Upload
dangminh
View
214
Download
1
Embed Size (px)
Citation preview
Image Processing for Skin Cancer Detection: Malignant Melanoma Recognition
Karen Cheung
A thesis submitted in conformity with the requirements
for the degree of Master of Applied Science
Graduate Depart ment of Electrical and Computer Engineering
University of Toronto
@Copyright by Karen Cheung 1997
Bibliographie Services services bibliographiques 395 Wellington Street 395, rue Wellington Ottawa ON K1A ON4 Ottawa O N KI A ON4 Canada Canada
Your lile Votre référence
Our fi18 Notre rddmnce
The author has granted a non- exclusive licence allowing the National Library of Canada to reproduce, loan, distribute or sel1 copies of this thesis in microfom, paper or electronic formats.
The author retains ownership of the copyright in this thesis. Neither the thesis nor substantial extracts from it may be printed or othemise reproduced without the author's permission.
L'auteur a accordé une licence non exclusive permettant à la Bibliothèque nationale du Canada de reproduire, prêter, distribuer ou vendre des copies de cette thèse sous la forme de microfiche/film, de reproduction sur papier ou sur fomiat électronique.
L'auteur conserve la propriété du droit d' auteur qui protège cette thèse. Ni la thèse ni des extraits substantiels de celle-ci ne doivent être imprimés ou autrement reproduits sans son autorisation.
V V
Malignant Melanoma Recognition A thesis for the Degree of Mizster of Applied Science, 1997
by
Karen Cheung
Depart ment of Electrical and Cornputer Engineering
University of Toronto
Abstract
In order t o achieve an effective way to identify malignant melanoma
stage without performing any unnecessary skin biopsies, digital image
at an early
analysis of
the images is investigated. For the detection of malignant melanoma, appropriate
analyses are done on the tumor images according to the clinical characteristics that
early melanoma possesses. In the last stage, classification is done based on the results
obtained from the above mentioned analyses. The tumors are classified as either
"potential malignant melanoma" or "non-melanoma". This can reduce the time spent
and pain received by the patients in detecting early malignant melanoma. The major
purpose of this thesis is to help in identifying early rnalignant melanoma and aid in
early diagnosis to help to reduce the death rate caused by this deadliest disease.
1 would like to thank Professor A. N. Venetsanopoulos for his excellent guid-
ance, supervision, and continued support provided throughout this research. 1 would
also wish to thank Professor A. Banerjea, Professor R. M. Iravani and Professor
E. S. Sousa for serving in the examination cornmittee. Invaluable suggestions and
timely help given by Dr. K. N. Plataniotis has been greatly appreciated. The help
with the computer system and related issues provided by Dimitrios Androutsos and
Lowe11 Winger is sincerely acknowledged. Finally, 1 would like to thank Mrs P. Acker-
man for encouragement, help and revision of the thesis. And also Dr. M. Mackenzie
for providing me information on skin cancer.
iii
Contents
1 Introduction
1.1 Characteristics of Malignant Melanoma and other benign pigmented
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . lesions 2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Objective 5
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 DetectionProcess 5
. . . . . . . . . . . . . . . . . . . . . 1.4 Images of Malignant Melanorna 6
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Thesis Outline 7
Overview of Techniques Currently in Use
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Introduction
. . . . . . . . . . . . . . . . . . . . . 2.2 Preprocessing of Tumor Images
2.3 A: Detection of Asymmetry in Skin Tumors . . . . . . . . . . . . . . 2.4 B: Detection of Skin Turnor Boundary Irregularities in Colour Images
. . . . . . . . . 2.5 C: Detection of Variegated Colouring in Skin Tumors
. . . . . . . . . . . . . . . . . . . . . . . . . . 2.6 Texture in Skin Image
. . . . . . . . . . . . . . . . . . 2.7 Identification of Malignant Melanoma
3 Preprocessing and Colour Variegation Analysis 26
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Introduction 26
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Colours 26
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Colour Spaces 27
. . . . . . . . . . . . . . . . . . . . . . . . 3.3.1 RGB Colour Mode1 28
. . . . . . . . . . . . . . . . . . . . . . . . 3.3.2 HSI Colour Mode! 28
. . . . . . . . . . . . . . . . . . . . . 3.3.3 L*u*u* or L*a*b* Space 30
. . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Image Segmentation 31
. . . . . . . . . . . . . . . . . . . . . 3.4.1 Gray Level Thresholding 32
. . . . . . . . . . . . . . . . . . . 3.4.2 Region-Based Segmentation 33
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Preprocessing 34
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5.1 Filtering 35
. . . . . . . . . . . . . . . . . . . . . . . . . . 3.5.2 Median Filters 35
. . . . . . . . . . . . . . . . . . . . . 3.5.3 Vector Directional Filter 36
. . . . . . . . . . . . . . . . . . . . . . 3.5.4 Modified Median Filter 37
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5.5 Results 39
. . . . . . . . . . . . . . . . . . . . . . . . . 3.5.6 Tumor Extraction 39
. . . . . . . . . . . . . . . . . . . . . . . 3.6 Colour Variegation Analysis 41
. . . . . . . . . . . . . . . . . . . . . 3.6.1 Gray-Scale Thresholding 41
. . . . . . . . 3.6.2 Vector Directional Segmentation in RGB Space 42
. . . . . . . . . . . . . . . . . . . . 3.6.3 Segmentation in HSI Space 42
. . . . . . . . . . . . 3.6.4 Segmentation in L*u*v* or L*a*b* Space 43
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7 Observations 44
. . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7.1 Preprocessing 44
. . . . . . . . . . . . . . . . . . . 3.7.2 Colour Variegation Analysis 45
4 Border Irregularity Analysis. Asymmetry Analysis and Analysis of
a Common Mole 61
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Introduction 61
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Edge Detection 62
. . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Gray-Scale Images 62
. . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.2 Colour Images 62
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.3 Results 64
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Border Irregularity 66
. . . . . . . . . . . . . . . . . . . . 4.3.1 Boundary Representation 67
. . . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Goodness of Shape 74
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.3 Results 77
4.4 Asymmctry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.4.1 Asymmetry Index . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.5 Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.5.1 Border Irregularity Analysis . . . . . . . . . . . . . . . . . . . 80
4.5.2 Asymmetry Analysis . . . . . . . . . . . . . . . . . . . . . . . 82
4.6 Analysis Results of Non-Melanoma Mole . . . . . . . . . . . . . . . . 83
5 Conclusions 97
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Discussion 97
5.2 Problems with the proposed methods . . . . . . . . . . . . . . . . . . 99
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Future Researcli 100
5.4 Benefits of Digital Imaging in Derrnatology . . . . . . . . . . . . . . . 100
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5 Conclusion 100
List of Tables
3.1 Filters Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
. . . . . . . . . . . . . . . . . . . . . . . 3.2 Gray Scale Thresliold values 40
. . . . . . . . . . . . . . . . . . . . 3.3 Angle and Magnitude Thresholds 40
. . . . . . . . . . . . . . . . . . . . . 3.4 Possible regions of segmentation 42
. . . . . . . . . . . . . . . . . . . . . . 3.5 Number of Segmented regions 44
Number of Vertices and Threshold used on the Approximated Polygon 69
. . . . . . . . . . . . . . . . . . . . . . . . . . Feature Interpretation 71
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Centoids 72
. . . . . . . . . . . . . . . . . . . . . . . . . . . . Irregularity Indices 78
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Order of shapes 78
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . Orientation Angle 80
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . Asymmetry Index 80
Border Irregularity Measures and Asymmetry Index . . . . . . . . . . 83
5.1 Summary of the Classification of Malignant Melanoma . . . . . . . . 97
vii
List of Figures
1.1 Structure of the detection of Malignant Melanoma . . . . . . . . . . . 1.2 (left) Superficial Spreading and (right) Nodular Malignant Melanoma
1.3 (left) Lentigo Maligiiant Melanoma and (right) Seborrheic Keratosis
Tumor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . 1.4 Canadian Cancer Society Alert Bookmark
2.1 New Colour Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 The Colour Triangle regions defined by 10 degree increments on Angle
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A and Angle B
2.3 Smooth block (2772m2) from a skin tumor . Luminance for each pixel,
range O -255. is displayed on the vertical axis. with pixel locations
within the 2 - mm2 block displayed on the x and y axes . . . . . . . . . . . . . . . . . . . . . . . 2.4 Rough block (2mm2) from a skin tumor
2.5 Relative spatial locations of the eight members of the circular neigh-
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . bour set
2.6 The hierarcliy of the plans used in the high-level analysis . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 RGB Colour Mode1
. . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 HSI Colour Triangle
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 HSI Colour Space
. . . . . . . . . . . . . . . . . . . . . . 3.4 Histogram with distinct peaks
3.5 Histogram with more than 2 peaks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Modified Median Filter
. . . . . . . . . . . . . . . . . . 3.7 Recusive Modified Generalized Filter
viii
3.8 Filtered Colour images (left) Super (right) Node . . . . . . . . . . . . 47
3.9 Filtered Colour images (left) Lent (right) Sebor . . . . . . . . . . . . 47
3.10 Tumor Extraction: Gray Scale Thresholding Histogram (left) Super
(right) Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.11 Tumor Extraction: Gray Scale Thresholding Histogram (left) Lent
(right) Sebor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.12 Turner Extraction: Gray Scale Thresholding Extracted Tumor (left)
Super (right) Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.13 Tumor Extraction: Gray Scale Thresholding Extracted Tumor (left)
Lent (right) Sebor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.14 Tumor Extraction: Vector Directional Segmentation Colour Histogranz
(left) Super (right) Node . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.15 Tumor Extraction: Vector Directional Segmentation Colour Histogram
(left) Lent (right) Sebor . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.16 Tumor Extraction: Vector Directional Segmentation Extracted Colour
Tumor (left) Super (right) Node . . . . . . . . . . . . . . . . . . . . . 51
3.17 Tumor Extraction: Vector Directional Segmentation Extracted Colour
Tumor (left) Lent (riglit) Sebor . . . . . . . . . . . . . . . . . . . . . 51
3.18 Colour Variegation Analysis: Gray-Scale Thresholding Histogram (left)
Super (right) Node . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.19 Colour Variegation Analysis: Gray-Scale Thresholding Histogram (left)
Lent (right) Sebor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.20 Colour Variegation Analysis: Gray-Scale Thresholded regions for Super 53
3.21 Colour Variegation Analysis: Gray-Scale Thresholded regions for Node 53
3.22 Colour Variegation Analysis: Gray-Scale Thresholded regions for Lent 54
3.23 Colour Variegation Analysis: Gray-Scale Thresholded regions for Lent 54
3.24 Colour Variegation Analysis: Gray-Scale Thresholded regions for Lent 55
3.25 Colour Variegation Analysis: Gray-Scale Thresholded regions for Lent 55
3.26 Colour Variegation Analysis: Super Colour regions . . . . . . . . . . . 56
3.27 Colour Variegation Analysis: Super Colour regions . . . . . . . . . . . 56
O.LU V U L U U ~ V ~ L I C ~ ~ L ~ L W I I r u ~ a ~ y ~ m . I Y V U C UVIUUI l G ~ l U 1 1 3 . . . . . . . . . . . 0 1
3.29 Colour Variegation Analysis: Node Colour regions . . . . . . . . . . . 57
3.30 Colour Variegation Analysis: Lent Colour regions . . . . . . . . . . . 58
3.31 Colour Variegation Analysis: Lent Colour regions . . . . . . . . . . . 58
3.32 Colour Variegation Analysis: Sebor Colour regions . . . . . . . . . . . 59
3.33 Colour Variegation Analysis: Sebor Colour regions . . . . . . . . . . . 59
3.34 Colour Variegation Analysis: Lab Segmentation (left) Super (right) Node 60
3.35 Colour Variegation Analysis: Lab Segmentation (left) Lent (right) Sebor 60
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Sobel Operators 62
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Window 63
4.3 Structuring element of the thinning process . . . . . . . . . . . . . . . 65
4.4 Structuring element of the 2nd thinning process . . . . . . . . . . . . 66
4.5 Structuring element of the 3rd thinning process . . . . . . . . . . . . 66
4.6 4- and 8- directional chain codes . . . . . . . . . . . . . . . . . . . . . 67
4.7 The first few iterative technique for generating a polygonal approxima-
tion to a curve . The initial nodes 1 and 1' are chosen arbitrary . Nodes
2 and 2' are generated. since neither satisfies the distance criterion . At
this point the approximation is given by 1.2.1'.2'.1 . Shen each of the
curve segments can be split independently . For example. segment 1-2'
is split into 1-3 and 3-2'; segment 1-3 into 1-4 and 4.3. and so on until
the optimization criterion is satisfied . . . . . . . . . . . . . : . . . . . 68
4.8 a) Circle's Signature b) Square's Signature . . . . . . . . . . . . . . . 72
4.9 Different shapes with the same area and perimeter . . . . . . . . . . . 75
4.10 Edge Detection using Sobel Operator (left) Superficial Spreading and
(right) Nodular Malignant Melanoma . . . . . . . . . . . . . . . . . 85
4.11 Detection using Sobel Operator (left) Lentigo Malignant Melanoma
and (right) Seborrheic Keratosis Mole . . . . . . . . . . . . . . . . . . 85
4.12 Preprocessed and Edge Detection (left) Superficial Spreading and (right)
Nodular Malignant Melanoma . . . . . . . . . . . . . . . . . . . . . . 86
A . L V A A V r L V V u u V u u u-iu -u 0" Y V V V V V I V I I \ A V L V / U V I I V A O' "'-" "'-'-"'-'-'-
and (right) Seborrheic Keratosis Mole . . . . . . . . . . . . . . . . . . 4.14 First Thinning (left) Superficial Spreading and (right) Nodular Malig-
nant Melanoma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.15 First Thinning (left) Lentigo Malignant Melanoma and (right) Sebor-
rheic Keratosis Mole . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.16 2nd Thinning (left) Superficial Spreading and (right) Nodular Malig-
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . nant Melanoma
4.17 2nd Thinning (left) Lentigo Malignant Melanoma and (right) Sebor-
rheic Keratosis Mole . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.18 Polygonal Approximation (Mt) Superficial Spreading and (right) Nodu-
lar Malignant Melanoma (dashed line: approximated tumor; solid line:
original tumor) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.19 Polygonal Approximation (left) Lentigo and (right) Sebor (daslied line:
approximated tumor; solid line: original turnor) . . . . . . . . . . . . 4.20 Incremental Curvature (left) Superficial Spreading and (riglit) Nodular
Malignant Melanoma . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.21 Incrernental Curvature (left) Lentigo Malignant Melanoma and (right)
Seborrheic Keratosis Mole . . . . . . . . . . . . . . . . . . . . . . . .
