troubles which cannot be solved through vile potential successfully [7, 9]. These
features or benefits are the motive from what the region regarding ANN’s application is
altogether wide or includes because of example:

• Pattern recognition,

• Object classification,

• Medical diagnosis,

• Forecast of economical risk, want costs changes, necessity because of electrical power, etc.

• Selection about employees,

• Approximation regarding feature value.

19
3.2 BIOLOGICAL NEURAL NETWORKS

The ethnical Genius consists about around nerve cells referred to as neurons. The core can stand
dealt with as like the computational fuehrer about a neuron. Here the predominant methods
receive place. The yield channel over a neuron is called axon as dendrite is its input. One neuron
is able bear deep dendrites but only certain axon; organic neurons have heaps regarding
dendrites. Connections within neurons are referred to as synapses; their content of a human
Genius is higher than [13]. A neuron receives electrified impulses thru its dendrites then sends to
them according to the subsequent neurons using axon. An axon is cut up among dense branches
ending with synapses. Synapses exchange monitoring concerning acquired sign before the
subsequent neuron pleasure obtain it. Changing the strengths concerning synapse outcomes is
disingenuous in imitation of stay a essential share about study manner and that faith is exploited
within models about a human brain among its artificial equivalent [9,1].The human talent
consists about round nerve cells known as neurons. The meat execute remain treated as the
computational fuehrer on a neuron. Here the major tactics absorb place. The output race about a
neuron is known as axon as dendrite is its input. One neuron can have deep dendrites but solely
certain axon; biological neurons have lots concerning dendrites. Connections between neurons
are known as synapses; theirs amount among a ethnical talent is increased than durability [13]. A
neuron receives electrical impulses thru its dendrites and sends to them in conformity with the
subsequent neurons using axon. An axon is split into many branches finish along synapses.
Synapses trade limit concerning obtained signal earlier than the next neuron choice get hold of it.
Changing the strengths about synapse effects is illusory after stay a indispensable part on
discipline procedure yet that creed is exploited of fashions regarding a human Genius into its
artificial equal [16, 1].

20
 Figure 3.1 Two connected biological neurons [7]

3.3 OVERVIEW OF ANN

Structure on a synthetic neuron, labored oversea by means of McCulloch and Pitts within 1943
[1], is comparable according to organic neuron (Figure 3.2).

Figure 3.2 Model of an artificial neuron [20]

It consists about couple modules: summation module Σ and activation module F. Roughly the
run-on module corresponds according to organic nucleus. There algebra run-on over weighted

21
input indicators is cooked and the output signal φ is generated. Output sign be able remain
thought the usage of the formula,

φ = ∑𝑚
𝑖=1 𝑤𝑖 = 𝑢 (3.1)

Where w is the vector about weights (synapses equivalent), u - vector concerning input signals
(dendrites equivalent), m - quantity regarding inputs. Signal φ is processed via the activation
module F, who can stand precise via different features according in conformity with needs. A
simple linear function do be used, below the output signal y has shape

y=Kφ (3.2)

Where k is coefficient, Networks the use of this characteristic are referred to as Madeline or
theirs neurons are known as Adaptive linear element. They are the simplest networks, which bear
located sensible application. Another kind on the activation module function is a beginning
feature

1, φ > φh
( 𝑥) = { (3.3)
0, φ ≤ φh

Where φh is regular beginning value, however, features which construct a non-linear plan of
organic neuron greater precisely are a sigmoid characteristic

y= 1
1+ 𝑒−𝛽φ
(3.4)

Where 𝛽 is given parameter and a tangensoid function,

𝛼𝜑 1− 𝑒−𝛼𝛽
y = tan h ( ) (3.5)
2 1+ 𝑒−𝛼𝛽

Where α is fond parameter [7, 16, 6]. Information capacity and technology potential over a odd
neuron is exceedingly small.

However, such be able stay advanced by the fabulous attachment of dense neurons. In 1958 the
advance ANN, called perceptron, was raised through Rosenblatt [2]. It was used because of

22
alphanumerical personality recognition. Although the outcomes were now not satisfactory, that
was once successful as the preceding regulation built, as false a neural network. Rosenblatt has
additionally standardized that salvo a fond hassle execute keep solved with the aid of a
perceptron, afterward the answer execute stand observed between a finite wide variety of steps
[7, 9 then 6]. After e-book about the e book “Perceptrons” by way of Minsky or Papert of 1969
[3] lookup of the location regarding ANNs had come in accordance with a standstill. The authors
measured as perceptrons can't remedy linearly non-separable issues (e.g. tell an XOR-function).
Research has been resumed after nearly 15 years by using a collection on publications displaying
so much these obstacles are no longer relevant according to non-linear multilayer networks.
Effective education strategies hold also been added.

Neurons between the multilayer ANNs are grouped within iii special sorts over layers: input,
output, or hidden seam (Figure 3.3). There do be certain and extra secret layers into the
community but only some outturn and some input layer. The wide variety over neurons in the
input ledge is designated through the type yet aggregation concerning statistics as pleasure stands
partial according to the input. The wide variety regarding output neurons corresponds to the type
over reply about the network. The total of hidden layers or theirs neurons is more tough in
imitation of determine. A network including certain black tier suffices in conformity with clear
up just tasks. None of the recognized problems wants a network along greater than iii lawful
layers of order in accordance with be solved (Figure 3.3) from the neurons within the enter ledge
(IL) signals are preached according to the stolen layer (HL) yet then in the end after the outturn
strata (OL). There is no helpful wrinkle for the number of secret neurons selection. One over the
techniques is described via formula

𝑁ℎ = √𝑁𝑖 𝑁𝑜 (3.6)

Where Nh durability is the variety on neurons between the hidden layer, or Ni then No are the
correspondent numbers because the enter then yield layers, respectively. However, typically the
volume about hidden neurons is determined empirically [7, 9 then 6]. Two kinds on a multilayer
ANNs may remain exclusive along regards according to the architecture: feed-forward then
remarks networks. In the feed-forward networks sign perform pace among certain path solely
then can't career in neurons in the same seam (Figure 3.3).

23
 Figure 3.3 Multilayer feed-forward ANN. [9]

Figure 3.4 Selection of the number of layers for solving different problems. [6]

24
Such networks do be chronic into the pattern recognition. Feedback networks are more
complicated, because a signal may lie sent returned in accordance with the enter regarding the
equal ledge including a changed virtue (Figure 3.5). Signals execute motion of it loops till the
good state is achieved. These networks are additionally called interactive or recurrent [6, 10].

