A Supervised Quantiﬁcation of the Color Names Characterizing the

Visual Component Color in the ABCD Dermatological Criteria for a

Further Melanoma Inspection

Jinen Daghrir

1,2

, Lotﬁ Tlig

, Moez Bouchouicha

, Noureddine Litaiem

, Faten Zeglaoui

and Mounir Sayadi

Universit ´e de Sousse, ISITCom, 4011, Hammam Sousse, Tunisia

Universit ´e de Tunis, ENSIT, Laboratory SIME, Tunisia

Aix Marseille Univ, Universit ´e de Toulon, CNRS, LIS, Toulon, France

University of Tunis El-Manar, Faculty of Medicine of Tunis, Department of Dermatology,

Charles Nicolle Hospital, Tunis, Tunisia

Keywords:

Melanoma Inspection, Medical Imaging, Color Name Extraction, Machine Learning, Computer Vision,

Image Processing.

Abstract:

Digital imaging is widely used for creating automated systems for medical purposes such as the diagnosis

of certain kinds of diseases. One typical use of these computer vision diagnosis systems in dermatology is

the inspection of melanoma skin cancer, which is one of the most fatal skin cancer. For the early detection

of melanoma, a lot of systems have been proposed. Most of them use some visual features through image

processing methods, such as color processing and border and texture inspection. Color variation is a good clue

to differentiate melanoma and benign lesions. Thus, it is important to process skin lesion images to extract

the various colors. The paper presents a new method that extracts the different color names from a skin lesion

in a supervised way based on observed skin condition types. These features can ensure accurate melanoma

detection with other types of features. To demonstrate the effectiveness of our suggested representation, we

construct a prediction system for inspecting the malignancy of skin lesions. The experimental results show a

consistent improvement in the prediction performance against other color representations.

1 INTRODUCTION

Skin cancer is the uncontrolled growth of abnormal

skin cells. It is caused by unrepaired DNA dam-

age that activates mutations or genetic defects, which

stimulate the skin cells to rapidly multiply and form

malignant tumors. Among the three main types of

skin cancer, two of them are frequently diagnosed,

which are Basal and Squamous cell carcinoma. These

are considered non-melanoma skin cancer and not

life-threatening (Khazaei et al., 2019). However,

melanomas, which is the deadliest form of skin can-

cer, are less common but they represent the most

fatal cancer since they can quickly spread to other

parts of the body. A melanoma arises through a

malignant transformation of melanocytes which are

derived from the neural crest neoplasia (Dimitriou

et al., 2018) causing about 60,000 cancer deaths in

2018 (Khazaei et al., 2019). It represents as 0.7% of

all cancer deaths. The incidence rate from 1973 to

2009 shows a rise in the number of cases (Heinzer-

ling and Eigentler, 2021) which is particularly wor-

rying. A particular interest in creating automated

systems for melanoma inspection has been the chal-

lenge of the healthcare management community. It

is now crucial to use supportive imaging to identify

melanomas at an early stage when the odds of cur-

ing it are completely high, thereby reducing mortal-

ity (Khazaei et al., 2019).

Computer-Aided Diagnosis (CAD), has been de-

signed to improve and facilitate a quick and accu-

rate diagnostic process based on strategies invented

by physicians. One widely used clinical clue is

the ABCDE signs, which is a useful indicator for

melanoma. The ABCD rule (Stolz et al., 1991) of

dermatoscopy, based on multivariate analysis of only

four criteria was introduced by Stolz et al. and ex-

panded to ABCDE in 2004. The rule represents an

analytical method for the evaluation of melanocytic

Daghrir, J., Tlig, L., Bouchouicha, M., Litaiem, N., Zeglaoui, F. and Sayadi, M.

A Supervised Quantiﬁcation of the Color Names Characterizing the Visual Component Color in the ABCD Dermatological Criteria for a Further Melanoma Inspection.