4.22 Signature (left) Superficial Spreading and (right) Nodular Malignant
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Melanoma
4.23 Signature (left) Lentigo Malignant Melanoma and (riglit) Seborrheic
Keratosis Mole . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.24 Power Spectrum (left) Superficial Spreading and (right) Nodular Ma-
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . lignant Melanoma
4.25 Power Spectrum (left) Lentigo M alignant Melanoma and (right ) Seb-
. . . . . . . . . . . . . . . . . . . . . . . . . . orrheic Keratosis Mole
4.26 Curvature Histogram (left) Superficial Spreading and (right) Nodular
. . . . . . . . . . . . . . . . . . . . . . . . . . . Malignant Melanoma
4.2 1 Lurvature nistogram tien) Lenugo lvlarignant luelanoma ana (ngnr;)
. . . . . . . . . . . . . . . . . . . . . . . . Seborrheic Keratosis Mole 93
. . . . . . . . 4.28 (left) Original Scler Image (right) Filtered Scler Image 94
. . . . . 4.29 (left) Extracted Scler Tumor (right) Segrnented Scler Turnor 94
. . . . . . . . . . . . . . . . . . . . . . . . . 4.30 Sclerosing Tumor border 95
4.31 (left) Polygonal Approximation (right) Incrernental Curvature Plot . 95
. . . . . . . . . . . . . 4.32 (left) Signature Plot (right) Fourier Spectrum 96
. . . . . . . . . . . . . . . . . . . . . . . . . . . 4.33 Curvature Histogram 96
xii
Chapter 1
Introduction
Among al1 of the different forms of skin cancer, Malignant Melanoma is the Ieading
cause of death nowadays. The incidence of melanoma has doubled during the last
20 years [62]. Fair-skinned people, who burn easily and rarely tan, are most a t risk.
Major causes of this disease are 1) The depletion of ozone layer caused by pollution 1261
and 2) The excessive exposure to Sun [26] . In order to reduce the death rate due to
malignant melanoma, it is necessary to diagnose it a t an early stage. The less mature
the malignant melanoma is, the earlier the surgical treatment, the greater the survival
rate. Skin biopsies a t an early stage are necessary to identify Malignant Melanoma.
Due to the rise of medical costs, especially the cost of skin biopsies, some better ways
to identify malignant melanoma a t an early stage are needed without increasing the
number of skin biopsies. The role of digital irnaging cornes into sight here; it was
hoped that image analysis could help in identifying early malignant melanoma and
aid in early diagnosis to help to reduce the death rate caused by this deadly disease.
1.1 Characteristics of Malignant Melanoma and
other benign pigmented lesions
Human skin is made up of the epidermis (the top layer) and dermis (the inner layer).
In the epidermis layer, there exist melanocytes which are cells that contain melanin,
and i t is the melanin which gives colour to the skin. Melanoma is often called cuta-
neous melanorna or malignant melanoma; it is a skin disease in which the cancer cells
are found in the melanocytes of the epiderrnis.
There are five types of malignant melanoma [29] which are classified by their
histologic features and are listed according to their frequency of occurrence:
0 Superficial Spreading Malignant Melanoma(SSM) is the most cornmon
type of malignant melanoina. I t may occur on any part of the body and
is usually greater than 0.5cm in diameter. I t is a tumor which is elevated
and spreads laterally. SSM usually develops as an asyrnmetric plague with
variation in colour as pigment pattern and irregularity or notching of bor-
ders. An example is shown in Figure 1.2 1291. It can be seen that this
tumor has a highly irregular border with fingers stretching out on the left.
It also has got variable degrees of pigmentation and a nodular amelanotic
component at the lower right.
Nodular Malignant Melanoma is the next frequent type, it is less common
but more malignant. I t is a raised papule or nodule, sometimes ulcerated.
The outline of the lesion rnay be irregular and its colour varied. Very
often, i t will have a well-defined border and symmetry in contrat to other
melanomas. An example is shown in Figure 1.2 [29].
Lentigo Malignant Melanoma is represented by varying admixtures of pink,
gray, blue and white. The borders are frequently highly irregular and
notched. The overall size rnay range from 1.0 to 20.0 cm or larger. Malig-
nant change is recognized by thickening and the development of discrete
tumor nodules. This is an invasive proliferation of malignant melanocytes
which arise in lentigo maligna. An esample is shown in Figure 1.3 [29].
Acral Lentiginous Malignant Melanoma is a very rare tumor. I t usually
arises in an acral location or on a mucous membrane and is initially flat and
irregular, but soon becomes raised and subsequently nodular. Colouration
is less varied than SSM, but borders may show marked irregularity and
notching. The size ranges from 0.9 to 12 cm or greater. Advanced tumors
exhibit raised papules or nodules that are blue, black or amelanotic and
often ulcerated.
Desrnoplastic Malignant Melanoma is a tumor consisting of spindle-shaped
cells with an increased dermal connective tissue component. This forrn of
rnalignant melanoma is aImost impossible to diagnose. Al1 four of the clas-
sic forms of malignant melanoma described above can show desmoplastic
changes.
In order to make an early diagnosis, the clinicians must be able to identify the
melanoma tumor, so the differences in characteristics between benign pigmented
lesions, rnalignant melanoma and precursor lesions that may give rise to malig-
nant melanoma are distinguished. Characteristic clinical features of early malignant
melanoma in general can be described by "ABCD", which stands for
A = Asymmetry
B = Border Irregularity
C = Colour Variegation which means that two or more colours exist within
the tumor border
D = Diameter generally greater than 6 mm
These can be seen on the book mark printed by the Canadian Cancer Society as
shown in Figure 1.4 to alert the patients about this kind of disease.
Moreover, its texture is smooth, too.
clinically are:
a In t radermal Nevus with characteristics:
- Colour Flesh-coloured, pink, tan or brown
- Shape Round or oval, may fade gradually into surrounding skin
- Surface Smooth, sometimes papiilomatous and raised
- Size < 6mm in diameter
Seborrheic Keratos is often known as age spots, age warts or liver spots. An
example is shown in Figure 1.3. It has these characteristics:
- Colour tan to brown, maybe flashy or pink
- Shape Borders often oval or round, but maybe irregular, often sharply
demarcated but will appear as gradually fading into surrounding
skin in fair persons.
- Surface Rough, raised surface and frequently sharp border
- Size Usually 5 - 15 mm
- Location Face, neck and trunk
Moreover, a kind of precursor lesion that rnay turn into a melanoma mole [14] is
0 Dysplast ic Nevus has characteristics:
- Colour Mixture of tan or brown, black and red/pink
- Shape Irregular borders that may include notches. May fade into
surrounding skin and include a flat portion level with the skin.
- Surface Smooth, slightly scaly or have a rough pebbly appearance
- Size < 5 mm
These descriptions indicate that melanoma and the above benign tumors differ
only slightly in their physical characteristics and colours, so a collection of these
features rather than a single feature is needed to obtain a satisfactory classification
of the tumor image.
The objective of the research is to find better and more efficient ways to automatically
detect early rnalignant melanorna using digital image processing techniques. This
thesis focuses on the preprocessing stage, the analysis of colour variegation, border
irregularity and asymmetry of the turnor. The ultirnate goal is to ease the doctor's
role in the detection of early malignant melanoma by providing better and more
reliable results, so that more patients can be correctly diagnosed.
This thesis is not intended only for engineers working in related field, but for
medical doctors, clinicians and other professionals interested in this field. This thesis
has been prepared in an easy-to-read and understandable manner, by providing the
basics behind the algorithms used, hoping the reader can get the most out of it from
reading this thesis.
1.3 Detection Process
The process of detection can be divided into three steps as shown in Figure 1.1:
Preprocess ing is to prepare suitable images for analysis by performing feature
enhancement and noise reduction. Tumor images may contain non-tumor
features like hairs, skin-mark and other noise that are acquired from taking
the photograph or digitizing process which will greatly affect the result of
the analysis; therefore, preprocessing is necessary.
0 Image Analysis is the stage to perform asymmetry, variegated colouring, bor-
der irregularity and textural analysis of the preprocessed tumor images.
0 Identif icat ion & Classification is to identify if the tumor image is malignant
melanoma or other sltin diseases. This stage is beyond the scope of this
research, because this is the task of the doctor whose expertise is in this
area. This research will provide better results to help the doctor to identify
the tumor more easily.
Preprocessing
1 Identification & Classification
Diagnosis
7;
Image Analysis
Y
Figure 1.1: Structure of the detection of MaIignant Melanoma
1.4 Images of Malignant Melanoma
Images used were digitized into 512x480 pixels, 24 bit/pixel colour resolution. Three
malignant melanoma tumor images representing the three types of most common
melanoma tumor mentioned before referred to as super, node, lent and one non-
malignant melanoma image Seborrheic Keratosis referred t o as sebor are used in this
thesis. They are shown in Figure 1.2 and Figure 1.3. They were obtained from a
wide variety of sources, thus there was no photography standard or control among
the images. So images photographed with varying magnifications or taken under
different environmental influences may exist; hence, the tumor size could not be
estimated from the images accurately and size analysis is purposely omitted in this
t hesis.
This thesis is basically divided iiito four sections:
1. Chapter 1 provides an introduction to the thesis topic and other relevant in-
formation about the research work.
2. Chapter 2 gives a brief overview of the existing techniques currently in use to
detect early malignant melanoma.
3. Chapters 3 and 4 discuss the simulation, design work done, results and ob-
servations obtained from preprocessing, colour variegation, border irregu-
larity and asymmetry analysis of the test images. Last of all, a common-
mole is investigated.
4. Chapter 5 concludes the whole thesis with a short discussion and by stating
the problems with the tested algorithms, the benefits of digital imaging in
dermatology and further research required.
Figure 1.2: (left) Superficial Spreading and (right) Nodular Malignant Melanoma
Figure 1.3: (left) Lentigo Malignant Melanoma (right) Seborrheic Keratosis n imor
LOOK FbR DANGER SlGNS PIGMENTE0 SPCITS ON THE SKlN
Asymm e t ~ y - one half unlike thc oiher half.
Border itregutar - scalkped or pooriy
C Color uafied fram one area to another; shades
of tan and brown: Mack; some tirnes white. mi or bluc. .
MlND THESE ABCD'S They may be signs of malignant melanama.
-ove;-
Consul t your doctor if you suspect any change which mrry indicatie skin cancer.
SKIN CANCER CAN BE . ,,
PREVENTED
Wear a hnt and pratcrtllrc clothing,
Usc a sunscrecn with an SPF of I 5 or higher that protects againrt UVA and UVB rays.
Avaid direct exposure bdnrewi 1 I A.M. - 4 P.M.
Figure 1.4: Canadian Cancer Society Alert Bookmark
Chapter 2
Overview of Techniques Currently
in Use
2.1 Introduction
There are many different techniques currently in use to process both gray-scale
and colour images to detect and identify malignant melanoma. Whether the technique
uses special equipment 119, 401, expert system [l], artificial intelligence or neural
network (3, 181, or otlier sophisticated algorithms 12, 4, 5, 6, 7, 8, 10, 20, 241, al1
have something in common: they attempt to diagnose patients as soon as possible.
Thus, it is important to investigate some of the more developed and commonly used
techniques from which many of the new techniques emerged. This chapter is divided
into sections according to the clinical features "ABC" of early rnalignant melanoma,
texture, preprocessing and identification stage. Methods used under each feature will
be briefly presented.
2.2 Preprocessing of Tumor Images
In order to study skin images, slides of skin tumor are digitized in true colour a t
512 x512 pixels, 24 bit/pixel colour resolution. In some of the processing like detection
of asymmetry and textural features, the resulting colour images are transformed to
luminance = 0.3 x red + 0.59 x green + O. 11 x blue (2.1)
Moreover, sometimes filtering of the image is required t o remove unwanted noise.
In detecting the boundaries of skin tumor, the image must be filtered through a me-
dian filter recursively to reduce noise and improve the visibility of the tumor borders
clearly by using a 3x3 window. Furthermore, in some cases, i t is necessary to remove
the background in images of skin lesions. This can be done by an adaptive histogram-
based region growing rnethod [l]. Pixel-by-pixel region-growing analysis is done on
the whole image and the histograms of the lesion and the background which are ob-
tained from some apriori information about the gray-scale image, are updated along
with the analysis. A binary mask is created after region-growing by showing which
pixels should be labelled as the lesion and which should be labelled as background.
The lesion can tlien be extracted from the background while preserving the texture,
and the boundary of the lesion can be traced accurately. Very often, to increase the
efficiency of the processing algorithm, image compression is applied on the image to
reduce colour information.
There may be much more preprocessing on the image to be done, but this depends
on the type, quality, ... etc of the tumor image obtained.
2.3 A: Detection of Asymmetry in Skin Tumors
To study asymmetry, al1 digitized colour images are transformed to black-and-white
images using the standard luminance formula. There are several algorithms to de-
termine asyrnmetry; for example, representing a closed curve or shape as a polygon
and then searching for the best symrnetric axis a t a hierarchical level by using an
eight-point neighbourhood method [54]. Or t o determine the rotational symmetry of
a closed curve S [55], by considering a circle C superimposed on S with center on the
centroid of S and radius equal to the average radius of S, so C and S will intersect a t
a set of points. The relationship between the order of the rotational symmetry of S
a L L U V L L C I L U L L L U G L U A 1 1 1 b G L J C L C i L W 1 1 ~ U L l l U O U b U V V b b L A W UiiU U buiii W b b U U C 4 U L L O L L G U i
The above algorithms are slow and complicated; in order to avoid the complexities
and slow performance of these algorithms, a new algorithm involving the calculation of
an asymmetry index is developed. This method is designed based on the characteristic
that any planar shape or closed curve possesses two principal axes and has a product of
inertia being zero. These principal axes are 90' apart and they intersect a t the centroid
of the area. For a nearly symmetrical shape, i t is assumed that one of the principal
centroidal axes is a sufficiently ciose approximation to the axis of symmetry because
for most benign tumors, there are one or two axes providing near symmetry. First of
all, Radial Search Algorithm of Golston el al [9] was employed to determine the
border of the tumor. This algorithm assumes that the center of the image lies within
a tumor and searches for jumps in luminance sustained for a specified length along
each of the sixty-four equally spaced radii from the center. When sixty-four candidate
boundary points are found, boundary points with large discrepensies in radii are
discarded and the border points are stored. A simple closed curve is constructed by
drawing lines between these stored points according to Sneighbour connectedness.
Then algorithms like Parity Check Filling or Connectivity Filling are used to
find the correct shape by filling enclosed pixel of the simple closed curve. Parity Check
Filling is based on the fact that a straight line intersects the contour of a region an
even number of times; that is, it assumes whenever a straight line goes into the
shape, it must also ieave. Therefore, it is possible to count boundary intersections
on a scanning Iine and decide if one is inside the shape or outside. Connectivity
Filling uses a point known to lie within the tumor (usually center of image) as a
seed point, and finds al1 points connected to it by 4-connectivity. This method may
require recursions for some arbitrary shapes and is often preferred, because i t does
not have tangential or acute angle problems which are present in parity check filling
method. The steps for calculating asymmetry index using the filled shape obtained
from above are as follows:
1. Translate or shift the final shape determined from previous steps so that
the x, y coordinates of image coincide with centroid of image.
3. Reflect image about abscissa and the ordinate.
4. Obtain the two area differences by subtracting the image on one side of
the axis from the refiected image.
5. Index of symmetry is obtained by dividing the least of the absolute values
of these area differences AAmi, by area.
AAmin x 100% Asymmetry index = - A
Al1 tumors with an asymmetry index of 6% or greater are classified as asymmetric.
From the experiment done using this method, it was found that 88% of the melanomas
in test have an asymmetry index above 8%, whereas this figure is only 47%, 62%
and 55% for intradermal nevi, seborrbeic keratosis and dysplastic nevi. Moreover,
correct determination of asymmetry was made in 93.5% of the tested tumors using
the automatic detection methods, which is quite a high success rate [6].