Figure 3.5 Multilayer feed-forward ANN. [9]

3.4 ANN CLASSIFIER

Wide degrees concerning classifiers are available and each certain regarding them has its
strengths or weaknesses. Classifier performance depends considerably of the characteristics on
the facts in imitation of remain categorized then like is no alone classifier that workshop best
overall attached problems. Various experimental checks have been observed to compare
extraordinary classifier performance and according to discover out the kin within characteristics
about statistics or the classifier performance. Determining a suitable classifier for a fond problem
is then again nonetheless more an art than a science. In it arrangement work, pair neural
networks are aged as classifier, Back-propagation neural network (BNN) then Auto-associative
neural community (AANN).

25
3.4.1 Back-Propagation Neural Network (BNN)

Back-propagation neural network is some concerning the near frequent neural community
structures, so such is easy then effective. Back-propagation is the generalization of the Widrow-
Hoff lesson administration to multiple-layer networks or nonlinear differentiable transfer
functions [20]. The black or output strata nodes regulate the weights cost relies concerning the
error among classification. The change about the weights is in accordance according to the
gradient concerning the frenzy curve, which points in the course in imitation of the local
minimum. BNN is advantage about reckoning then array however the technology pace is slower
in contrast according to mean study algorithms.

3.4.2 Auto-Associative Neural Network

Figure 3.6 Structure of AANN. [20]

Auto-associative neural community is a network committed concerning even volume of the input
strata and the outturn layer, and smaller size of stolen layer. The shape on AANN is proven on
configuration 3.6. The input strata yet mapping strata form a depth network yet the De-mapping

26
yet outturn seam employment as much a decoder network. Unlike the others neural network, the
desired yield over AANN is identity out of the input.

3.5 CONCLUSION

This section explains an overview concerning artificial neural network; BNN and AANN are two
fantastic classifiers of proposed algorithm. According to it work, these classifiers have multiplied
the propriety about pores and skin cancer detection system.

27
 CHAPTER 4
METHODOLOGY AND
PROPOSED ALGORITHMS

28
 CHAPTER 4
METHODOLOGY AND PROPOSED ALGORITHMS

4.1 INTRODUCTION

This chapter proposes an algorithm for Automatic discovery of pores and skin most cancers who
is one of the close challenging issues of medical image processing. It helps doctors in imitation
of determine whether a pores and skin melanoma is effective yet malignant. So, identifying the
more efficient techniques about detection after limit the rate over blunders is a necessary trouble
among researchers. The performance concerning this assignment is according to suggest and
evaluate the modern yet innovative scientific expert systems to that amount routinely in a
position after align skin most cancers as like inoffensive or dangerous, precisely; If possible,
together with less blunders than ethnic experts.

The proposed Algorithm regarding this thesis uses a novel strategy because of classifying the
pores and skin cancer. Most regarding the structures consist regarding following procedures:
image pre-processing to removes the noise yet great hair; Post-processing: in accordance with
beautify the structure about image; Segmentation: in accordance with gets rid of the wholesome
skin beside the image then find the region regarding interest, usually the cancer cell, stays in the
image; Feature extraction extracts the useful information then picture homes from the segmented
images; Finally, this facts is old of the array law for education and checking out purpose. The
classification provision is typically supported through shrewd classifier, such so neural
community and assist vector machine. Different strategies among exclusive steps hold been back
to attain that environment friendly system. The cause concerning this lookup is in conformity
with recommend contributions between distinctive stages regarding it system. The algorithms try
to speed on the detection together with less error than sordid regular ones. It is meant so the
proposed algorithms hold make a contribution within populace health systems or help medical
experts according to veil skin most cancers detected in early stages.

29
In summary, the subject regarding dialogue execute remain categorized of stages as like
illustrated of parent 4.1:

 Pre-processing (Image algorithm)

 Segmentation process
 Feature extraction then Selection
 Classification

Generally, the results are analyzed with the aid of evaluating the proposed algorithm with the
current ones according to show the truth regarding consequences or reducing the computational
cost.

Figure 4.1 Skin Cancer detection system methodologies

4.2 PRE-PROCESSING

The advance step in expert systems ancient because of skin examination entails the achievement
concerning the tissue digital image because of pores and skin most cancers detection. The pores
and skin most cancers picture typically contain best hairs, uproar and wind bubbles. These
characteristic so much is now not piece over the cancer cell or would minimize the rigor
regarding the answer discovery then segmentation. The advance foot in imitation of do is appeal
some image processing strategies in conformity with the images. Thus pre-processing old to
referring in accordance with remove the undesirable applications about the skin or post-
processing referring after rise after the structure regarding image.

30
The accessible techniques certain so Karhunen-Loève (KL) seriously change histogram
equalization yet exceptional sorts over filter are back in conformity with attain these goals. In
addition, distinction access can sharpen the image answer or improve the precision because
segmentation. Since the image database consists on both digital photo yet dersmocopy. These
pictures are mated beyond special source yet the altar regarding the photos is non-standard. The
advance step is to resize the image to have a constant breadth 360 pixels however moving altar
over height. The 2D footsie is after quote the history maze out of the pictures. The method
chronic here is wavelet de-noise by way of two-dimensional bior3.3 wavelet. Biorthogonal (bior)
is a linear wavelet who advanced back within image reconstruction or decomposition.

Figure 4.2 True image identification of skin cancer

31
Figure 4.3 Image de-noised by wavelet

Figure 4.4 Image from median filter

32
Idea reconstruction then decomposition, configuration 4.2 (original image) indicates an unique
digital photo and Figure 4.3 setup as Idea de-noised by way of wavelet transforms. Thus the
result image addicted a “blur” image yet the detail concerning the photo do stand retained.

Uptake processing algorithm incorporate next footsie is picture in conformity with skip for
smoothing with the aid of median filter. The median filter may stand aged to keep edge, remove
the maze best with the aid of the image shooting yet recover the nice hairs. This filter yield
suggests namely easy photo as much figure 4.4.

The size of filter windows is calculated with the aid of the method out of the equation refers
according to a common volume including 768 x 512 pixels image. M then N speaks according to
the dosage about resized image.