2.4 B: Detection of Skin Tumor Boundary Irregu-
larities in Colour Images
Boundaries represent a major fraction of the information content in an image. There
are however very many methods used to detect skin tumor borders. Conventional
luminance boundary detection algorithms are 1) Applying the Sobel O p e r a t o r by
using a 3 x 3 window size on turnors with luminance as a major border primitive
determinant and 2) Marr-Hi ldre th edge detec t ion m e t h o d which convolves input
image with the Laplacian of any size Gaussian function. I t was later found that
Radial Search is more efficient and reliable. It operates on gray-scale images and
searches outward radially frorn the center of the object, trying to find points that
are likely border points, by looking for jumps in luminance that are sustained for a
sufficient length. Shen the points are connected to form a border.
If the tumor to background transition is not sharp enough, wrong border points
will be found and often, rather than the tumor border, hairs or other skin mark-
ings are detected. So a method called the Low-pass filter-Radial edge detector
was developed hoping to overcome problerns faced in classical radial search rnethod.
First, the Fast Fourier Transform (FFT) of the image is computed. Then the Fourier
coefficients are multiplied by appropriate weighting factors, zero is used for al1 FFT
coefficients except those corresponding to the five lowest frequency components to
achieve low-pass filtering. Inverse FFT of the image is then computed. This low-pass
filter wiil remove hair and skin markings leaving the tumor border alone. After low-
p a s filtering, ordinary radial search is performed on the filtered image. This method
has proven to work quite well, but it requires that the tumor border be connected
radially; that is, each radial line must cross the turnor boundary only once. Violation
of this assumption will result in a false border being detected.
As technology advances, there is another method which detects a connected tumor
segment from border points, rather than detecting individual border points. It is
called Image Segmentation and it h a . been proven to be more efficient. Two
image segmentation algorithms are discussed: one is based on thresholding, the other
is based on the coXour information. Before applying image segmentation, in order to
improve the visibility of the turnor border and to ensure correct borders are identified,
the image should be filtered through a median filter recursively to reduce noise which
h a . corrupted the clarity of the image.
Thresholding After filtering, a transformation from RGB to X = w,R + w,G + wbB [Il] system where w,, wb, wg are weights, is done to identify
which pixels have large discriminant power in order to separate out tumor
pixels from the background ones; that is, to improve the bimodality of the
image histogram. Then a segmentation algorithm based on thresholding is
used to determine border of skin tumor. Thresholding is a widely used tool
in image segmentation for identifying different components of an image,
where and p ~ , , are the means of the known tumor and background
portions of the transformed image X.
The segrnented image may contain some background areas falsely îdentified
as tumor clusters due to noise along with the true tumor cluster. So
a recursive region growing algorithm that starts with a tumor point as a
seed point and recursively searches for tumor points in its neighbourhood is
used to identify the correct tunior segment. A contour-following algorithm
which assumes 4-connectedness is then used to determine border points in
their right order. The spline operation is used for smoothing the resulting
border, because even small irregularities will cause error in finding the
actual shape of the perimeter.
Colour Space Segmentation Colour Averaging is done on the orginal 512 x 512 x 3
byte image to reduce the amount of data to be processed by later pro-
cess and at the same time to reduce any noise present. Unwanted por-
tions (nonskin artifacts) of image are masked out by filling those por-
tions of the image with zeros or sometimes to mask out blocks of image
containing specific feature, so that each module can be processed inde-
pendently. Different ways of feature masking are applied depending on
different needs. The segmentation algorithm used is a combination of the
Median Split Colour Segmentation [56,57] with the Principal Com-
ponents Transform(P CT) 1271 t o optimally segment the colour image
to find the tumor border. The PCT is based on statistical properties of
the image, and is applied to the three-dimensional (3-D) colour space to
compress information into a reduced directionality. After PCT is applied
to the image, the median split colour space segmentation performs a colour
split along each of the three new axes from PCT, and a parallelpiped solid
is chosen for each of the colour regions. Averages are calculated for al1
pixels falling within a single parallelpipe within the PCT image space.
Then each pixel is mapped to the closest average colour, based on a Eu-
clidean distance measure. This process continues until the desired number
of colours is reached. Tlien small objects not to be identified as coIour
objects are filtered out, holes filled in or rounding out objects. The last
step is to determine the border using radial search technique. The criteria
for potential border points are I)A colour change exists a t the point and
2) The colour change has a sustained length of two pixels.
The last step to al1 border determination algorithms is to calculate the irregularity
index to determine the irregularity of the border.
The Irregularity indes is cafculated by the equation
where P is the perimeter of the tumor calculated as the number of pixels along the
border and A is the area of the tumor within the defined border obtained by counting
the number of points within and on the border.
The irregularity index for a circle is 1. From reference [7], it was found that a
threshold of 1.8 gives best separation between regular and irregular border. Most
melanomas have a high irregularity index, which means that they have an irregular
shape. Some experiments were done using this method and it was found that for
detecting border of melanoma images, a success rate is only 77%; while for intradermal
nevus, 100% is achieved. As a whole, image segmentation method has improved the
overall success rate to over 66% as compared to the radial search method.
T'mors
In order to detect variegated colouring in skin tumors, many algorithms have been
developed. Among these techniques, the most popular one is the automatic colour
segmentation by segmenting the image based on colour information. Before colour
segmentation can be applied, the digitized images are to be transformed from RGB
to spherical coordinates as follows:
where L is the vector length
BLUE
t 1=-> Angle A ?
u RED A Angle B
Figure 2.1: New Colour Space
This transformation splits the colour space into a two-dimensional (2-D) colour
space represented by LA and LB, and a one-dimensional (1-D) intensity (brightness)
space represented by the vector length L to avoid splitting colour objects that were
partially in shadow into separate objects. Figure 2.1 [5] shows the new colour space.
Now the image is ready for processing. Colour averaging is done on the orginal
5 1 2 x 5 1 2 ~ 3 byte image to reduce the amount of data to be processed by later process
C Y I I U L Y U YI. . , YU I I I IY "III." V V IV--"" CYILJ A A V I Y V y* v v v r - v - v II i. uirruvu Y"* U A V I A " \ AAwA.UA-AA*
artifacts) of image are masked out by filling those portions of the image with zeros or
sometimes this step is used to mask out blocks of image containing specific feature, so
that each module can be processed independently. Different ways of feature masking
are applied depending on different needs. The Colour Space Segmentation is outlined
as follows:
1. Obtain the minima and maxima of LA, LB and define a subspace.
2. Divide the subspace, into equal-sized blocks and number the blocks as
1,2, ... .
3. Make a note of the number of pixels that fell into a block.
4. Calculate the means and variances for R, G, B, L, LA, LB of each block.
5. Create a n image file with each colour vector replaced by the means of
(R,G,B) of the colour into which the corresponding pixel or group of pixels
fell in the colour space segmentation.
Object Filtering is done on the new image file to filter out small objects that will not
be identified as colour objects and filling in holes or rounding out objects to aid in
segmenting the image into colour objects. A six-connectivity mode1 containing the
upper left and lower right corners in addition to the four edge neighbours is used to
define neighbouring blocks. The filter goes through the image, block by block. If
a block has four or more neighbours that are the same, this block is replaced with
the value of the neighbours that are the same. A block is not replaced if i t has al1
different neighbours. A binary sequential labeling algorithm [58] is then used to label
the objects and finding the area of each colour object. Last of all, the identification
of variegated colouring based on the results obtained from previous steps can be done
via some decision criteria.
A problem always haunting this technique is the definition of the subspace created
by LA and LB as shown in Figure 2.2. As can be seen from the subspace, the closer
range; tha t is, for a region defined by a range of minima and maxima, on LA and LB,
the side of the region that is closest to the blue vertex is shorter than the side that is
closest t o the line that joins the red and green vertices. This distortion will facilitate
the perception based aspect of the image segmentation. If colour vectors near the
bIue vertex were in the image, the colour quantization scheme would probably need
to be modified.
BLUE O
RED GREEN
CIO* O
Angle B = 90
Figure 2.2: The Colour Triangle regions defined by 10 degree increments on Angle A and Angle B
2.6 Texture in Skin Image
Smooth texture is a characteristic of malignant melanoma and it is to be distinguished
from other textures including normal texture, regular hyperkeratosis, irregular hyper-
keratosis, warty hyperkeratosis(tal1 peaks) , papillomatous(numerous small bumps) , and lobular(severa1 larger burnps). The images are divided into 2 mm2 blocks con-
taining 32x32 pixels each for easier analysis. 3-D contour graphs are used to display
the luminance for each pixel as shown in Figure 2.3 and Figure 2.4
In the graphs, brightness is represented as peaks and darkness as valleys. Fig-
Figure 2.3: Smooth block (2mm2) from a skin turnor. Luminance for eacli pixel, range O -255, is displayed on the vertical axis, with pixel locations within the 2 - mm2 block displayed on the x and y axes
Figure 2.4: Rough block (2mm2) from a skin tumor
--- -.- L-'J ---r--J- - --------a --Vu-- - - O - - w L 7 - , --- - -., - - - - -O -- - - - - - - . , .' -----
Since the texture in a coloured image is assumed to be a function of black-and-
wl.iite(luminance) image, textural information is riot dependent upon colour. So
black-and-white images are obtained from colour images via the standard luminance
formula.
Three most popular statistical methods to determine the presence or absence of
smoothness will be discussed here. Al1 of them are fast, directly applicable to digital
images and rotationally invariant.
0 Circular Syrnmetric Autoregressive random field model (CSAR) This
is a model-based inethod for extracting two rotational invariant features
from a texture image [24]. AH features are obtained by fitting various types
of 2-D parametric random field models to the given texture. This model
operates within a 3 x 3 square matrix segment of the image subimaged on
the "circular neighbour set" where symmetrical points are located on a unit
radius circle centered at (0,0), as shown in Figure 2.5 [24]. The intensity
of a pixel is the summation of eight other pixel values in its neighbour-
hood. The intensity values of four of these neighbours are known since
their locations correspond to grid corners of the digitized image, but the
intensity values of the black dots in Figure 2.5 are to be calculated frorn
intensity values of nearby elements. The least square method is used to
estimate the two rotational invariant statistical parameters, namely &, ,6. Li is a statistical parameter used in linear estimation in the model, without
physical correlate; B is a measure of the degree of roughness of the texture.
gives the best experimental result. With this method, microtextures as
well as macrotextures can now be handled efficiently.
Neighbouring gray-level dependence m a t r i x (NGLDM) This method finds
a value for the (i, j )Lh position in the NGLDM matrix that is the count of
Figure 2.5: Relative spatiaI locations of the eight members of the circular neighbour set
the number of pixels in the image that have gray level i and have j neigh-
bouring pixels within a predefined radius or range of the gray level of the
index pixel [59]. The smoothness measure (Ni) is obtained by summing
over the matrix for each (i, j ) th entry the value of entry divided by j2. The
coarseness measure ( N z ) is obtained by summing over the matrix for each
(i, j ) th entry the value of the entry multiplied by j2. It was found that
gives the best result.
0 Number of peaks versus variance Two measures, "number of local peaks"
and "variance" are used for texture analysis. This method will do a scan
over each image block and determine the nurnber of local peaks, and the
variance of the image luminance. For a smooth block, the variance is small,
therefore there will be a smaller difference in the range of pixel intensities
encountered. It was found that when number of peaks = O and variance
5 11.9387, this method gives the best experimental results.
It was found that CSAR is the weakest of the three methods in detecting smooth-
I l G U J CülIU 1 T U U U I V I pblrvr iriu u r i v v v u u - i AU A u r uriu, u u u u r r r b vr v ~ * v uiyyrrvwvrrruj u r uvr-vur
algorithms is done to derrnatology, the NGLDM has been found to be suitable for
detection of shininess too 1241.
There is a major flaw in these methods, which is the lack of standard definition
of smoothness, because different dermatologists may have a different interpretation
of smoothness which will cause inconsistent results.
2.7 Identification of Malignant Melanoma
After appropriate image processing and low-level analysis to examine al1 features like
vertical thickness, coior variegation, pigmentation pattern and boundary characteris-
tics have been completed, high-level analysis of the above information is to be done
for the dermatologist to make a final diagnosis, so as to identify malignant melanoma
tumor arnong the other skin diseases. Thus a Knowledge-Based image analysis
and interpretation system, a high-level frame-based and rule-based expert system
is developed to interpret and analyze images of the skin lesion with respect to a set of
features, colour, boundary and surface characteristcis. This analysis when combined
with the patient's history, such as occurrence of melanoma or dysplastic nevi in the
family, is used by the knowledge-based expert system to detect early or potentially
malignant lesions. Figure 2.6 [l] shows the hierarchy of plans used in the rule-based
high-level analysis and interpretation system.
First, the given skin lesion is tested to find out whether or not the lesion is nor-
mal and benign. The charactistic features of the normal and benign lesion, such as
sharp outline, usually less than 5-mm surface, uniform pigmentation, fair uniformity
in colour of the lesion, etc. are analyzed. If it is not a normal and benign lesion,
i t is tested for being dysplastic nevus on the basis of the characteristic features of
dysplastic nevi. A plan maker will provide a top-down process with the clues on
what knowledge (such as boundary features) could be applied in each step of the
hierarchy. Each time a diagnostic variable feature is analyzed and if the condition
is met, a score will be assigned and will be added up in each stage of analysis. The
FHOM - Family History of Malignant Melanoma SIP - Suspicious - (Probability ?)
MCCSM- Minimum Characteristics Criterion Score o f Melanoma BP - Benign - (Probability ?)
MM - Mahgnanl Melanoma SIE - Suspicious - Rccxaminc in lime
decision is made on the basis ot the measurements of Ieatures, their analysis and the
risk factor like family history of malignant melanoma. At the end, the total score
is analyzed. A tliresholding number of this score is called a "Minimum Character-
istic Criterion Score for Melanoma" (MCCSM). The final decision is made on the
basis of MCCSM and other conditions. The final decisions as "BENIGN LESION",
" MALIGNANT LESION", " MALIGNANT PROBABLE LESION", "SUSPECT LE-
SION" and "SUSPECT AND REEXAMINE LESION" will be issued to the user with
the probability figures, and al1 the reasoning and feature measurernents utilized in
drawing the conclusion will be delivered to the user dermatologist or physician.
Apart from expert systems, there is another technique in the identification of
tumors which out rules the expert systems, this is the Neural Network. A neural
network can learn and gain experience on its own about the malignant melanoma
diagnosis problem. The ability to select pertinent features for a particular problem
on their own is an advantage which neural networks possess over expert systems when
solving such diagnostic problems.
Neural networks are used as pattern classifiers to classify skin tumors as malig-
nant or nonmalignant from colour photographic slides of the tumors. A multilayered
feedforward neural network trained using the generalized delta rule(backpropagation
training algorithm [GO, 611) is suitable, because it is built to classify digital images
of tumors into a small number of fixed categories. Nodes in this backpropagation
neural network form a weighted sum of the inputs and is mapped t o the output of a
neuron via the hyperbolic tangent function. A gradient-descent training technique,
called backpropagation, which minimizes the squared error between actual outputs
of the network and the desired outputs is used to train the network. After learning,
the network cari generalize, giving correct responses even in the presence of patterns
tha t are not included in the training set. The overall diagnostic test results were very
promising, with an accuracy as high as 86% in detecting malignant melanoma [3].
Chapter 3
Preprocessing and Colour
Variegat ion Analysis
3.1 Introduction
ince Early processing techniques were only concerned with monochrome images. '3'
colour conveys variable information about the objects in a scene and this information
can be used to further refine the performance of imaging system, the processing of
colour images is becoming more and more important. Coloured tumor images are
used, so a brief discussion about colours and colour spaces is presented here which
will help in the understanding of colour images.
For the colour information about the object in an image to be useful, it is necessary
to identify individual regions or objects by a certain criteria; image segmentation
is one good technique and will be discussed in brief. Following the discussion of the
basics of image processing are the preprocessing and colour variegation analysis.
3.2 Colours
If light is achromatic, intensity is its only attribute. Gray-level is a measure of in-
tensity that ranges from black to gray and finally white. For chromatic light, the
w W" * a *
quality of light can be described by three basic quantities:
Rad iance is the total amount of energy that flows from the light source.
0 Luminance is the rneasiire of the amount of liglit an observer perceives from
a Iight source.
a Brightness embodies the achrornatic notion of intensity.
The three primary colours defined by CIE (Commission International de 1'Eclairage
- the international commission on illumination) are red(R) = 700nm, green(G) =
546.1nin and blue(I3) = 435.8nm. These çolours can produce al1 kinds of visible
colours wlien mixed in various intensity proportions. A colour is specified by its
trichromatic coefficients:
where r + b + g = 1
The charact eristics used to distinguish one colour from another are brightness,
hue, which is an attribute associated with the dominant wavelength in a mixture of
light waves, and sa tura t ion, the relative purity or the amount of light mixed with a
hue. Hue and saturation can be taken together as chromaticity.
3.3 Colour Spaces
A colour space is a geometrical and mathematical representation of colour. The most
commonly used colour spaces in practice for image processing are the RGB and the
HSI (hue, saturation, intensity) model. Many other coIour coordinate systems like
- -. - - - - 1 - - - - - - - - - -
c - - d - -- a - generally linear or nonlinear versions of the RGB and HSI colour models [IO].
3.3.1 RGB Colour Model
Figure 3.1: RGB Colour Model
This model ensures that there is no distortion of the initial information and it can
be viewed as in Figure 3.1 [27]. In this model, the gray scale extends from black to
white along the line joining the black and white points, while colours are points on
or inside the cube defined by vectors extending from the origin.
3.3.2 HSI Colour Model
HSI model is defined with respect to the colour triangle as shown in Figure 3.2 1271. A
3-D vertical ellipse structure shown in Figure 3.3 [27] can be constructed by combin-
ing hue, saturation and intensity into a 3-D colour space. The hue of a colour varies
along the circumference, saturation varies along the radial distance and its intensity
is determined by its perpendicular distance from the black point. HSI is often pre-
G
ferred over RGB, because
Figure 3.2: HSI Colour Triangle
White
Line of Grays
Pure Colours
Saturation
Black
Figure 3.3: HSI Colour Space
the 1 (intensity) component is decoupled from chrominance
information (H,S) and H,S are related to the way human beings perceive colour.
I t is possible to convert from RGB model to HSI model by this set of equations:
H = cos-l( $[ (R - G) + (R- B)]
[(R - G)? + ( R - B)(G - B)]; )
A disadvantage in using HSI is that peaks of hue histogram are often split due to
the angular character of the hue CO-ordinate system [37]. So care must be taken when
using HSI space, and this has created difficulty in many of the hue-based process.
These are uniform colour spaces which account for the nonlinear response to lurni-
nance. One advantage of using this space is that the perceptual colour distance can
be forrnulated easily. Here L* represents the perceptual response to luminance, u*v*
and a* b* represent chrominance.
The following are the expressions to obtain L*u*v* and L*a*l* [38]:
Colour differences can be easily obtained via L2 norrn as follows:
These spaces are very useful in the precise evaluation of perceptual closeness
between two colours like in colour matching systems [38].
3.4 Image Segmentation
Image Segmentation is a popu1a.r technique used in image processing t o identify indi-
vidual regions or objects in an image. It is used in rnany areas like tumor extraction,
bviu u r i iwr~vb-vrvrr, v r U A A AAA A U ~ A A U A L J L A A S O ~ ~ - A I L ~ L U U U K A A U U VIA ~ U ~ I I A V A U LAILU U L ~ U L , b~ u u u ,
telangiectasia and reflections. Segmentation is basically a process of segmenting a
picture into subsets by assigning the individual pixels to classes, known as pixel
classification. As a result, the regions should be homogeneous with respect to the
segmentation criterion. For exarnple, dark objects can be distinguished from their
light background or vice versa by segmenting a picture via thresholding its gray level,
which is classifying the pixels into dark and light classes.
The basic segmentation techniques in colour image processing are:
* Gray Level Thresliolding
Region-Based Segmentation
3.4.1 Gray Level Thresholding
Figure 3.4: Histogram with distinct peaks
The gray level histogram of pictures that is composed of only two kinds of regions
at different gray level ranges will display peaks corresponding to the two gray level
ranges as shown in Figure 3.4 [27]. These kinds of pictures can be segmented by
choosing a threshold that separates these peaks. Selection of an optimum threshold
is not an easy task; the end result of the segmentation depends very much on the
t hreshold selected.
For pictures tha t contain more than two types of regions with histograms as shown
in Figure 3.5 [27], multilevel thresholding is possible by applying several thresholds.
However, when regions overIap in a picture, segmentation by thresholding becomes
difficult. l t 1s impossible to cieaniy separaLe me overiapyirig regluus; urtxtxuic, D U ~ L K
preprocessiiig like smootliing or averaging is to be done on the image before thresh-
olding. Sometimes, when a single threshold does not give good segmentation results,
especially for uneven illumination in a picture, dividing the picture into blocks and
applying threshold selection techniques to each block is possible.
Figure 3.5: Histograrn with more than 2 peaks
3.4.2 Region-Based Segmentation
There are basically two types:
Region Growing This algorithm will start with a set of "seed" points and
from thcse, regions will grow by appending to each seed point those neigh-
boring pixels that have similar properties.
Region Splitting The image is considered as one region to start with. It will
look for groups of pixels of similar properties and partition the image into
a set of small region. Then a uniformity test is applied to each region: if
the test fails, tlie region is subdivided into srnaller elements. The unifor-
mity test is applied again; this is repeated until al1 regions are uniforrn,
thus splitting the image into smaller regions. A specific region splitting
segmentation intended for colour images that is being used in this research
is Vector Directional Image Segmentation [39]. It employs the prop-
erties of a vector field where the vector magnitude is proportional to the
intensity of the image, and the direction represents colour chromaticity.
The vector direction is represented by
where m, = Jr2 + g2 + 6 2 is the vector magnitude, O 5 ai 5 90°,i =
fi, G,
Two angles out of c r ~ , c x ~ , CYG plus mv, the vector magnitude are used to
represent the colour attributes. Firstly, al1 the peaks of (ai, aj i # j )
histogram are obtained and then a group of vectors whose directions are
around each of the peak obtained are segmented out as a possible region
according to a certain tolerance level. Since each vector is characterized
also by its magnitude, the vectors with the same direction may form more
than one region depending on their magnitude. So the region is further
segrnented if the magnitude is different. The above is repeated until no
histogram has any significant peaks.
3.5 Preprocessing
Before any detection or analysis algorithm can be done on the images, preprocessing
rnust be applied to prepare suitable images for further processing. Since the tumor
images may contain hairs, skin-marks, skin background, and other noise acquired from
photography taking or digitizing process, two types of preprocessing are proposed for
the tumor images obtained:
+ Filtering
m Tumor Extraction
Noise in tumor images are very different from the standard type of noise which are
used to corrupt standard images. The noise in a tumor image is often unknown,
positive or negative spikes, mixed and more than one pixel wide; for example, thick
hairs, wrinkles, skin-marks. Baically, the noise is not as simple as Gaussian or
impulsive noise which many researchers have tested with or developed filters to get
rid of them. Apart from the capability of removing noise, the filter should also
preserve tumor border. So a number of filters were tested and compared.
3.5.2 Median Filters
It is known that median filter removes impulsive noise both positive and negative ef-
fectively while preserving edges. Moreover, it does not reduce the brightness difference
across steps, because the values available are only those present in the neighbourhood
region, not an average between those values. Hence, two types of median filters are
tested.
Multivariate Ordering There are two types:
0 M-Ordering Median Filters (marginal ordering) The multivariate sam-
ples are ordered along each one of the colour component independently,
so the correlation between signal components is not utilized and there-
fore rnay not preserve well the edges.
R-or dering Median Filters (reduced or dering) Each multivariate ob-
servation within a window is reduced to scalar values di according to
a certain distance criterion. The samples are arranged in ascending
order of di and the median one is chosen.
Vector Median Filters The aggregate distance of X h o the set of vectors X1, X2, ..., Xn
js defined as
The ordering becomes
d(1) l 4 2 ) 5 . . -1 d(n)
wliich implies the ordering
and the vector median XVM =
Other filtering techniques tested were basic and generalized vector directional
filter.
3.5.3 Vector Directional Filter
Vector Dircctional Filters (VDF) is a class of multichannel image processing filters
that are based on vector ordering principles. Angle between the image vectors is
the ordering criterion. Basically, the processing of vector data in VDF is separated
into directional processing and magnitude processing. This separation of process
will establish a link between multichannel signal processing and single channel image
processing ivhich is a major reason why VDF is chosen to process colour images.
There are two types of VDF: Basic form only considers the vector direction; while
the General ized form deals with both vector direction and magnitude processing.
Basic Vector Direct ional Filter (BVDF) BVDF employs a window W which
slides on the image plane. For each window position, the vector a t the central
pixel is replaced by the vector that minirnizes the sum of the angles with al1 the
other vectors within W. Hence, the vector most centrally located is chosen as
output of BVDF.
For fi a vector in W,
where A(fi fj) denotes angle between the vectors fi and fj, O 5 A(fi fj) 5 a for
each pixel in the wiiidow of size n
An ordering of the ai's:
< a.. Q(i) 5 q 2 ) - I Q(n)
implies the saine ordering:
The output BVDF = f(')
Generalized Vector Directional Fi l ter (GVDF) The generalized vector direc-
tional filter (GVDF) is the general form of the Vector Directional Filter. In
directional processing, a set of vectors that is centrally located in population
with approximately the sarne direction in the vector space is the output. The
first k terms of expression (3.31) are chosen as the set of output vectors. This
is often considered as a single-channel signal. Then magnitude processing is
applied in cascade; a single output is produced a t each pixel by passing the
signal from the previous process through a magnitude filter. In this research,
two operators have been used for "magnitude" processing:
0 Average
Max-Min
The results obtained with the various types of filtering were still not as promissing,
so a recursive modified version of each kind of filters is proposed and tested. This
version of filter is derived from the Modified Median Filter.
3.5.4 Modified Median Filter
This filter is implemented by an ordinary median filter followed by thresholding as
shown in Figure 3.6. If the difference between the filtered output and the original
Figure 3.6: Modified Median Filter
input is larger tlian the threshold, the final output is the filtered output; if the differ-
ence is smaller than the threshold, the input will not be filtered. An enhancernent to
the modified median filter can be done by replacing the Media11 Filter block by Vec-
tor Median, M-ordering or R-ordering Median Filter, or even non-median filters like
Vector Directional Filters. Apart from replacing the block, the whole filtering process
can be made to run recursively until the root signal is found. The final form of the
proposed filter looks like Figure 3.7 which is a Recursive Modified Generalized
Filter, not necessarily a modified median filter anyrnore.
Figure 3.7: Recusive Modified Generalized Filter
The most popular method used in industry now is the ordinary recursive median
filter which wilI erase the fine details like hairs or wrinkles, and large regions will talte
on the same brightness values; while the edges remain in place and well defined, there
will still be considerable blurring of the edge. The method proposed in this thesis
would be better because the amount of filtering done on the image is controlled by
the threshold. So unnecessary filtering is avoided, and will better preserve the edge
of the tumor.
3.5.5 Results
The four test images have very different noise content. The noise basically cornes
from various sources, and depends on the part of the body from which the image was
taken. For example, the node image was talen from the face of the lady, so there
were no observable hairs on the image, but the cheek lines and mouth have created
a problem in the analysis. Therefore, different filters are needed for different images.
The filters that give the best result for the particular image are listed in Table 3.1.
The filtered images are shown in Figure 3.8 and Figure 3.9.
Very obvious differences can be seen in the super image. The thick, dark hairs on
the image are filtered out; while the image border is not corrupted and the sharpness
of the border is preserved. For the node image, the deep cheek line on the face is
made lighter and thinner by filtering. This will help in tumor extraction.
Table 3.1: Filters Used
3.5.6 Turnor Extraction
Image Superficial Spreading Nodular Lentigo Seborrheic Keratosis
The simpliest way of Gray-Scale thresholding was tested first because of its fast and
easy computation. The histograms for each tumor image are shown in Figure 3.10
and Figure 3.11. The thresholds for each image are listed in Table 3.2.
Results are shown in Figure 3.12 and Figure 3.13. I t can be seen that the back-
ground and the tumor are not separated very well, especially for the node image.
As well, dark lines on the background can be seen. This may be because only pixel
Filt ers Recursive Modified Max-Min Vector Directional Filter
Basic Modified Vector Directional Filter Modified Vector Median Filter
Recursive Modified Max-Min Vector Directional Filter
Table 3.2: Gray Scale Threshold values
Image Threshold Value Superficial Spreading Nodular Lentigo Seborrheic Keratosis
brightness is considered; the histogram only counts pixeI in the entire image, losing al1
information about the original location of the pixels, the brightness values of al1 their
neighbours and the colour information of colour images. Even if a series of histograms
for each of the RGB colour planes is used, i t is difficult or often impossible to judge
which of the peaks correspond to the tumor of interest. So the use of three separate
histograms and sets of threshold levels do not help to see which pixels have various
combinations of values. This rnethod wilI only work well for monochrome images.
Although two or more-dimensional thresholds can be used, i t would be difficult to
interpret the meaning of the settings and it is often complex to use, too.
Since colour conveys a lot of information about the image, colour information can
be used to extract the tumor from the background. Vector Directional Segmenta-
tion [39] is a kind of Region Splitting Segmentation algorithm. It would be a suitable
technique, because this technique has put into consideration both the vector direc-
tion and magnitude. This is much more difFerent and accurate than simply histogram
thresholding. The angle histograms are shown in Figure 3.14 and Figure 3.15. The
corresponding thresholds used to extract the tumors are listed in
Table 3.3: Angle and Magnitude Thresholds
1 Imaoe 1 Anqle 1 Maqnifude 1 " 1 Y , - Superficial Spreading 1 28,66 1 370
Table 3.3.
Nodular Lentigo Seborrheic Keratosis
35,73 42,67 43,66
220 270 290
With proper postprocessing (usually subtraction), the filtered tumor can be iso-
lated from the background. Then using the filtered tumor as a mask, the tumor
from the original image can be masked out. Thus an unfiltered tumor on a white
background is obtained. This tumor may contain some noise, too, but much more
less than the background, minor median filtering is done to eliminate the noise. The
extracted tumor is shown in Figure 3.16 and Figure 3.17
The extracted tumor obtained via vector directional segmentation was much better
than the ones obtained from Gray-Scale Thresholding. The most obvious improve-
ment can be seen from the node image. Moreover, the super image extracted here is
more complete, bccause holes can be seen al1 over the gray-scale thresholded image.
It can be concluded that colour information is needed in tumor extraction.
3.6 Colour Variegation Analysis
Colour Variegation means more than one colour existing within a tumor border and
is one of the clinical features of early malignsnt melanoma. I t is very difficult to
distinguisli between two colours that Vary only slightly by using bare human eyes.