𝑀 𝑁
n = floor x√(786) 𝑥 √512 (4.1)

The fine hair is capable in imitation of mask oversea by using median filter as like proven within
configuration 4.4. Median filtering is a nonlinear filtering technique so is known for retaining
acute adjustments between signals or for weight particularly high quality among casting off
emotive noise. One fantastic usage over median filters has been the reduction of high- frequency
or impassioned confusion into digital pictures besides the tremendous blurring then part
consumption related with linear filters. Because the median filter is nonlinear, spectral evaluation
offers baby perception of the filtering technique as like well. Replace each pixel price along the
median price regarding the mature values among the place about the pixel.

33
 Figure 4.5 Computation of mean value

4.3 SEGMENTATION PROCESS

In this stage, the greater pores and skin picture is segmented to solve the tumour beyond the
heritage (skin). Segmentation eliminates the wholesome pores and skin beside the image then
finds the region of interest. Usually the most cancers cells remains in the image since
segmentation. Segmentation old here is Threshold Segmentation. Thresholding gives a
convenient or handy course to function the segmentation regarding the basis regarding the
exclusive intensities yet shades among the foreground or heritage regions about an image. The
enter in imitation of a thresholding operation is typically a grayscale yet shade image. After
segmentation, the yield is a double image. Segmentation is realized by scanning the total image
pixel with the aid of pixel yet labeling each pixel so target then history in accordance according
to its binarized inveterate level.

Segmentation algorithms are based on some regarding couple basic properties of depth values
discontinuity yet similarity. First category is to part a image based totally about abrupt
adjustments of intensity, such namely edges between an image. Second category is based
concerning partitioning a photo between areas up to expectation are comparable according in

34
conformity with predefined criteria. Histogram Threshold approach fount underneath that
category.

The equalization on image, Histogram equalization (HE) algorithm is a smart contrast access
approach [29]. The histogram equalization can aline within Global Histogram Equalization
(GHE) then Local Histogram Equalization (LHE). GHE’s transform characteristic C(rk) usage
the histogram information on the total enter image. The downside regarding GHE is it omits the
brightness capabilities about the image; thus, the ripe level including excessive frequency
pleasure selects the mean frequency one. LHE is in a position in conformity with unravel the
hassle concerning GHE however also increase the noise. LHS employs little window up to
expectation scans though each pixel concerning the image. Only the blocks concerning pixels as
posture among its windows are ancient because of the histogram. Then increase by means of
inveterate stage mapping solely sues in accordance with the centre pixel about that window.

Threshold is some concerning the broadly methods old because image segmentation. It is useful
among discriminating foreground beyond the background. By selecting an enough beginning
charge T, the ripe degree image is converted to double image. The geminate photo should
include all concerning the crucial records about the role then shape of the objects about interest
(foreground). The skill of obtaining advance a double picture is so much that reduces the
complexity on the facts or simplifies the manner on attention yet classification. The near
common access to alter a gray-level picture after a double image is in imitation of pick out a
unaccompanied threshold value (T). Then every the skilled level values past this T charge choice
lie categorized as much fuscous (0), or those upon T charge choice keep classified namely bright
(1). The segmentation hassle will become some concerning selecting the good price because the
introduction T. A customary approach old in imitation of choose T is through analyzing the
histograms concerning the type on pictures up to expectation want according to remain
segmented. The best suit is when the histogram presents only joining dominant modes yet a clear
glen (bimodal). In certain suit the cost concerning T is selected as the delve factor within the
twain modes. The histogram based totally techniques is dependent over the success regarding the
estimating the introduction cost as separates the twins homogenous vicinity over the objective
then historical past on an image.

35
The interesting applications on melanoma are covered inside the resemble seeing that most about
the cancer cells are nodule structure. The answer shape offers crucial information because of
correct diagnosis. Many scientific purposes including asymmetry and resemble irregularity are
considered out of the border. In that thesis, onset then statistical region merging (SRM) are
carried out and compare their propriety along neural community classifier.

Figure 4.6 An ideal histogram [30]

Figure 4.7 Segmented image

36
The threshold values are thought via histogram about RGB coloration bandage (Red, Green or
Blue) regarding the images. Formal 4.6 show a perfect histogram. Statistical region merging
(SRM) algorithm is primarily based about place rising yet merging. The performance over SRM
has a higher ROI segmentation end result evaluate after sordid famous segmentation techniques .
The idea over region merging starts at a truss point and examine its four near points or pixels.
The vicinity is major now the neighbor’s pixels piece the same homogenous properties. SRM
generally employment with a statistical takes a look at to decide the merging over regions.
Configuration 4.7 indicates the result of segmentation the place gray shade represents the place
on interest.

4.4 FEATURE EXTRACTION AND SELECTION

At that stage, the vital purposes about picture data are extracted out of the segmented image. By
extracting features, the picture statistics is narrow below after a engage on features as can
characterize into Malignant then Benign melanoma. The extracted features ought to remain both
representatives concerning samples yet ample in conformity with keep classified. 2D wavelet
radically change is ancient for the feature extraction. In this system, 2-D wavelet custom is
chronic yet the stronger photo in inveterate scaled so an input.

Assume a digital image sized M x N pixels is transformed by the discrete wavelet as shown in
Figure 4.7 which produced by the level decomposition, the result of the decomposition L and H
stand for low and high frequency components. FL and FH represent low-pass and high-pass
filters. Perform discrete wavelet transform to the image. LL (0) is the original image. LH (1),
HL(1) and HH(1) are the output of high-pass filter that’s represent the horizontal details, vertical
details and diagnosing details. LL (1) represents the approximation with the same size of LH(1),
HL(1) and HH(1) that’s use to perform the second-level decomposition. The images LH(2),
HL(2) and HH(2) have finer detail than in LH(1), HL(1) and HH(1). Moreover, the image energy
is distributed according to the resolution. Each of these nodes is represent one feature, which
then can be used as an input for classification stage. The second level decomposition can
generate 16 nodes or features. Two-dimensional wavelet packet returns the coefficients in 2
dimensions matrix. The features are calculated by their mean, maximum, minimum, median,

37
standard deviation and variance. Therefore, 144 features (16 nodes x 9 features) are produced.

Figure 4.8 (a) Original image; (b) First-level decomposition; (c) Second level decomposition

4.5 CLASSIFICATION

Artificial neural network (ANN) architectures have been recognized for a number of years as a
powerful technology for solving real-world image processing problems. The primary purpose of
this special issue is to demonstrate some recent success in solving image processing problems
and hopefully to motivate other image processing researchers to utilize this technology to solve
their real-world problems.