Visual limitation will cause an image to be under-segmentated. Very often, computer
algorithms will deal much better than humans with scenes containing more than one
type of colour. Image Segmentation is typically helpful in this area.
Different methods were tested:
3.6.1 Gray-Scale Thresholding
The coloured tumor images obtained from the 1 s t section are converted to gray-scale
images and the histograms are plotted as shown in Figure 3.18 and Figure 3.19. It
can be seen that pealts of very similar intensity exist on the graphs. By choosing
values between the peaks as thresholds, the different intensity of the images can be
extracted. The result is shown in Figure 3.20, Figure 3.21, Figure 3.22, Figure 3.23,
Figure 3.24 and Figure 3.25. For both the super and node image, the image can only
be divided into two regions, while the lent image is divided into five regions with
difierent intenslty. ror the sebor image, since tne nistogram snows oniy one tail pearr,
it is assumed that oniy one intensity exists within the tumor border.
Looking at their corresponding colour images, it can be seen that this method
of thresholding can only distinguish between regions of different intensity, but many
tirnes, more than one colour does exists within a region of sarne intensity. So colour
information has to be introduced as a segmentation criteria, rather than just the
intensity alone.
3.6.2 Vector Directional Segmentation in RGB Space
Thc correlation betwecn the colour components is taken into consideration. Possible
regions are listed in Table 3.4.
Table 3.4: Possible regions of segmentation
Image Super
With appropriate tolerance level, the image is segrnented into different regions
as shown in Figure 3.26, Figure 3.27, Figure 3.28, Figure 3.29, Figure 3.30 and Fig-
ure 3.31, Figure 3.32 and Figure 3.33. I t can be seen that there is more than one
colour in sorne of the region; it is obvious that the method under-segmented the image
which is often not preferred. There is an improvement here when compared to gray-
scale thresholding, because more regions can be segrnented out using this method.
The most obvious difference can be seen from sebor image.
Possible Regions (a, p, mv) 55,55,20 1 55,55,70 1 42,76,135 )
Node Lent Sebor
3.6.3 Segmentation in HSI Space
Because segmentation in RGB space has caused an under-segmentation problem,
other colour spaces have been tried. HSI is cllosen in this section because of its
35,73,30 42,67,30 43,66,lO
35,73,50 42,67,70 43,66,100
35,73,100 55,55,50 55,55,240
advantage as aiscussea eariier in tnis cnapcer. fi rnu1r;isr;age reglon sprir; anu Irrerge
segmentation known as Valley Searching and Merging algorithm [53] was tried.
Valley-seeking routine is applied to the hue component to locate al1 valleys of
the hue histogram. These valleys are then used as threshold to segment the image.
The whole thing is repeated twice for al1 the segmented region but applied to the
saturation and intensity component simultaneously. A11 three-components are used
because different things will affect the cornponents differently, like uneven illumination
has little effect on the consistency of hue, but has a significant effect on the consistency
of saturation.
RGB tumor images were converted to HSI space according to the expressions
given earlier. When the hue histogram \vas plotted, a problem which existed was
that discontinuities in the hue component introduced a visible split of the histogram.
HSI systems give rise t o singularities which result in undesirable instabilities, notably
with respect to the statistical properties of hue distributions [37]. Although ways have
been soughted t o overcome such a problem, significant amount of numerical overhead
will be added. So some other ways of segmentation were looked at.
3.6.4 Segmentation in L*u*v* or L*a*b* Space
L*u*v* and L*a*b* are uniform colour spaces and their associated colour-difference
formulae can help to discriminate two colours even if they differ only slightly. This has
provided great reason for choosing segmentation in this space. The proposed method
will make use of the magnitude of each pixel and the colour difference between them
by region growing from a seed point. The image is converted to L*u*v* or L*a*b*
space via expressions given earlier in this chapter.
The seed point is chosen to be the pixel with smallest magnitude value. Once
this is found, the colour difference between this seed and other pixels are found.
For al1 the difference values which are smaller than a predefined threshold (obtained
from experirnentations), the pixels are grouped as one region. Then the rest of the
ungrouped pixels will go through the whole procedure again, defining a new seed
point and finding the colour difference during every round. The recursion will stop
when the last region to be segmented has less than a certain amount of pixels. Then
a merging technique is used by comparing colour difference between two adjacent
regions. If they differ oniy slightly, they can be merged to become one region. This
way, the image will not be over-segmented. The result of applying this algorithm in
the L*a*b* space to the turnor images is shown in Figure 3.34 and Figure 3.35.
Table 3.5 will give the number of different colour regions found on the tumors.
Table 3.5: Number of Segmented regions
I Irnaqe I Number o f re.qions I - I - - 1 Superficial Spreading 1 9
From the segmented images, different regions on the tumor are shown by
Nodular Lentigo Seborrlieic I<eratosis
a differ
9 10 3
ent
colour. Very often, it is liard to distinguish between two regions with our human eyes,
because the colour difference between two regions is too small; but the computer
algorithm here tells us tliat the two regions are different. For sebor image, only this
fine separation can distiguish three regions on the tumor, whereas in the previous
rnethods, only one or two regions were found. So this method is a good choice for
colour image segmentation.
3.7 Observations
3.7.1 Preprocessing
At the very beginning of al1 the analyses, filtering is done on the images to reduce
noise and other artifacts on the image. The most popular method used in this area of
melanoma recognition is the median filter with a 3x3 window. I t is well known that
this filter can erase fine details but it will blur the edge of the image a considerable
amount. So in order to find a filter that out performs this older filter, rnany different
kinds of filters were tried, but a simple common one that can filter al1 the images
equaily weil ~ u u l u ~ i u c ut: luuiru. A ~ I G L G L U L G , A U U - a A "lluu u ~ A A w A u . ~ w .,+
noise will require different types of filters. The best filters for each of the test images
are listed in Table 3.1 and the filtered images were shown in Figure 3.8 and Figure 3.9.
Al1 these filters perform much better than the ordinary median filter by which they can
remove noise effectively while preserving the edge and sharpness of the turnor quite
well. I t can be seen that the Recursuve Generalized Max-Min Vector Directional
Filter is best in removing hairs; this can be observed from image super and sebor.
For tumor extraction, the industry is only working with gray-scale images, because
techniques are more developed in this area. In this thesis, i t was found that colour
segmentation is better t han gray-scale t hresholding, because i t is often difficult to
find a thresliold that can exactly subdivide the image into two different regions; it
is not enougli to rely only on brightness information of the individual pixel. The
extracted tumors found from both methods can be found in Figure 3.12, Figure 3.13,
Figure 3.16 and Figure 3.17. For the gray-scale thresholded ones, i t can be seen
tha t some background is included as the tumor or some tumor part is included as
background as long as tliey have the same intensity. Therefore, one cannot depend
solely on pixel brightness, thiis losing al1 information about the original location of
the pixels, brightness values of al1 their neighbours and colour information of the
pixels. For the extracted tumor ob tained from vector directional segmentation, the
exact tumor with no other undesirable portions is obtained.
3.7.2 Colour Variegation Analysis
For the detection of variegated colouring in skin tumors, the most popular technique
used nowadays work with spherical coordinates, splitting the colour space into a 2-
D space and a 1-D space to ease in the analysis. There is a major problem witli
this method and that is the distortion caused by the transform which will facilitate
the perception based aspect of the image segmentation. The proposed and tested
as "best" method in this thesis is much more easier to use and does not introduce
any distortion a t all. I t is the segmentation in either L*u*v* or L*a*b* space that
can segment the tumor into different regions according to their colour content with
V A A b l l b A p V A V I A " V V I V U L U I A A b A L A A b - ~ " A I A A U I U U + &&.&V ru-ri----- r u v r r u vvuu ru---- i i i -
this uniform colour space is chosen. This method also out performs segmentation in
RGB space because RGB segmentation depends on too many experimental thresholds
which will create a big error margin. Moreover, gray-scale thresholding is also tested,
although it is the easiest method to use, but it under-segments the images, which is
often not preferred.
Al1 the segmented images are shown in Figure 3.20, Figure 3.21, Figure 3.22, Fig-
ure 3.23, Figure 3.24, Figure 3.25, Figure 3.26, Figure 3.27, Figure 3.28, Figure 3.29,
Figure 3.30, Figure 3.31, Figure 3.32, Figure 3.33, Figure 3.34 and Figure 3.35. They
can be compared and it is noted that L*u*v* or L*a*b* segmentation is the best.
Figure 3.8: Filtered Colour Images: (left) Super (right) Node
Figure 3.9: Filtered Colour Images: (left) Lent (right) Sebor
47
Figure 3.10: Tumor Extraction: Gray Scale Thresholding Histogram (left) Super (right) Node
SOM
zsw
Figure 3.11: Tumor Extraction: Gray Scale Thresholding Histogram (left) Lent (right) Sebor
Figure 3.12: Tumor Extraction: Gray Scale Thresholding Extracted Tumor (left ) Super (right) Node
Figure 3.13: Tumor Extraction: Gray Lent (right) Sebor
Scale Thresholding Extracted Tumor (left)
Figure 3.14: Tumor Extraction: Vector Directional Segmentation Colour Histogram (left) Super (right) Node
Figure 3.15: Tumor Extraction: Vector Directional Segmentation Colour Histogram (left) Lent (right) Sebor
Figure 3.16: Turnor Extraction: ~ & o r Directional Segmentation Extracted Colour Tumor (left) Super (right) Node
Figure 3.17: Tumor Extraction: Vector Directional Segmentation Extracted Colour Turnor (left) Lent (right) Sebor
Figure 3.18: Colour Variegation Analysis: Gray-Scale Thresholding Histogram (left) Super (right) Node
Figure 3.19: Colour Variegation Analysis: Gray-Scale Thresholding Histogram (left) Lent (right) Sebor
t, .. r .. .-. - ,. .. .
Figure 3.20: Colour Variegation Analysis: Gray-Scale Thresholded regions for Super
Figure 3.21: Colour Variegation Analysis: Gray-Scale Thresholded regions for Node
Figure 3.22: Colour Variegation Analysis: Gray-Scale Thresholded regions for Lent
Figure 3.23: Colour Variegatiori Analysis: Gray-Scale Thresholded regions for Lent
Figure 3.24: Colour Variegation Analysis: Gray-Scale Thresholded regions for Lent
Figure 3.25: Colour Variegation Analysis: Gray-Scale Thresholded regions for Lent
Figure 3.26: Colour Variegation Analysis: Super Colour regions
Figure 3.27: Colour Variegation Analysis: Super Colour regions
Figure 3.28: Colour Variegation Analysis: Node ~ o l o u r regions
Figure 3.29: Colour Variegation Analysis: Node Colour regions
Figure 3.30: Colour Variegation Analysis: Lent Colour regions
Figure 3.31: Colour Variegation Analysis: Lent Colour regions
Figure 3.32: Colour Variegation Analysis: Sebor Colour region
Figure 3.33: Colour Variegation Analysis: Sebor Colour region
Figure 3.34: Colour Variegation Analysis: Lab Segmentation Node
of (left) Super (right)
Figure 3.35: Colour Variegation Analysis: Lab Segmentation of (left) Lent (right) Sebor
Chapter 4
Border Irregularity Analysis,
Asymmetry Analysis and Analysis
of a Common Mole
4.1 Introduction
Border Irregularity and Asymmetry are two critical features in the identification of
malignant melanoma. Tliese can be regarded as shape analysis of a simple closed
curve. The shape of an object is entirely described by its boundary, so it is important
to obtain a clean border for use in the analysis stage. The best classical method used
to detect the edge of an object is the Radial Search Method [4] or other rnodified
version of it [4], but there are problems when the turnor image is complex such
as intersecting with itself. So in this thesis, some edge detection techniques will
be examined to obtain necessary information for border irregularity and asymmetry
analyses. The most widely used irregularity and asymmetry indices will be computed
and compared to the shape factor measures proposed in this thesis. Last of all, a
common-mole will go through al1 the analyses stages and the results will be examined.
4.2.1 Gray-Scale Images
In gray-scale images, an edge is the boundary between two regions with relatively
distinct gray-level properties. Edge Detection will detect edges or curves, basically
local features that involve abrupt changes in gray level in the image. For example,
when an edge exists, the gray level changes abruptly as the border between regions is
crossed; for lines or curves, the gray level is constant except along a thin strip, which
yields a sharp spike; for spots, the gray level is relatively constant except at one loca-
tion. Gradient operators like 3x3 Sobel operator as shown in Figure 4.1 [27], Prewitt,
Roberts, ... etc., or other Laplacian operators are used, followed by a threshold oper-
ation on the gradient in order to decide whether an edge or other local features exist.
The pixels that have been identified as edges must be linked to form closed curves
surrounding the region.
Figure 4.1: Sobel Operators
Sobel operator is often more attractive than the other operators because it has
the advantage of providing both a differencing and smoothing effect.
4.2.2 Colour Images
For colour images, a nurnber of different algorithms exist for colour edge detection.
Some of them treat the three colour components separately, thus the correlation
between the three colours is lost. This rnethod is referred to as Scalar Edge Detection;
while the other algorithms treat the colour cornponents as vectors which is the Vector
E d g e Detection.
>orne examples 01 cne vecr;or bage uer;ecr;iwi are;
0 Vector Gradient Operator which is an extension of gradient operators from
gray-scale images to colour images.
0 Vector Range Operator (VR) which is based on vector order statistics de-
fined as VR = I I x ( ~ ) - x ( ' ) I I where N is the number of image pixels in the
window, X" i = 1,2, ..., N are the sample vectors and ~ ( ' 1 , i = 1,2, ..., N
are the ordered vectors as shown:
VR gives good response to edges but is sensitive to noise.
a Difference Vector Operators [39] which is extremely effective because they
yield high gradients a t the location where the image attributes are changing
rapidly while not sensitive to noise.
Figure 4.2: Window
Consider the window in Figure 4.2 [27]:
The following changes of Xo is obtained from the use of difference vector
operators in the horizontal and vertical directions.
These expressions represent a gradient value of Xo in four directions. The
maximum value among the four directions is taken as the gradient mag-
nitude and the corresponding direction is taken as the direction of the
gradient. A threshold is applied to the vector gradient magnitude to ex-
tract edges from an image background. The vectors with larger gradients
are regarded as the colour edges; the other vectors are considered to be
image background pixels.
4.2.3 Results
Gray-scale edge detection is chosen because the tumor has been extracted from the
background in the preprocessing stage and colour information is not needed in finding
the edge from a n extracted turnor. This will save computation time and the technol-
ogy is much more well developed in this area. The colour turnor is first converted to a
gray-scale image, then Sobel operator is applied to the image. The results are shown
in Figure 4.10 and Figure 4.11. It can be seen that too much details exist and it's
difficult to define the edge of the tumor, so some preprocessing is needed to be done
on the gray-scale images. The details of the edge within the image appear because
there is more than one kind of intensity level on the tumor, so filtering is done to
remove outliers; smoothing is done to smooth the intensity of the image; erosion is
done to get rid of lioles on the image that may cause a fault edge to be detected.
The processed tumor edge is shown in Figure 4.12 and Figure 4.13. It is found that
the edge is too thick to be useful; the existence of parallel edge pixels is one of the
undesired artifacts of edge detection algorithms. A thinning algorithm used t o re-
move the inherent edge broadening in the gradient image without destroying the edge
where A is the set for thinning and B is the structuring element
Defining I3 as
{ B ) = {BI, B ~ , . -. , Bn)
where Bi is a rotated version of Bi-'
Therefore thinning of the set A by a sequence of structuring element B is
A €3 { B ) = ((. * - ( ( A €3 B') 8 B ~ ) - - .) 63 B") (4.9)
In other words, the process is to thin A by one pass with B1, then the result is
thinned with one pass of B2 and so on, until A is thinned with one pass of Bn. The
entire process is repeated until no further changes occur. The
sequence of structuring elements is shown in Figure 4.3.
most comrnonly used
Figure 4.3: Structuring element of the thinning process
The thinned result is sliown in Figure 4.14 and Figure 4.15. I t can be seen that
65
the edge is still not clean enough. To obtain a one-pixel wide edge, another sets of
structuring elements are used. They are shown in Figure 4.4 and Figure 4.5. The
final edge is one pixel wide as shown in Figure 4.16 and Figure 4.17.
Figure 4.4: Structuring element of the 2nd thinning process
Figure 4.5: Structuring element of the 3rd thinning process
4.3 Border Irregularity
According to the Canadian Cancer Society, border irregularity means scalloped or
poorly circumscribed border. So paying close attention to the contour complexity
and side regularity are necessary. Actually, these can be considered as principal
factors in the description of a good shape [41]. Border Irregularity analysis can be
considered as "goodness of shape" analysis. To describe a particular object shape,
the border of the object must be represented in some useful way, so that important
information can be extracted from the border, too. Polygonal approximation, chain-
code representation, signature and incremental curvature will be examined. Then, the
angle regularity, circularity or roundness (the most widely used index), side regularity,
minimum energy measures and the geometric mean of al1 of them will be computed
and compared to describe the shape of an object.
Chain-Codes
Chain-codes are connected sequences of straight-line segments of specified length and
direction used to represent a boundary. The two basic types of chain codes are shown
in Figure 4.6 [52].
Figure 4.6: 4- and 8- directional chain codes
A chain code can be generated by following a boundary and assigning a direction to
the segments connecting every pair of pixels. This method is generally unacceptable
because 1) the resulting cliain of codes is usually quite long 2) any noise or imperfect
segmentation may create small disturbances along the boundary affecting the codes,
the codes may not be reflecting the actual shape of the boundary. So care must be
taken when using this method; smoothing of the border may be required beforehand
to get rid of errors created during the digitization process. Chain code representation
is quite useful in computing the minimum bending energy which will be discussed in
the next section.
Polygonal Approximat ion
Polygonal Approximation is better in a way because it does not depend on the fine
details of the boundary, as the chain-code approximation does. It is also information
preserving because the original data rnay be represented as closely as one wants by
choosing a high enough nurnber of polygon vertices. This method will find an ap-
proximating polygon to a digital curve provided the Euclidean distance (1) between
two curves is compared to a threshold (d). This method starts off by finding two
~ " I I A V Y VU "I*V V-- '" "LI-" - IV LUI* UIIYUV U ) ~ U I A V -II- VY-----vu--- O --a"" 'J - '-'-a O--" ------
Then for al1 points on the curve, I is cornputed from the curve to this line. The point
with largest 1 that is the point on the original curve farthest from the approximating
curve will be chosen to be the splitting point. Then the initial point is connected
to the splitting point and the whole procedure is repeated until al1 of the 1 on the
curve obtained are less than the threshold. An example of this method is shown in
Figure 4.7.
2' I ~araliel to coordinate
I axes
Figure 4.7: The first few iterative technique for generating a polygonal approximation to a curve. The initial nodes 1 and 1' are chosen arbitrary. Nodes 2 and 2' are gen- erated, since neither satisfies the distance criterion. At this point the approximation is given by l-2-1'-2'-1. Then each of the curve segments can be split independently. For example, segment 1-2' is split into 1-3 and 3-2'; segment 1-3 into 1-4 and 4-3, and so on until the optimization criterion is satisfied.
The threshold (d) will limit the number of sides or vertices of the approximated
polygon, thus it will control the similarity of the polygon to the actual curve. The
smaller the threshold, the approximation resembled the actual curve more, thus pre-
serving information on the curve. Although, there are many other polygonal approx-
imation methods such as split-and-rnerge method [42], functional approximation [43],
conventional techniques such as Newton's method [44] and minimax method 1451,
... etc, the method proposed in this thesis is chosen because its convergence is fast
and the metliod is quite robust, the result not being too sensitive to noise on the
boundary. The obtained polygonal approximation to the original boundary is shown
in Figure 4.18 and Figure 4.19. On these graphs, the polygonal approximation is laid
on top of the original tumor; the number of vertices and threshold are provided in
the Table 4.1.
Table 4.1: Number of Vertices and Threshold used on the Approximated Polygon
Incrernental Curvature
F'rom classical geometry, inaxima, minima and points of inflection indicate the curva-
ture of a curve [49]. Wliile, discontinuities in curvature, endpoints and intersections
are also considered as informative critical points [50]. The incremental curvature plot
presented here will be useful to derive important features on the boundary curve that
can help to determine border irregularity. The technique used to complete the plot
is called line-segment-scaii method [51] which is based on the computation of the dis-
crete average slope a t eacli point along the object boundary. The slope is measured
based on a moving average involving W nodes in the chain. The window size W will
govern the amount of noise filtering and will normally lie in the range 4 to 9. Angle
0, is cornputed with respect to the coordinate axes of a line segment as follows:
Threshold 5 5 5 5
Image Superficial Spreading Nodular Lentigo Seborrheic Keratosis
Number of Vertices 31 36 32 13
the variables xl, yl represent the x, y components of the chain-link vectors and can
assume values of 1, 0, -1. Tlierefore, the incremental curvature bk at node k is given
The feature interpretations [50] are listed in Table 4.2 which gives the interpreta-
tion of the incremental curves indicating the shape properties of the object boundaries.
Figure 4.20 and Figure 4.21 show the incremental curvature plot of the four tumors.
Places where there is a high curvature and other discontinuities or properties can be
observed easily. Thus, border irregularity can be determined from the plot without
any difficulty.
Looking a t the plots, variable amount of peaks and valleys can be seen on the plots
of the three malignant inelanoma tumors. These can be interpreted as curvature
discontinuities and high curvature area. For a Seborrheic Keratosis mole, only a
nearly straight line is sliown in the plot which means it has a constant curvature.
Thus this particular mole has a regular border.
Signature
A boundary may be represented as a 1-D functional representation known as signa-
ture. It may be generated in various ways; one of the simplest is to plot r(0) curve.
~ ( 0 ) is the distance froin the centroid to the boundary as a function of angle, an
example of a circle and square's signature are shown in Figure 4.8.
'l'able 4.2: Feature lnterpretation
Incremental Cuwature Plot Horizontal Line Large-Magnitude Value f ositive Value Negative Value Zero Value Zero-crossing Peak or vaïIey of width w+3. D
and sum val1
Pairs of opposite-sign peaks of width 2 and magnitude arctan l/w, separated by w-2 points of constant value Increasing (decreasing) mean slope
Shape Interpretation
The definition of ceiitroid is:
Constant Curvature High Curvature Curvature toward left (bay) Curvature toward right (peninsula) Straight Line Point of Inflection Curvature discontinuity w/2 units towa right of peak (valley) center and of angular change of D/2 degrees Straight line or gentle curve
Inward (outward) spiral
where mi are the mass of each pixel, (xi, yi) are the coordinates of each pixel.
Because al1 pixels are assumed to be of the same mass ml = m2 = . = m,, the
definition of centroid becomes the following:
Figure 4.5: a) Circle's Signature b) Square's Signature
Centroids of the tumors are listed in Table 4.3.
Table 4.3: Centoids
Nodular (269,304) Lentigo (262,264)
The corresponding signatures are shown in Figure 4.22 and Figure 4.23. Smooth-
ness of the border can be observed from the plots. These curves are then represented
by polynomials and FFT is performed on them. Figture 4.24 and Figure 4.25 show
the PowerSpectrurn of the signatures. It can be seen that the more and talIer the side
loops, the more irregular the border is. An experimental threshold can be chosen for
the Power Spectrum as a criteria to decide if the plot indicates a regular or irregular
border.
It can be observed that the super tumor has the most irregular border under this
representation because its corresponding Power Spectrum hm the most nurnber of side
ioops. Again, SebOr 1s the least irregular shape with a nearly horizontal signature.
Distance Vs angle is not the only way to generate a signature. For example, the
angle between a line tangent to the boundary and a reference line can be plotted
against the position along the boundary. The resulting signature would carry in-
formation about basic shape characteristics. A variation of the above approach in
obtaining the signature of an object is the slope density function which is basically
a histogram of tangent-ailgle. I t responds strongly to sections of the boundary with
constant tangent angles and has deep valleys in sections producing rapidly varying
angles (corners or sharp inflections).
Curvature Histogram
Curvature is defined as the rate of change of slope. Curvature histogram can be
used to obtain information on the " wiggliness" of a curve. The curvatures will be
more concentrated near zero for a smooth curve; while more spread out for a wiggly
curve. To obtain the slope, the chain-code representation of the curve is used. Let
a point on the boundasy be P, k-slope a t P is the average of the unit slopes (i.e.
codes) a t a sequence of 1-points centered at P. This way of defining the slope is a
type of smoothing in order to rneasure a more continuous range of slopes. Therefore,
k-curvature of P is the difference between its left and right k-slopes. Figure 4.26 and
Figure 4.27 shows the 1-curvature histogram of the four tumors. I t can be seen that
al1 of them shows wiggly curves because the curvatures are spreaded out around zero.
k=l is used here in order to avoid too much smoothing which would affect the result.
Actually, different k values can be used depending on the curve in question. A note
worth mentioning is that the k-curvatures fluctuate for small k and approach O as k
gets large. Simply looking a t the zero-point on the plots, the regularity of the curve
can be determined. A tlireshold can also be set as a tolerance margin around the
zero point to classify between regular and irregular border. It can be seen that sebor
is the least irregular, because its curve is more concentrated around zero.
Circularity or Compactness
The most popular method used to measure the compactness or circularity of a tumor
is
where P is the nurnber of pixels on the tumor border and A is the number of pixels
within and on the border
Al takes on a, minimum value of unity for a circle and larger values for distortions
therefrom. The ratio increases when the shape becomes irregular or if its border
becomes wiggly. It was found that a threshold of 1.8 gives best separation between
regular and irregular border as mentioned in chapter 2. Therefore any index of value
1.8 or greater indicates irregular border. This method is simple t o use but i t has
a problem that Iimits its usefulness. Rom Figure 4.9, it can be seen shapes that
are very different should have a different degree of irregularity, but they possess an
identical value Al. This is because the way P and A is computed are not affected by
the specific order of the shape discontinuities.
So different shapes witli the sarne area and perimeter will have the same circularity
AI. Therefore this method is not information preserving because it achieves the same
value of Al for widely different shapes and is not accurate enough to detect border
irregularity. Other shape measures are proposed below.
Angle Regular i ty
Contour complexity is concerned with the extent of jaggedness of the boundary, that is
whether the contour contains any undulations or if it is smooth. One might refer this
as the "texture" of the object contour. The simplest measure of contour complexity
is angle regularity 1461.
Figure 4.9: Different shapes with the same area and perimeter
n is the number of vertices obtained from polygonal approximation described
earlier.
A . = { ' for least complex contour
1 or larger for most complex shapes
Side Regularity
This factor measures the reguiarity of the "sides" of the shape; that is, "are they uni-
formly distributed or are there large variations in side lengths?" This method applies
- - * A * Y - , - " - polygon having exactly the same number of sides and total perimeter P. P is obtained
from its chain code representation by counting horizontal and vertical moves as 1 and
diagonal moves as fi. For 8-directional chain code, al1 the odd codes movement are
considered to have a length of fi; while the even ones are 1 unit length for a unit
length grid.
For a regular polygon with n sides of length L = e, the side regularity [47] is:
where lk is the length of the kth side of the approximated polygon. Note that
A . = { ' for a regular polygon (4.25) max value of 1 for the rnost nonuni f orm shape
Minimum Bending Energy
This category of shape descriptor encompasses the macroconcepts of roundness with
the assumption that a circle represents the most stable figure. Clearly, the objective
is to be able to find an appropriate feature which can extract the essence of the shape
while ignoring al1 irrelevant surface irregularities. The minimum bending energy of
the boundary curve [47] is an interesting shape feature in this category.
where P is the perimeter of the arbitrary shape, m is the total number of boundary
points and
where CC(k) is the chain code for the kth link and
2 for 1 = even Li ' ) = { '
a 2 for z =odd
= { "'-
1 for
This feature is particularly useful
biomedical image processing.
a circle
most complex shape
for the bloblike shapes that often occur in
Geometric Mean
Of al1 the above, no simple shape attribute is sufficient to describe an object [48].
Clearly, each of them proposes a different ordering, depending on the unique shape
measure and a combination is often useful. The geometric mean of the angle regularity
(contour complexity), side regularity (shape uniformity) and global shape (average
bending energy) is defined as Ag = (A2 x A3 x A ~ ) $ A threshold can be set by
experimentation, so that when Ag is below the threshold, the object shape is not
complex but simple, while anything above is considered to be complex.
4.3.3 Results
The five measures for the four tumors and the interpretation are listed in the Table 4.4
Looking a t Al, only the superficial spreading and nodular melanoma have irregular
borders because they have a value greater than 1.8. For the other measures, al1 the
malignant melanorna mole shows variable degree of irregularity in the border and
complexity of the shape. The most obvious result is from A4 (Minimum Bending
energy), where the values for super, node and lent al1 approach 1 which means that
the shapes are cornplex. The results for sebor are very different from the rest of
the tested tumors: they show that the mole is less complex and less irregular in the
border. One can order the shapes according to specific features as shown in Table 4.5.
Table 4.5: Order of shapes
From Table 4.5, it can be seen that the result of A3 is very different from the rest,
of the measures. This may be because this measures the side regularity of a shape; it
compares the side length of the approximated polygon with that of a regular polygon.
The approximated polygon is a regular polygon only if the length between each pair
of vertex is the same. While the other measures A2 is a measure of the interna1
angle of a vertex and Aq lneasures the average bending energy along the length of the
boundary curve in terms of the curvature. Both of them do not deal with the actual
length of each segment of curve between the vertices. Actually al1 three of them
approach the same problem in three different ways and no simple one is suficient to
describe an object. Therefore, the geometric mean is designed to combine the good
essence of each of them. It can be seen that Lentigo Malignant Melanoma has got a
least irregular border among the other tested malignant melanoma mole.
A5 Super
2 3 4
Rank 1
A2 Super
Al Super ode Lent
Sebor
A3 Node
Node Lent
Sebor
A4
Super Lent
Sebor Sebor
Node Lent
Sebor
According t o the Canadian Cancer Society, asymmetry of a tumor means "one-half
unlike the other half." Tlie newest and less complex method is used to find the
principal axes and rnatching the two halves of the tumor about the principal axes to
determine the degree of symmetry they have to decide if the tumor is symmetrical or
not. The degree of asymmetry is determined from:
AAmin x 100% Asymrnetry index = - A
4.4.1 Asymmetry Index
The steps to calculate the Asymmetry Index are as follows:
Step-1 The orientation angle is defined as the angle between the x-axis and axis
around whicli the object can be rotated with minimum inertia is
where mlSl,m2,o and rno,:! are the second order moments or moment of
inertia defined as
where (xo, go) is the centroid.