Artificial neural community (ANN) architectures bear been identified for a quantity of years
namely a husky technology for fixing real-world photo technology problems. The most
important motive on this extraordinary difficulty is after display some current advancement
within fixing image processing problems and with a bit of luck to motivate mean image
processing researchers to turn to advantage that technological know-how in accordance with
clear up their real-world problems.

Back-propagation neural network is one of the most common neural network structures, as it is
simple and effective. Back propagation (BPN) Algorithm is used for training. There must be
input layer, at least one hidden layer and output layer. The hidden and output layer nodes adjust
the weights value depending on the error in classification.

38
Auto-associative neural network is a network instituted regarding even bulk of the input bed and
the output layer, then smaller quantity between secret layers.

4.6 CONCLUSION

This chapter reviewed the proposed algorithms between distinctive tiers about pores and skin
cancer discovery system. This system consists of pre-processing, segmentation, feature
extraction or selection, yet classification.

39
 CHAPTER 5
RESULTS AND DISCUSSION

40
 CHAPTER 5
RESULTS AND DISCUSSION
5.1 INTRODUCTION

This chapter presents and discusses the obtained effects regarding the proposed algorithm of
pores and skin most cancers detection. In order in imitation of look into the proposed helpful
steps based neural network classifier, MATLAB or its amenities is used.

5.2 SIMULATION TOOL

The algorithm was carried out in MATLAB simulation tool. MATLAB presents GPU computing
performance that is well perfect because of a hunch about features such so facts analysis, sign
processing, then image processing. All community training yet testing techniques of that treatise
was rendered in MATLAB, as offers a strong ANN toolbox. This toolbox has built-in purposes
to construct, train, yet retailer the network.

5.3 RESULT AND ANALYSIS

This section obtained results of the proposed algorithm.

5.3.1 Results of Pre-Processing

In this section, the exclusive filters concerning different noises hold been experimented after find
the better filter for improving the pixel quality. This manner present in accordance with real
identification on pores and skin cancer. Figure 5.1 and figure 5.2 show as like authentic
difference on skin cancer.
Figure 5.1 show a true detected image of skin cancer. It is clear and real image of skin cancer.
Figure 5.2 show a false image concerning skin cancer. It is evident Image but no longer
performed an actual syndrome over skin cancer.

41
Figure 5.1 True Identification of skin cancer

Figure 5.2 False Identification of skin cancer

42
 5.3.2 Results of Segmentation process

Segmentation is realized by scanning the complete photo pixel by way of pixel and labeling each
pixel as aim and history in accordance in accordance with its gray level. This book computes
segmentation by means of SRM.

Figure 5.3 Segmentation from SRM

5.3.3 Results of Feature extraction and Selection

This step of thesis intends to rank the available extracted features by attention to their impact on skin
cancer detection.

Figure 5.4 Gray image and BW image composition

43
 Figure 5.5 DWT Gray image and DWT BW image composition

These part exhibit consequences a first part on composition yet 2nd portion about decomposition
together with executed distinct wavelet transforms.

5.3.4 Result of classification

These portion exhibit penalties a first portion on arrangement but 2nd piece respecting
decomposition together with rendered awesome wavelet transform.

44
 Figure 5.6 Some training image results of detected skin cancer

5.3.4.1 Results of BNN classifier

Table 5.1 shows as a best result with highest overall accuracy is 90.2%. The best BNN is three
hidden layer with 40, 25 and 10 neurons for each hidden layer. The accuracy is increase with
number of neuron in hidden layer. However, number of hidden layer cannot improve the result
but it could reduce the probability of over-fitting.

45
5.3.4.2 Results of AANN classifier

The best AANN testing result found is 20 neurons in the first and third layer with overall
accuracy 81.5% as table 5.2 illustrated. Unlike BNN, ANN provides a stable classification result
in different number of neuron. However, when the layer 1 and layer 3 have different size of
neuron, the classifier result has a significant low accuracy diagnosing result

Table 5.1 BNN classification results with different layers

No of No of Neuron Training (%) Testing (%) Validation (%) Total (%)

Phase

1 10 82.06 56.10 64.09 75.05

1 20 99.06 61.10 67.10 84.63
1 30 99.05 53.28 78.02 87.06
1 40 98.75 50.11 78.06 88.07
2 10,5 81.34 47.3 57.80 70.17
2 20,10 98.89 51.90 75.06 85.83
2 30,20 98.70 54.28 68.80 85.04
2 40,20 97.09 52.09 78.20 89.13
3 10,8,6 93.50 61.20 70.18 86.30
3 20,12.8 98.16 63.10 77.20 88.19
3 30,20,10 98.82 61.82 71.45 88.60
3 40,25,10 98.12 62.30 82.53 89.62

46
 Table 5.2 AANN classification results with size of neurons

Layer 1 to 4 Training Validation Testing Total

10 4 10 4 86.12 58.83 70.04 78.09

10 5 10 4 81.85 55.94 69.18 75.15
20 4 20 4 89.11 59.62 67.09 78.81
20 10 20 10 90.74 60.85 69.06 80.93
30 4 30 4 90.19 58.95 73.00 82.24
30 10 30 4 87.85 54.74 63.92 77.09
40 4 40 4 88.82 62.73 69.15 80.08
40 20 40 4 90.75 53.81 62.83 78.04
40 10 30 4 40.84 40.08 40.04 39.74

47
 Figure 5.7 Neural network training tool simulations.

5.4 CONCLUSION

This section reviews the empirical consequences over proposed algorithms of distinctive tiers
regarding pores and skin most cancers detection system. In pre-processing stage, the outcomes
exhibit the higher overall performance on images. Then end result regarding segmentation show
a SRM image. In characteristic extraction stage, show a arrangement and decomposition images.
Classification basic end result is 90.2 % for back-propagation neural community yet 81.3%
because of auto-associative neural network. These outcomes are between accidents along the
effects obtained with the aid of analytical solution.

48
 CHAPTER 6

CONCLUSION AND FUTURE

WORK

49
 CHAPTER 6

CONCLUSION AND FUTURE WORK

6.1 INTRODUCTION

This chapter summarizes the research and discusses potent directions because of after lookup
between the areas.