The orientation angles for the 4 tumors are listed in the TabIe 4.6.
Step-2 The image is rotated by -O0 to align the x, y coordinates with centroidal
principal axes.
Step-3 The image is tlien reflected about the principal axes.
Step-4 The image is tlien subtracted on one side of the axis from the reflected
image and the miniinum area difference is obtained as AAmi,.
Image Orientation Angle Superficial Spreading Nodular Lentigo Seborrheic I<eratosis
Step-5 Area of the tumor image is obtained by counting the number of pixels
on the tumor as A.
Step-6 The values of asyirimetry index are listed in the Table 4.7
It was found thût tumors with an index of 6% or more are classified as
asymmetric [3]. So the only symmetric mole is seborrheic keratosis, a
non-inelanorna mole.
Table 4.7: Asymmetry Index
1 Nodular 1 6.37 1 Asymmetric 1 Image Superficial Spreading
4.5 Observations
index 8.6
Lentigo Seborrlieic Iha tos i s
4.5.1 Border Irregularity Analysis
interpetution Asymmetric
For border irregularity analysis, many of the existing techniques divided the analysis
into two stages, the first is to detect the tumor boundary; the second is to calculate
an irregularity index. For the first stage, the most popular technique used is Radial
Search Method, but it assumes a tumor boundary does not intersect with itself. A
11.12 2.1
Asymmetric Symmetric
violation 01 tnis assumption wiii resulG 111 a laise ouruer ueiiit; uewueu. A M ~ U U ~ I ~
there are other methods like thresholding or colour space segmentation to obtain
the border, but they are not accurate and are complicated to use. For the second
stage, an irregularity index Ai = & is used. Although this is simple to use, it was
found that very different shapes should have different degree of irregularity, but they
possess an identical value AI. So any shape with the same area and perimeter will
have the same irregularity index; consequently, this method is not accurate at al1
in determining border irregularity. In this thesis, some new rnethods were tried to
perform border irregularity analysis.
Edge Detection
Since the tumor has been extracted from the background already, colour information is
not needed anymore, a simple Sobel operator is performed on the gray-scale extracted
tumor image. This operator is chosen because i t has the advantage of providing
both a differencing and srnoothing effect. Hit-or-Miss 'Ikansform is then applied to
remove the inherent edge broadening in the gradient image without destroying the
edge continuity of the image and other post-processing is done to the tumor edge to
obtain a one-pixel wide edge for use in the border irregularity analysis. The obtained
one-pixel wide images are shown in Figure 4.16 and Figure 4.17.
Border Representation
Boundary of an image is represented in many useful ways like polygonal approxima-
tion, incremental curvature, signature or curvature histogram, t o represent the 2-D
curves in a 1-D way. They can be divided into two types, one type of border represen-
tation is just a representation of the border used for further analysis; while the other
type can actually classify a border to be irregular or not like the signature, curvature . .
histogram and incremental curvature. For example, Table 4.2 gives a full feature in-
terpretation of the incremental curvature plot. From the plot, different critical points
on the curve can be observed. The irregularity of the border can be easily observed
frorn these plots. Cornbining the results from al1 three of them, it was found that
sebor has a regular border and super rias a most cornplex Poraer. nll tnree or tne
plots show similar result which means that they are quite consistent in determining
border irregularity.
Goodness of Shape
Apart from observations from the plot, some factors like circularity index Al, angle
regularity mesure A2, side regularity A3, minimum bending energy A4 and the geo-
metric mean Ag. Ai is similar to the irregularity index mentioned before, so it is not
useful at all. Referring to Table 4.4 and Table 4.5, most of the rneasures show the
same result or ranking, except A3. Since no simple shape attribute is sufficient to de-
scribe an object, it is better to use a combination of them which is As, the geometric
mean of A2, A3, Aq. An experirnental threshold can be obtained to determine which
shape has an irregular border or not. The final result shows that super has the most
irregular border; while sober has the least irregular border. This result is consistent
with the results obtained from various plots in the previous section.
4.5.2 Asymmetry Analysis
An important aspect of shape understanding is symmetry, which is very useful in
pattern analysis. For a symmetric pattern, one needs store only one half of the
pattern along with the mis of symmetry. If a part of the pattern is missing or noisy,
with the help of symmetry one can complete the pattern or rid the pattern of noise.
To check for the degree of symmetry of a tumor, an asymmetry index was used in
this thesis. First of all, the tumor shape was obtained by Colour Segmentation, this
is a more simple and faster algorithm than going through steps of radial search, line
drawing, connectivity filling just to get a planar shape. Then an orientation angle
is obtained via the calculation of second order moments. Second order moments
are not affected by object transition, rotation and size; they also carry a "physical"
interpretation of the boundary shape. The orientation angle will he1p in locating the
axis of symmetry. The values of the asymmetry index are shown in Table 4.7. I t was
found that only Seborrheic Keratosis is symmetric.
4.6 Analysis Result s of Non-Melanoma Mole
The tested tumor is a Sclerosing Hemangioma shown in Figure 4.28. This dome
shaped nodule is notable for its striking symmetry and well-defined borders. The
lesion also exhibits a dark brown, dusky colour. Here, the proposed methods in this
thesis will be used to perform the different analyses on the turnor and compare the
experimental results to the clinically obtained results mentioned above. As can be
seen from the original image, there are many non-tumor artifacts around the tumor,
a filtered image is shown in Figure 4.28. The filter used is the modified vector median
filter, the skin marks are gone and the shape is preserved. With appropriate masking
and vector directional segmentation, the extracted image is shown in Figure 4.29.
On the right is the L*a*b* space segmented tumor, only one colour exists within the
tumor. This is consistent with the clinical description that it has only got a dark
brown colour, with no other colour exists within the border.
Table 4.8: Border Irregularity Measures and Asymmetry Index
Figure 4.30 shows the one-pixel wide tumor border, Figure 4.31 to Figure 4.33
shows the plots to Border Irregularity Analysis. Combining al1 the results from the
plots, a conclusion can be drawn about the lesion border - Regular. Apart from the
plots, other factors are shown in Table 4.8 to support the above argument.
Al1 the goodness of shape factors are consistent with the results from the plots.
Factors Al A2
A3
A4
A5 Orientation Angle Asyrnmetry Index
Values 3.7691
0.83810 0.03972 0.28100 0.21071
31 2.55%
'l'hese resuits snow tmt tiie tumor nas a well-aennea Doraer. tasr; 01 ail, m e asym-
metry index of 2.55% is less than 6% has indicated that the tumor is syrninetric.
Al1 tliree criteria show consistency with the clinical results which indicates that the
methods proposed in this thesis are capable of recognizing potential melanoma moles
or distinguishing between melanoma and non-melanoma moles.
Figure 4.10: Edge Detection using Sobel Operator (left) Superficial Spreading and (right) Nodular Malignant Melanoma
Figure 4.11: Detection using Sobel Operator (left) Lentigo Malignant Melanoma and (right) Seborrheic I<eratosis Mole
Figure 4.12: Preprocessed and Edge Detection (left) Superficial Spreading and (right) Nodular Malignant Melanoma
Figure 4.13: Preprocessed and Edge Detection (left) Lentigo Malignant Melanoma and (right) Seborrheic Keratosis Mole
Figure 4.14: First Thinning (left) Superficial Spreading and (right) Nodular Malig- nant Melanorna
Figure 4.15: First Thinning (left) Lentigo Malignant Melanoma and (right) Seborrheic Keratosis Mole
Figure 4.16: 2nd Thinning (left) Superficial Spreading and (right) Nodular Malignant Melanoma
Figure 4.17: 2nd Thinning (left) Lentigo Malignant Melanoma and (right) Seborrheic Keratosis Mole
Figure 4.18: Polygonal Approximation (left) Superficial Spreading and (right) Nodu- lar Malignant Melanoma (dashed linc: approximated tumor; solid line: original tu- mor)
row s
Figure 4.19: Polygonal Approximation (left) Lentigo and (right) Sebor (dashed line: approxirnated tumor; solid line: original tumor)
-200 -- zoo r a o eoa am t o m i z m t roa pkrlnumbu
Figure 4.20: Incremental Curvature (left) Superficial Spreading and (riglit) Nodular Malignant Melanoma
-200 1 I 1 I 1 I I I a m o r o a aoo too tooo m o 1400 IWO
pk4 n m b u
Figure 4.21: Incremental Curvature (left) Lentigo Malignant Melanoma and (right) Seborrheic Keratosis Mole
Figure 4.22: Signature (left) superficial Spreading and (right) Nodular Malignant Melanoma
300 500 -
250 - 250 -
O - 2 2 0 0 - U
- g m -
= too- 3 3
Figure 4.23: Signature (left) Lentigo Malignant Melanoma and (right) Seborrheic I<eratosis Mole
5 0 -
'0 50 IO0 150 200 250 500 350 100 '0 ;O O !;O ttk i l 0 5 h &O 400 Anglo mado hem tdh 8? r4X.k Angl. mido llom idlrn m a ~ n k
50
I , I I I I
-
Figure 4.24: Power Spectrum (left) Superficial Spreading and (right) Nodular Malig- nant Melanoma
Figure 4.25: Power Spectrum (left) Lentigo Malignant Melanoma and (right) Sebor- rheic Keratosis Mole
Figure 4.26: Curvature Histogram (left) Superficial Spreading and (Nght) Nodular Malignant Melanoma
Figure 4.27: Curvature Histogram (left) Lentigo Malignant Melanoma and (right) Seborrheic Keratosis Mole
Figure 4.28: (left) Original Scler Image (right) Filtered Scler Image
Figure 4.29: (left) Extracted Scler Turnor (right) Segmented Scler Tumor
Figure 4.32: (left) Signature Plot (right) Fourier Spectrum
chah code 8llfrance
Figure 4.33: Curvature Histogram
Chapter 5
Conclusions
5.1 Discussion
F'rorn Chapter 3 and 4, the Colour Variegation, Border Irregularity, Asymmetry and
Common Mole Analysis results are summarized in Table 5.1 for each of the lesion
tested.
Table 5.1: Surnmary of the Classification of Malignant Melanoma
A nalysis Colours Al A2 A3 A4 A5 Border Asym Index Interpretation Classification
Super 9
2.1736 2.8924 0.1528 0.9999 0.7617
Irregular 8.6%
Melanorna
9 2.1260 1.5420 0.2017 0.9999 0.6775
Irregular 6.37%
Asymmetric Melanoma
Lent 10
1.5312 1.1832 0.1123 0.9999 0.5103
Irregular 11.12%
Asymmetric Melanoma
Sebor 3
1.1210 0.9276 0.0531 0.7050 0.3263
Regular 2.1%
For Border Irregularity Analysis, the decision is made from observations ob-
tained from the different boundary representation plots together with the measures
A2, AS, A4, A5. Ai is not included because of its inherent problem. The plots can
Scler 1
0.7690 0.8381 0.3972 0.2810 0.2107
Regular 2.55%
I
Symmetric Melanoma
Symmetric Non-melanom
provide an accurate visualization of the border irregularities, while the factors mea-
sured can provide a numerical representation of irregularity. Al1 of them have to be
taken into consideration, each of them will provide something important. For exam-
ple, incremental curvature can be used to locate critical points on the border; that
is, places where there is a curvature discontinuity or high curvature. Signatures can
show the overall wiggliness of the border and the main advantage of tliis plot is that
the process is reversible. The tumor border can be redrawn from the plot. As dis-
cussed before, A5 is a unique shape measure because i t is a combination of contour
complexity, shape uniformity and global shape measures. So rnost of the time A5 will
be looked at first and the graphs will help in making the decision. It is necessary to
refer to the individual measures only when there is doubt in making the decision.
Note that al1 three categories are necessary to identify rnalignant melanoma tu-
mor, because a number of the pigmented lesions look very similar to malignant
melanoma. For example, Seborrheic Keratosis can be described as a turnor that
mimics melanoma. Very often, it is mistaken as melanoma turnor because of its sim-
ilarity. This is why a Seborrheic Keratosis tumor is chosen as a test image. It can
be seen here. The test sebor image is symmetric and has a regular border, but it
is colour variegated, and would be classified as a potential melanoma tumor. So if
a positive result is obtained in any one of the analyses, the tumor is classified as a
possible malignant melanoma tumor. We cannot take any chance in these classifi-
cations because the tumor is a risk to the life of the patient. Until now, there has
only been a discussion on melanoma tumor or other lesion that mimics melanoma.
The reader may think that the proposed method will only classify everything as a
possible melanoma mole, but the previous section has provided the test results of a
non-melanoma-like tumor. It can be'seen that the scler tumor has shown negative
results in al1 the analyses, which means it is a non-colour variegated turnor, having a
regular border and symmetric in shape. For the other melanoma tumors, the overall
result of al1 the analyses has correctly identified them as a melanoma mole, except
Seborrheic Keratosis, which often appears similar to a melanoma mole.
5.x rromerns wim &ne proposeu rxlecnous
œ Reflections or Shadows As can be seen from the original images in Figure 1.2
and Figure 1.3, there is a t least one area on the tumor with a bright white
spot. This is not a kind of colour on the tumor but is a reflection caused by
the electronic Aash unit during photograph taking. It has created problem in
turnor extraction, because the white part is often not included as part of the
turnor. The reflections and shadows can be avoided by taking the photogrepli
in a controlled environment.
Views of Tumor The view of the tumor is very important, because the top view
or side view of an image will give a different result in the analysis. Especially
with border irregularity and asymmetry analysis, a wrong border could be found
if a side view of the tumor is taken. This can be avoided by controlling the
photograph taking procedure.
Location of the t u m o r The test images used were al1 on the skin that is easy
to analyze, but when the tumor is on some hidden area, for example under a
finger or toe nail, or on either the upper or lower eyelid, it is difficult to obtain
an accurate picture of the tumor.
Colour of the Skin So far, the test images used were from fair skin patients.
There would Lie a problem in colour segmentation if a dark brown coloured
tumor is grown on a dark skin person.
Size Analysis Size Analysis could not be done because the images obtained
were from a wide variety of sources. They may be taken with varying rnagni-
fication or under different environmental disturbances. Tumor size could not
be estimated from the images accurately. This can again be avoided if the
photograph taking procedures and environment are controlled.
r Take photographs of the images under a controlled environment.
a Investigate and find ways to eliminate the reflections present on the images.
r 'Tky Neural Network in the classification of the tumor given the results from the
different anal ysis.
5.4 Benefits of Digital Imaging in Dermatology
a Biopsies are not required in the detection of malignant melanoma. Moreover,
digital images of the tumor can be stored and sent more efficiently.
A 3-dimensional extent of a tumor is possible by analyzing images of the tumor
taken from different views.
r Diagnostics can be done by means of a cornputer diagnostic system - expert
system.
5.5 Conclusion
The incidence of malignant melanoma has risen drarnatically in recent years. Fast
and effective detection methods are needed desperately to Save thousands of life each
year. It has been known for many years that most skin cancers including melanoma,
are curable if treated a t an early stage [62]. Many algorithms for detecting different
features " ABCD" of early malignant melanorna have been discussed.
For A = Asymmetry, asymmetry remains undefined up to now. Dermatologists
only consider a nearly symmetric tumor symmetric - there is no absolute division
between symmetric and non-symmetric tumor yet. There is still many disagreement
between clinicians on decision for tumors that are moderately asymrnetric. Since
asymmetry is a critical feature in the diagnosis of malignant melanoma, an asymmetry
index was tested in this thesis as a measure to determine the degree of symmetry,
A I V p A A A 6 U V bllV U I I V u u A r r r u u v 1 v ~ r v v u W &\&CC- u- y----u------- - ----- "- S v --- m . -wu. - --- method is simple and fast; it is designed based on the geometrical characteristic of
a planar shape. In this thesis, colour segmentation was used to obtain the tumor
shape. With the principal axes found, asymmetry was determined about the near-
axis of symrnetry by cornparing absolute area differences to the total area of the
shape. The results in this thesis have shown that this method is adequate; in fact,
clinically test results have shown quite a high success rate [ô].
For B = Border Irregulari ty, the border of the turnor was obtained from gray-
scale edge detection. Then an overall measure for detecting border irregularity was
found to be the Geometric Mean measure and the results obtained were in agreement
with the tumors tested. It was also stated that a combination of the visualization
of the borders and other goodness of shape measure will be useful in making tlie
decision. In this way, the decision of whether the border is irregular or not can be
made more accurately.
For C = Colour Variegation, results in this thesis demonstrated the success
of image segmentation by colour using uniform colour spaces. Therefore, this colour
segmentation algorithm is successful as a n aid in finding the tumor border and dif-
ferent colours in a tumor; moreover, it may be successful as an aid in finding ulcer
or reflections. This colour spaçe's interpretation of colour is most closely related to
human perception because it has accdunted for the nonlinear response to luminance.
Together with the colour-difference formulae, this space is very useful in the precise
evaluation of perceptual closeness between two colours.
For D = Diameter generally grea ter than 6mm, since the slides used in
the study were mostly obtained from various sources and photographed with varying
magnifications, tumor size could hardly be estimated from the images accurately. So
tumor size or diameter was accordingly omitted in the discussion. This is actually one
of the important features of early Malignant Melanorna, so inclusion of this analysis
in the classification of the tumor is a must.
When a mole is suspected to be a melanoma mole, it must go through al1 four
analyses " ABCD" , not just the first t h e which were described in this thesis. Sliow-
i i i ~ ~ J U J L ~ J L V C LGSUAU LH a u y u i r c UL birc auuryaio , u r i b u u r r i u i r a L L ~ J J L L L G U aa a p v u b i i u r u r
melanoma mole. In fact, the classification result of this thesis is not complete, be-
cause only three clinical features of early malignant melanoma are being looked at. A
classified "non-melanoma" may be found to be a melanoma tumor after it hm gone
through the size analysis, so size analysis should not be omitted.
After all, the best way to prevent ourselves from getting this disease is to look after
ourselves carefully, to stay away from the Sun, and keep our body as healthy as we
can. I t is better and easier to prevent than to find a cure. According to research done
by the Canadian Dermatology Association, it was found that most skin cancers are
preventable [62]. Moreover, understanding ourselves, being observable, and staying
alert to pigmented spots on the skiii are al1 good ways to reduce the chance of getting
any slcin cancer which is a risk to human life.
Bibliography
[l] A. Dhawan, An Expert Systern. for the Early Detection of Melanoma Using
Knowledge-Based Image Analysis. Analytical and Quantitative Cgtology and His-
tology, vol. 10, pp. 405 - 416, 1989.
[2] R. Kashyap, A. Khotanzad, A Model-Based Method for Rotation Invariant Tex-
ture Classification. IEEE Transactions on Pattern Analysis and machine intel-
ligence, vol. 8, no. 4, pp. 472 - 480, July, 1986.
[3] S. Umbargh, R. Moss, W. Stoecker, Applying Artificial intelligence to the iden-
tification of variegated coloting in skin tumors. IEEE Engineering in Medicine
and Biology, vol. 10, pp. 57 - 62, 1991.
[4] S. Umbaugh, R. Moss, W. Stoecker, An Automatic Color Segmentation Algo-
rit hm with application to identification of skin tumor borders. Co mputerized
Medical Imaging and Graphics, vol. 16, no. 3, pp. 227 - 235, 1992.
[5] S. Urnbaugh, R. Moss, W. Stoecker, Automatic Colour Segmentation of images
with application to detection of variegated colouring in skin tumors. IEEB En-
gineering in Medicine and Biology, vol. 8, pp. 43 - 52, December 1989.
[6] W. Stoecker, W. Li, R. Moss, Automatic Detection of Asyrnmetry in skin tumors.
Computerized Medical Imaging and Graphics, vol. 16, no. 3, pp. 191 - 197, 1992.
[7] J. Golston, W. Stoecker, R. Moss, Automatic Detection of Irregular Borders in
Melanoma and other slcin turnors. Computerized Medical Imaging and Graphics,
vol. 16, no. 3, pp. 199 - 203, 1992.
101 11. vv IIIIK, U. 1- I;Lt;uIlla, IL. IJLIIUWCII~CLUI~, f i u ~ i u ~ ~ l a b t m I~;CLLULT; u~;~ir;~lrluil 111 u q y b a i
images of skin. Computer Methods and Prograrns Biomedicine, vol. 34, pp. 41 -
60, 1991.
[9] J. Golston, R. Moss, W. Stoecker, Boundary Detection in Skin Turnor Images:
An Overall Approach and A Radial Search Algorithm. Pattern Recognition, vol.
23, no. 11, pp. 1235 - 1247, 1990.
[IO] Y. Ohta, T. Kanade, T. Sakai, Colour Information for Region Segmentation.
Computer Graphics and Image Processing, vol. 13, pp. 222 - 241, 1980.
[ I l ] F. Ercal, M. Moganti, W. Stoecker, R. Moss, Detection of Skin Turnor Boundaries
in Color Images. IEEE Transactions on Medical Imaging, vol. 12, no. 3, pp. 624
- 627, September 1993.
[12] W. Stoecker, R. Moss, Editorial: Digital Imaging in Dermatology. Computerzzed
Medical Imaging and Graphics, xol. 16, no. 3, pp. 145 - 150, 1992.
[13] A. Dhawan, Early detection of cutaneous malignant melanoma by three-
dimensional nevoscopy. Computer Methods and Programs, in Biomedicine, vol.
21, pp. 59 - 68, 1985.
[14] R. Friedman, D. Rigel, A. Kopf, Early Detection of Malignant Melanoma: The
Role of Physical Examination and Self-Examination of the Skin. CA-A Cancer
Journal for Clinicians, vol. 35, no. 3, pp. 131 - 151, MayIJune, 1985.
[15] M Pietikainen, A. Rosenfeld, Image Segmentation by Texture using Pyramid
Node Linking. IEEE fiansactions on Systems, Man, and Cybernetics, vol. 11,
no. 12, December, 1981.
[16] H. Antonisse, Image Segmentation in Pyramids. Computer Graphics ande Image
Processing, vol. 19, pp. 367 - 383, 1982.
[17] D. Perednia, R. White, R. Schowengerdt, Localization of lesions in digital images
of skin. Cornputers and Biomedical Research, vol. 22, pp. 374 - 392, 1989.
LlUj 1'. U L L L t l ) fi. ullawia, v v . UiluGbnbr, A r . ubb, ~ b . ~ v r v u u , A ~ b u r u i A I ~ U V ~ U A A ~ L I I C i U & , A A V U I U
of Malignant Melanoma From Color Images. IEEE Transactions on Biomedical
Engineering, vol. 41, no. 9, pp. 837 - 844, September 1994.
1191 A. Dhawan, R. Gordon, R. Rangayyan, Nevoscopy: Three-Dimensional Com-
puted Tornography of Nevi and Melanomas it In Situ by Transillumination. IEEE
Transactions on Medical Imaging, vol. MI-3, no. 2, pp. 54 - 60, June 1984.
[20] P. Chen, T. Pavlidis, Segmentation by Texture Using a Co-Occurrence Matrix
and a Split-and-Merge Algorithm. Computer Graphics and Image Processing,
vol. 10, pp. 172 - 182, 1979.
[21] A. Dhawan, A. Sicsu, Segmentation of Images of Skin Lesions using Colour and
Texture Information of Surface Pigmentation. Computeriaed Medical imaging
and Graphics, vol. 16, no. 3, pp. 163 - 177, 1982.
[22] R. Moss, W. Stoecker, S. Lin, S. Muruganandhan, K. Chu, K. Poneleit, C.
Mitchell, S kin Cancer Recognition by Computer Vision, Computerized Medical
Imaging and Graphics, vol. 13, no. 1, pp. 31 - 36, 1989.
[23] J. Du Buf, M. Kardan, M. Spann, Texture Feature performance for Image Seg-
mentation. Pattern Recognition, vol. 23, no. 314, pp. 291 - 309, 1990.
[24] W. Stoecker, C. Chiang, R. Moss, Texture in Skin Images: Comparison of three
nlethods to determine smoothness. Computerized Medical Imaging and Graphics,
vol. 16, no. 3, pp. 179 - 190, 1992.
[25] S. Pavlidis, Algorithms for Graphies and Image Processing. Computer Science
Press, 1982.
[26] W. Stoecker, Computer Applications in Dennatology. Igaku-Shoin Medical Pul-
ishers, Inc, 1993.
[27] R. Gonzalez, R. Woods, Digital Image Processing. Addison-Wesley Publishing
Company, 1993.
[29] E. Holzle, P. Kind, G. Plewig, W. Burgdorf, Malignant Melanoma Diagnosis and
Differential Diagnosis. Schattauer, New York, 1993.
[30] A. Rosenfeld, Multiresohtion Image P~ocessing and Analgsis. Springer-Verlag
Berlin Heidelberg, New York, 1984.
[31] A. Emmett, M. O'Rourke, Malignant Skin Tumors. Churchill Livingstone, New
York, Second Edition, 1991.
[32] A. Vivier, Atlas of Sliin Cancer. Gower Medical Publishing, New York, 1991.
[33] A. Kopf, R. Bart, R. Andrade. Atlas of Tumors of the Skin. W . B. Saunders
Company, Toronto, 1978.
[34] R. Barnhill, T. Fitzpatrick, K. Fandry, R. Kenet, M Mihm, A. Sober, Color
Atlas and Synopsis of Pigmented Lesions. McGraw-Hill, Inc, Health Professions
Division, Toronto, 1995.
[35] D. Roses, M. Harris, A. B. Ackerman, Diagnosis and Management of Cutaneous
Matignant Melanoma. W . B. Saunders Company, Toronto,, 1983.
1361 R. Friedman, D. Rigel, A. Kopf, M. Harris, D. Baker, Cancer of the skin. W . B.
Saunders Company, Toronto, 1991.
[37] D. Crevier, Computing Statistical Properties of Hue distributions for Color Image
Analysis. Proceedings of the SPIE Conference, vol. 16, no. 3, pp. 179 - 190,
September 4 - 7, 1993, Boston.
[38] G. Wyszecki, W. S. Stiles, Color Science: Concepts and Methods, Quantative
Data and Formulae, 2nd Edition. John Wiley and Sons, New York, 1982.
[39] Yi Yang, Colour Edge Detection and Segmentation Using Vector Analysis.
M.A.Sc. thesis, University of Toronto, Department of Electrical and Engineering,
Toronto, May 1995.
L i
pp. 96 - 97, August 1996.
[41] S. Ishikawa, Geometrical Indices Characterizing Psychological Goodness of Ran-
dom Shapes. Proceedings of the 4th International Joint Conference on Pattern
Recognition, pp. 414 - 416, November 7 - 10, 1978, Kyoto, Japan.
[42] T. Pavlidis, S. Horowitz, Segmentation of Plane Curves. IEEE Transactions on
Computers, vol. C-23, no. 8, pp. 860 - 870, August, 1974.
[43] T. Pavlidis, Waveform Segmentation through Functional Approximation. IEEE
Transactions on Computers, vol.'C-22, no. 7, pp. 689 - 697, July, 1973.
[44] T. Pavlidis, Polygonal Approximations by Newton's Method. IEEE Transactions
on Computers, vol. C-26, no. 8, pp. 800 - 807, August, 1977.
[45] Y. Kurozumi, Polygonal Approximations by the MiniMax Method. Computer
Graphics and Image Processing, vol. 19, no. 3, pp. 248 - 264, July, 1982.
[46] B. W. York, Shape Representation in Computer Vision. Ph.D. thesis, COINS
Technical Report 82 -13, University of Massachusetts, Depatment of Computer
and Information Sciences, Amherst, Mass., May, 1981.
[47] Y. M. Youssef, Quantification and Characterization of the Motion and Shape
of a Moving Cell. Ph. D. thesis, Department of Electrical Engineering, McGill
University, Montreal, 1982.
[48] T. Pavlidis, Algorithms for Shape Analysis of Contours and Waveforms. IEEE
Tkansactions on Pattern Analysis and Machine Intelligence, vol. PAMI-2, no. 4,
pp. 301 - 312, July, 1980.
[49] J. R. Slagle, J. K. Dixon, Using Descriptions to find figures that approximate the
given points. Proceedings of the 5th International Conference on pattern Recog-
nition, pp. 1049 - 1054, December 1 - 4, 1980, Miami Beach.
LW", AI. L.d"AAAUAA, YAIAbU, VU& .UV L U I A U V A A U I L U ) " U " A . Y C Y Y A V A A V A V A I - y V . *""y", Y O " " . V V VVL),
Image Processing Laboratory, Electrical and Systems Engineering Department,
Rensselaer Polytechnic Institute, Troy, N. Y., March, 1980.
[51] H. F'reeman, Shape Description Via the Use of Critical Points. Pattern Recogni-
tion, vol. 10, no. 3, pp. 159 - 166, 1978.
[52] D. H. Ballard, C. H. Brown, Computer Vision. Prentice-Hall, Inc., New Jersey
1982.
[53'] S.H. Ong, C.C. Hew, Segmentation of Colour Images based on iterative thresh-
olding and merging. ICIP '92. Proceedings of the 2nd Singapore International
Conference on Image Processing, pp. 721 - 725.
[54] S.K. Parui, D.D. Majumder, Symmetry Analysis by Computer. Pattern Recog-
nition, vol. 16(1), pp. 63 - 67, 1983.
[55] J. J. Leou, W.H. Tsai, Automatic Rotational Symmetry determination for Shape
Analysis. Pattern Recognition, vol. 20(6), pp. 571 - 582, 1987.
[56] P. Heckbert, Color Image Quantization for frame buffer display. Comput. Graph.,
vol. 16(3), pp. 297 - 304, 1983.
[57] G. Campbell, Two bit/pixel full color encoding. SIGGRAPH '86 Conf PTOC.,
Comput. Graph., vol. 20(4), pp. 215 - 223, Dallas, TX, 1986.
[58] B.K.P. Horn, Robot Vision. MIT Press, Cambridge 1986.
[59] C. Sun, W. Wee, Neighboring Gray Level Dependence matrix for Texture Clas-
sification. Comput. Vision Graph. Image Process, vol. 23, pp. 341 - 352, 1982.
[60] M.L. Astion, P.Wilding, Application of neural networks to the interpretation of
laboratory data in cancer diagnosis. Clinical Chemistry, vol. 38, no. 1, pp. 34 -
38, 1991.
[ G l ] R.P. Lip, Pattern Classification using Neural Networks. IEEE Commun. Mag.,
vol. 27, no. 11, pp. 47 - 64, Nov. 1989.
APPLIED A I M G E . lnc 1653 East Main Street - -.
, Rochester, NY 14609 USA -- -- - - Phone: 7161482-0300 -- -- - - Fax: 71 61288-5989
0 1993. Applied Image. Inc.. All Righls Reserved