During recent decades, the sexual intercourse regarding cruel melanoma as much the deadly
structure on pores and skin most cancers has been raised. This most cancers do keep cured
correctly condition it is detected among quickly stages. Therefore, that early analysis is an
necessary issue in imitation of reduce the mortality rates. Although there are many developments
in imaging technological know-how kind of dermoscopy, that prognosis suffers challenging
subjectivity, specifically of initial physician’s ponderabil as would remain the old entrance of
patients into dermatology community. Since the prognosis about melanoma out of efficient is
now not a convenient technique of before long stages, the dermatologist have to stay trained as
an expert. Therefore, the pc based analysis systems can also lay a best tool because of physicians
with much less experience.

An image classification system for early skin cancer discovery includes unique stages, pre or
submits processing, feature extraction yet classifier. The preprocessing resizes the photo up to
expectation improves the velocity performance or removes the superfluous characteristic such as
much the noise then great hair. Pre-processing enhances the image attribute or sharpens the
outline of the most cancers cell. Feature extraction decomposes the beneficial feature barring
scientific knowledge. The classifier use neural community measuring increase regarding predict
latter image.
The work presented a study which can be concluded that there are some possible factors of low
classification result. The image database is not feasible and too small; the variation between dermoscopy
and digital image is large. Since dermoscopy and digital image are used both in testing. The imaging
processing methods are not unique and their variation is large. However for our study with both types of
image and for two types of skin cancer, overall result is 90.2% for back-propagation neural network and
81.3% for auto-associative neural network.

50
This dissertation proposed innovative then tremendous algorithms which do enhance the overall
performance about computer-aided diagnostic structures because of melanoma detection. This
chapter covers a universal review concerning it thesis with respect after the elements of system, the
proposed algorithms, contributions yet empirical achievements.

Thesis Contribution

 First chapter is the explanation about the introduction of skin cancer (Melanoma & Non
Melanoma).
 Second chapter concludes the previous work to be done for skin cancer detection using
different type of techniques.
 Third chapter concludes basic techniques used in skin cancer detection process.
 Fourth chapter concludes about proposed methodology of thesis.
 Fifth chapter concludes final results coming from simulation tool on matlab.

6.2 FUTURE SCOPE

Following are the areas of after instruction which perform remain regarded because of further
research work.

It may want to keep excellent in accordance with provide a vast dataset include exceptional
pictures concerning the equal lesion. These similar snap shots perform be instituted beside
different imaging modalities certain as much extremely sound, dermoscopy or etc. in imitation of
think about the range of element of lesion. This may apply the one of a kind statistics in regard to
the same tumor certain as like deep over lesion, then floor concerning the blow or sordid criteria.
Thus, the acquired data would keep used to score or forecast extra accurately. On the lousy hand,
between some cases the sequential pictures done in a length concerning era would stay a proper

51
option for detection. The hybrid segmentation algorithms perhaps applied regarding pathologist
pictures in conformity with improve the segmentation results.

52
REFERENCES

53
 REFERENCES

[1] W. S. McCulloch and W. H. Pitts. A logical calculus of the ideas immanent in nervous
activity. Bulletin of Mathematical Biophysics, 5:115–133, 1943.
[2] F. Rosenblatt. The perceptron: a probabilistic model for information storage and
organization in the brain. Psychological Review, 65:386–408, 1958.
[3] M. Minsky and S. Papert. Perceptrons. Cambridge, Mass.: MIT Press, 1969.
[4] R. Anderson, A. Parrish, "The optics of human skin", The Journal of Investigative
Dermatology, Vol. 77, pp 13-19, 1981.
[5] McGovern TW, Litaker MS. Clinical predictors of malignant pigmented lesions“a
comparison of the Glasgow seven-point checklist and the American Cancer Society’s ABCDs
of pigmented lesions” J DermatolSurgOncol 1992,18:22–6.
[6] Ryszard Tadeusiewicz. Sieci neuronowe (Neural Networks). Akademicka Oficyna
Wydawnicza, Warszawa, PL, 1993.
[7] Chris M. Bishop. Neural Networks and their Applications. Review of Scientific
Instruments, 65(6):1803–1832, 1994.
[8] T. Lee, V. Ng, D. McLean, A. Coldman, R. Gallagher, J. Sale 1995, A Multi-Stage
Segmentation Method for Images of Skin Lesions, IEEE Pacific Rim Conference on
Communication, Computer and Signal Processing, pp. 602-605.
[9] Christos Stergiou and Dimitrios Siganos. Neural Networks. SURPRISE 96 Journal, 4,
1996.
[10] S. Agatonovic-Kustrin and R. Beresford. Basic concepts of artificial neural net- work
(ANN) modeling and its application in pharmaceutical research. Journal of Pharmaceutical and
Biomedical Analysis, 22:717–727, 2000.
[11] W. Stolz, O. Braun-Falco, P. Bilek, M. Landthaler, W. H. C. Burgdorf, and A. B.
Cognetta. colour Atlas of Dermatoscopy, volume 1. Blackwell, Berlin, 2 edition, March 2002
[12] P. Rubergni, G. Cevenini, M. Burroni, R. Perotti, G. Dell Eva, P. Sbano, C. Miracco, P.
Luzi, P. Tosi, P. Barbini, L. Andreassi 2002, Automated diagnosis of pigmented skin lesions,
International Journal of Cancer, Volume 101, Issue 6, pp. 576-580
[13] Caroline M Owen, Nicholas R Telfer, ”Skin Cancer”, Medicine, Volume 33, Issue 1, 1
January 2005, Pages 64-67.
[14] M. Abdullah-Al-Wadud, Md. H. Kabir, M. A. A. Dewan, O. Chae 2007, A Dynamic
Histogram Equalization for Image Contrast Enhancement, IEEE Transactions on Consumer
Electronics, Vol. 53 Issue 2, pp. 593-600
[15] M. E. Celebi, H. A. Kingravi, B. Uddin, H. Iyatomi, Y. A. Aslandogan, W. V. Stoecker,
R. H. Moss 2007, J. M. Malters, J. M. Grichnik, A. A. Marghoob, H. S. Rabinovitz, S. W.
54
Menzies 2008, Border detection in dermoscopy images using statistical region merging, Skin
Research and Technology 14, pp. 347-353
[16] U. Wollina, M. Burroni, R. Torricelli, S. Gilardi, G. Dell’Eva, C. Helm, W. Bardey 2007,
Digital dermoscopy in clinical practice: a three-centre analysis, Skin Research and Technology
Journal, Volume 13, pp. 133-142
[17] M. E. Celebi, H. A. Kingravi, B. Uddin, H. Iyatomi, Y. A. Aslandogan, W. V. Stoecker,
R. H. Moss 2007, J. M. Malters, J. M. Grichnik, A. A. Marghoob, H. S. Rabinovitz, S. W.
Menzies 2008, Border detection in dermoscopy images using statistical region merging, Skin
Research and Technology 14, pp. 347-353
[18] Science Daily, “Cancer Projected to Become Leading Cause of Death Worldwide In
2010”, http://www.sciencedaily.com/releases/2008/12/081209111516.htm, Dec 2008.
[19] Robert S. Porter, ”The Merck Manual”, July 20, 2011. & D. L. Kasper, E. Braunwald,
and A. Fauci, “Harrison’s Principles of Internal Medicine”, 17th ed., McGraw-Hill, New York
(2008).
[20] H. Demuth, M. Beale, M. Hagan 2008, Neural Network Toolbox User's Guide, The
Mathworks, Version 6.
[21] Dhawan Ap, Dalessandro B, Patwardhan S,Mullani N,” Multispectral optical imaging of
skin-lesions for detection of malignant melanomas”, Annual International Conference of
Engineering in Medicine and Biology Society, EMBC, 2009.
[22] T. DiChiara. “Pictures of Moles and Melanoma skin cancer – Learn o tell the difference
with pictures.”, Sept. 2010.
[23] Daniel Ruiz, Vicente Berenguer, Antonio Soriano and Belén Sánchez “A decision
support system for the diagnosis of melanoma: A comparative approach” 13 June 2011.
[24] Cancer Research UK, “Cancer Worldwide – the global picture”,
http://www.cancerresearchuk.org /cancer-info/cancerstats/world/the-global-picture /, Sept. 2012.
[25] Arul N, Cho YY,” A Rising Cancer Prevention Target of RSK2 in Human Skin Cancer”,
2013 Aug 5; 3:201. doi: 10.3389/fonc.2013.00201. e Collection 2013.
[26] H American Cancer Society. Skin Cancer Prevention and Early Detection, in: Annual
Report. (Atlanta, GA: American Cancer society) (2013).
[27] Randy Gordon,” Skin Cancer: An Overview of Epidemiology and Risk Factors”,
Seminars in Oncology Nursing, Vol 29, No 3 (August), 2013: pp 160-169.
[28] Maurya R, Surya K.S,"GLCM and Multi Class Support Vector Machine based
Automated Skin Cancer Classification, "IEEE journal, vol 12, 2014.
[29] A.A.L.C. Amarathunga,” Expert System For Diagnosis Of Skin Diseases”, International
Journal Of Scientific & Technology Research, Volume 4, Issue 01, 2015.

55
 APPENDIX
MATLAB Simulation
Phase1.m

clear
close all
clc

global CC
[imagesName] = rdir('skin_data\*\*\*.jpg');

verbose = true;
k=0;
h0= figure(1);
h1= figure(100);
h2= figure(200);
h3= figure(300);
h4= figure(400);
h5= figure(500);
h6= figure(600);
h7= figure(700);
h8= figure(800);
[~,idx]=sort(randi(5*numel(imagesName),numel(imagesName),1));
ktrain=0;
ktest=0;
for img=1:numel(imagesName)

image_orig = imresize(imread(imagesName(idx(img)).name),[256,256]);

if size(image_orig,3)==3
imgray=rgb2gray(image_orig);
else
imgray=image_orig;
end

image_orig=KL_histrogram_equilizer(image_orig);
image=double(image_orig);

for i=1:size(image,3)
x=image(:,:,i);
y = denoising_dwt(x);
image(:,:,i) = medfilt2(y,[5 5]);
end

56
 % Choose different scales
% Segmentation parameter Q; Q small few segments, Q large may segments
Qlevels=2.^(6:-1:6);
% This creates the following list of Qs [256 128 64 32 16 8 4 2 1]

% Creates 9 segmentations
[maps,images]=srm(image,Qlevels);
% And plot them
[Iedge, precision]=srm_plot_segmentation(images,maps,h1,h2,h3);
% figure(h6);

BW=Iedge==precision;

if size(image_orig,3)==1
imgx=image;
imgx(~BW)=255;
else
for i=1:size(image_orig,3)
imgxy=image(:,:,i);
imgxy(~BW)=255;
imgx(:,:,i)=imgxy(:,:,1);
end

end
min_th=50;

[Seg,BW,flag]=enclosed_region(imgray,Iedge,precision,min_th);

if flag
figure(h0),imshow(image_orig);
figure(h6);

figure(h4),imshow(uint8(imgx));
figure(h6);

figure(h5),imshow(uint8(Seg));
figure(h6);

figure(h6),imshow(uint8(double(image_orig).*double(repmat(BW,1,1,3))));
figure(h6);

[row,col] = find(BW);
% ic=floor(sum(row)/length(row));

57
% jc=floor(sum(col)/length(col) );
%
% ly=ic-100;ry=ic+100;
% lx=jc-100;rx=jc+100;
lx=min(col(:))-5;
ly=min(row(:))-5;
rx=max(col(:))+5;
ry=max(row(:))+5;
if lx<1
lx=1;
end
if ly<1
ly=1;
end
if rx> size(BW,2)
rx=size(BW,2);
end
if ry> size(BW,1)
ry=size(BW,1);
end

BWx=BW(ly:ry,lx:rx);
imgrayx=imgray(ly:ry,lx:rx);

BWx=imresize(BWx,[64,64]);
imgrayx=imresize(imgrayx,[64,64]);

figure(h7),imshow(uint8(imgrayx));
figure(h6);

figure(h8),imshow(uint8(255.*BWx));
figure(h6);

name=imagesName(idx(img)).prepath;
name=name(10:end);
cl=name(2:9);

Sname=imagesName(idx(img)).Sname;
Sname=Sname(1:end-3);
if img>0.75*numel(imagesName)

name=['test' name];

ktest=ktest+1;
Data_Test(ktest).img=image_orig;

58
 Data_Test(ktest).BWx=BWx;
Data_Test(ktest).imgrayx=imgrayx;
test_class(ktest)=all(cl=='melanoma');

else
name=['train' name];
ktrain=ktrain+1;
Data_Train(ktrain).img=image_orig;
Data_Train(ktrain).BWx=BWx;
Data_Train(ktrain).imgrayx=imgrayx;
Data_Train(ktrain).Class=all(cl=='melanoma');

end

mkdir(name);
imwrite(BW,sprintf('%s/%s_BW.bmp',name,Sname),'bmp');
imwrite(BWx,sprintf('%s/%s_BWx.bmp',name,Sname),'bmp');

imwrite(uint8(image_orig),sprintf('%s/%s_orig.jpg',name,Sname),'jpg');

imwrite(uint8(imgray),sprintf('%s/%s_gray.jpg',name,Sname),'jpg');
imwrite(uint8(imgrayx),sprintf('%s/%s_grayx.jpg',name,Sname),'jpg');

imwrite(uint8(imgx),sprintf('%s/%s_srm.jpg',name,Sname),'jpg');
end
end
save('DATA.mat','Data_Train','Data_Test','test_class');

KL_histrogram_equilizer.m

function imout=KL_histrogram_equilizer(I)
I=double(I);
imout=I;
[h w dd]=size(I);
m=1;
for ki=1:dd
for i=1:8:h
for j=1:8:w
for x=0:7
for y=0:7
img(x+1,y+1)=I(i+x,j+y,ki);

59
 end
end
k=0;
for l=1:8
img_expect{k+1}=img(:,l)*img(:,l)';
k=k+1;
end
imgexp=zeros(8:8);
for l=1:8
imgexp=imgexp+(1/8)*img_expect{l};%expectation of E[xx']
end
img_mean=zeros(8,1);
for l=1:8
img_mean=img_mean+(1/8)*img(:,l);
end
img_mean_trans=img_mean*img_mean';
img_covariance=imgexp - img_mean_trans;
[v{m},d{m}]=eig(img_covariance);
temp=v{m};
m=m+1;
for l=1:8
v{m-1}(:,l)=temp(:,8-(l-1));
end
for l=1:8
trans_img1(:,l)=v{m-1}*img(:,l);
end
for x=0:7
for y=0:7
transformed_img(i+x,j+y)=trans_img1(x+1,y+1);
end
end
mask=[1 1 1 1 1 1 1 1
11111111
11111111
11111111
11111111
11111111
11111111
1 1 1 1 1 1 1 1 ];
trans_img=trans_img1.*mask;
for l=1:8
inv_trans_img(:,l)=v{m-1}'*trans_img(:,l);
end
for x=0:7
for y=0:7
inv_transformed_img(i+x,j+y)=inv_trans_img(x+1,y+1);

60
 end
end

end
end
imout(:,:,ki)=inv_transformed_img;
end
% figure,
% imshow(transformed_img);
% figure,
% imshow(inv_transformed_img);
%

Denoising_dwt.m

function y = denoising_dwt(x)

windowsize = 7;
windowfilt = ones(1,windowsize)/windowsize;

% Number of Stages
L = 6;
% forward transform
W = dwt2D(x,L,'bior3.3');

% Noise variance estimation using robust median estimator..

tmp = W{1}{3};
Nsig = median(abs(tmp(:)))/0.6745;

for scale = 1:L-1

for dir = 1:3

% noisy coefficients
Y_coefficient = W{scale}{dir};

% noisy parent
Y_parent = W{scale+1}{dir};

% extent Y_parent to make the matrix size be equal to Y_coefficient

Y_parent = expand(Y_parent);
% Signal variance estimation

Wsig = conv2(windowfilt,windowfilt,(Y_coefficient).^2,'same');
Ssig = sqrt(max(Wsig-Nsig.^2,eps));

61
 % Threshold value estimation
T = sqrt(3)*Nsig^2./Ssig;

% Bivariate Shrinkage
W{scale}{dir} = bishrink(Y_coefficient,Y_parent,T);

end
end

% Inverse Transform
y = idwt2D(W,L,'bior3.3');

bishrink.m

function [w1] = bishrink(y1,y2,T)

R = sqrt(abs(y1).^2 + abs(y2).^2);
R = R - T;
R = R .* (R > 0);
w1 = y1 .* R./(R+T);

SRM

% Statistical Region Merging

%Segmentation parameter Q; Q small few segments, Q large may segments

function [maps,images]=srm(image,Qlevels)

% Smoothing the image, comment this line if you work on clean or synthetic images

h=fspecial('gaussian',[5 5],1);
image=imfilter(image,h,'symmetric');

%smallest_region_allowed=10;

size_image=size(image);
n_pixels=size_image(1)*size_image(2);
smallest_region_allowed=n_pixels/500;
% Compute image gradient

62
[Ix,Iy]=srm_imgGrad(image(:,:,:));
Ix=max(abs(Ix),[],3);
Iy=max(abs(Iy),[],3);
normgradient=sqrt(Ix.^2+Iy.^2);

Ix(:,end)=[];
Iy(end,:)=[];

[~,index]=sort(abs([Iy(:);Ix(:)]));

n_levels=numel(Qlevels);
maps=cell(n_levels,1);
images=cell(n_levels,1);
im_final=zeros(size_image);

map=reshape(1:n_pixels,size_image(1:2));
% gaps=zeros(size(map)); % For future release
treerank=zeros(size_image(1:2));

size_segments=ones(size_image(1:2));
image_seg=image;

%Building pairs
n_pairs=numel(index);
idx2=reshape(map(:,1:end-1),[],1);
idx1=reshape(map(1:end-1,:),[],1);

pairs1=[ idx1;idx2 ];
pairs2=[ idx1+1;idx2+size_image(1) ];

for Q=Qlevels
iter=find(Q==Qlevels);

for i=1:n_pairs
C1=pairs1(index(i));
C2=pairs2(index(i));

%Union-Find structure, here are the finds, average complexity O(1)

while (map(C1)~=C1 ); C1=map(C1); end
while (map(C2)~=C2 ); C2=map(C2); end

% Compute the predicate, region merging test

g=256;
logdelta=2*log(6*n_pixels);

dR=(image_seg(C1)-image_seg(C2))^2;

63
 dG=(image_seg(C1+n_pixels)-image_seg(C2+n_pixels))^2;
dB=(image_seg(C1+2*n_pixels)-image_seg(C2+2*n_pixels))^2;

logreg1 = min(g,size_segments(C1))*log(1.0+size_segments(C1));
logreg2 = min(g,size_segments(C2))*log(1.0+size_segments(C2));

dev1=((g*g)/(2.0*Q*size_segments(C1)))*(logreg1 + logdelta);
dev2=((g*g)/(2.0*Q*size_segments(C2)))*(logreg2 + logdelta);

dev=dev1+dev2;

predicat=( (dR<dev) && (dG<dev) && (dB<dev) );

if (((C1~=C2)&&predicat) || xor(size_segments(C1)<=smallest_region_allowed,
size_segments(C2)<=smallest_region_allowed))
% Find the new root for both regions
if treerank(C1) > treerank(C2)
map(C2) = C1; reg=C1;
elseif treerank(C1) < treerank(C2)
map(C1) = C2; reg=C2;
elseif C1 ~= C2
map(C2) = C1; reg=C1;
treerank(C1) = treerank(C1) + 1;
end

if C1~=C2
% Merge regions
nreg=size_segments(C1)+size_segments(C2);

image_seg(reg)=(size_segments(C1)*image_seg(C1)+size_segments(C2)*image_seg(C2))/nreg;

image_seg(reg+n_pixels)=(size_segments(C1)*image_seg(C1+n_pixels)+size_segments(C2)*im
age_seg(C2+n_pixels))/nreg;

image_seg(reg+2*n_pixels)=(size_segments(C1)*image_seg(C1+2*n_pixels)+size_segments(C
2)*image_seg(C2+2*n_pixels))/nreg;
size_segments(reg)=nreg;
end
end
end

% Done, building two result figures, figure 1 is the segmentation map,

% figure 2 is the segmentation map with the average color in each segment

64
 while 1
map_ = map(map) ;
if isequal(map_,map) ; break ; end
map = map_ ;
end

for i=1:3
im_final(:,:,i)=image_seg(map+(i-1)*n_pixels);
end
images{iter}=im_final;

[clusterlist,~,labels] = unique(map) ;
labels=reshape(labels,size(map));
nlabels=numel(clusterlist);
maps{iter}=map;

bgradient = sparse(boundarygradient(labels, nlabels, normgradient));

bgradient = bgradient - tril(bgradient);
idx=find(bgradient>0);
[~,index]=sort(bgradient(idx));
n_pairs=numel(idx);
[xlabels,ylabels]=ind2sub([nlabels,nlabels],idx);
pairs1=clusterlist(xlabels);
pairs2=clusterlist(ylabels);
end

----------------------------------------------------------------------
SRM_plot_segmentation.m

function [Iedge, precision]=srm_plot_segmentation(imseg,mapList,h1,h2,h3)

precision=numel(mapList);
Iedge=zeros([size(imseg{1},1),size(imseg{1},2)]);

quick_I1 = cell(precision,1);
quick_I2 = cell(precision,1);

for k=1:precision
map=reshape(mapList{k},size(Iedge));
quick_I1{k} = srm_randimseg(map) ;

65
 quick_I2{k} = imseg{k} ;
figure(h1);vl_tightsubplot(precision, k) ;
imagesc(quick_I1{k});axis off;
figure(h2);vl_tightsubplot(precision, k) ;
imagesc(uint8(quick_I2{k}));axis off;
borders = srm_getborders(map);
Iedge(borders) = Iedge(borders) + 1;
end

figure(h3);
Iedge=precision-Iedge;
imshow(Iedge,[0 precision]);

---------------------------------------------------------------

Enclosed_region.m

function [Seg,BW,flag]=enclosed_region(image,Iedge,precision,min_th)
global CC
[h,w]=size(image);
BW=Iedge==precision;
CC = bwlabel(BW, 4);
%figure,imshow(255*CC./max(max(CC)));
BWx=zeros(size(BW));
BWx(:,1)=1;
BWx(1,:)=1;
BWx(:,end)=1;
BWx(end,:)=1;

idx=find(BWx==1);

for j=1:length(idx)

i=idx(j);
if BW(i)~=0
val=CC(i);
idxx=[];
idxx=CC==val;
BW(idxx)=0;
CC(idxx)=-1;
% figure,imshow(255*BW);
end
end

66
%CC = bwlabel(BW, 4);

valX=[];
for i=0:max(max(CC))
idx=[];
idx=CC==i;
mX=mean(image(idx));
valX(i+1)=mX;

end

[valX,idxx]=sort(valX,'ascend');
idx=zeros(size(CC))==1;

i=1;
while i<=max(max(CC)) && ( ( valX(i)-valX(1) )<min_th || valX(i)< mean(valX))

idr= CC==(idxx(i)-1);
%figure, imshow(idr);

idx(idr)=1;
i=i+1;

end

% [valX',idxx']

BW=zeros(size(BW));
BW(idx)=1;
BW = bwmorph(BW,'majority',10);

CC = bwlabel(BW, 4);
if max(max(CC))>1
for i=0:max(max(CC))
idx=[];
idx=CC==i;
if sum(idx)< h*w/2500;
BW(idx)=0;
end

end
end
if sum(sum(BW))==w*h | sum(sum(BW))==0

67
 flag=0;
Seg=0;
else
flag=1;
Seg=image;
idx=BW>0;
mX=mean(image(idx));
Seg(idx)=mX;

end

%figure,imshow(255*BW);
end

68
 LIST OF PUBLICATIONS
Shanu Gaura, Farah Shan Khan “ADVANCED TECHNIQUE FOR MELANOMA AND
NONMELANOMA SKIN CANCER DETECTION USING NUERAL NETWORK”
International Journal of Engineering and Technical Research (IJETR) ISSN: 2321-0869 (O)
2454-4698 (P), Volume-8, Issue 5, May 2018

Shanu Gaura, Farah Shan Khan “ADVANCED TECHNIQUE FOR MELANOMA SKIN
CANCER DETECTION USING ARTIFICIAL NEURAL NETWORK: A SURVEY”
International Journal of Engineering and Technical Research (IJETR) ISSN: 2321-0869 (O)
2454-4698 (P), Volume-8, Issue 5, May 2018

69
 CURRICULUM VITAE

I (Shanu Gaura) have received my B.Tech degree in Information Technology from Uttar Pradesh
Technical University, Lucknow in 2015. I am currently pursuing M.Tech in Computer Science
and Engineering from Kanpur Institute of Technology (KIT), Kanpur affiliated to Dr. A. P. J.
Abdul Kalam Technical University, Lucknow, Uttar Pradesh.

Currently I am doing my dissertation on Advanced Technique for Melanoma and Non-melanoma

Skin Cancer Detection Using Neural Network.

My area of interest is Digital Image Processing, Algorithm, and Data Structure, Discrete
Mathematics.

70