DOI: 10.5220/0010865300003188

In Proceedings of the 8th International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AWE 2022), pages 147-154

ISBN: 978-989-758-566-1; ISSN: 2184-4984

147

lesions that clinicians and the general public can uti-

lize to help detect melanoma (Abbasi et al., 2004).

Melanoma often manifests some or all of the ABCDE

features, namely asymmetry (A), border irregularity

(B), color variability (C), diameter greater than 6 mm

(D), and evolution (E) or change in the color or size

(see Fig. 1). The color variability means that the col-

Figure 1: ABCDEs of detecting melanoma: Aspects and

differences between benign and malignant skin lesions.

ors of a skin lesion are not uniform. Since most be-

nign lesions mainly contains one color, often a single

shade of brown, having a variety of colors is a warn-

ing signal as given by Fig. 1. Melanoma can con-

tain different shades of brown, black, red, white, or

blue colors. Due to the sensitivity of the ABCDE rule,

physicians tend to use other new features to recognize

melanoma. The blue-black color is one of these fea-

tures. It is deﬁned by the presence of a combination

of blue and black pigmented areas involving at least

10% of the lesion surface (Argenziano et al., 2011).

While some melanomas begin within an atypical

mole, though it can be hard to observe the different

colors in an atypical mole. However, it may be ame-

lanotic, not having any of the skin pigment that typ-

ically turns a mole brown or black. Thus, deﬁning

the different colors in a skin lesion is not a straight-

forward process. Accordingly, it becomes harder to

recognize the melanoma. Even dermatologists will

not be able to examine it by the naked eye which can

lead them to remove a portion of a mole for exami-

nation in a lab, which can delay diagnosis. The auto-

matic inspection of melanomas is composed of a vari-

ety of steps including preprocessing, extraction of the

region of interest, post-processing, and lesion inspec-

tion, which are the various steps of a classical pattern

recognition system, including image acquisition, im-

age processing, segmentation, characterization, and

classiﬁcation of the lesion in question (Maglogiannis

and Doukas, 2009a).

One important step is skin lesion characterization,

which consists in extracting a set of relevant and dis-

criminative primitives that can describe precisely a

skin lesion. These characteristics must ensure non-

redundancy, relevance, discrimination, and robustness

to noise.

The ABCD rule is widely used in automated com-

puter diagnosis systems, which investigates the asym-

metry, the border, the color, and the diameter or differ-

ential structures (Maglogiannis and Doukas, 2009a).

Other features can be employed like the seven-point

checklist (Spalding, 1993) which contains three ma-

jor aspects ( change in size, shape, and color) and

four minor aspects (diameter, inﬂammation, crusting

or bleeding, and sensory change). These criteria can

be quantitatively determined by the change of the tex-

ture, color, and structure of the skin lesion. A lot

of studies have been introduced to examine the color

characteristics inside a lesion and to deﬁne the num-

ber of colors. Skin diseases are restricted to only six

deﬁned colors, for which the color name features are

extracted to achieve accurate inspection under any il-

lumination condition.

A lesion may include light brown, dark brown,

black, red, white, and slate blue (Maglogiannis and

Doukas, 2009b). Nevertheless, the lesion can be ame-

lanotic in some cases. Thus, these color names be-

come worthless for melanoma inspection, and other

degrees of colors should be examined. Besides, the

human perception of color is very complex, as men

and women can differently describe a color. Women

can distinguish even the tiniest differences between

two colors, contrarily to men who see them identi-

cal. Hence, learning about the different color shades

which can differentiate that a malignant and benign

lesion is not easy, Hence, learning about the different

color shades which can differentiate that a malignant

and benign lesion is not easy, so our main contribution

is to build a machine learning-based method to extract

the most relevant color names. This color representa-

tion will be used then for skin lesion classiﬁcation.

Our aim is to design a low-dimensional representa-

tion that can efﬁciently detect the different colors in a

skin lesion, more speciﬁcally the black and blue color

which has been proved that it is an accurate clue for

melanoma inspection.

This paper is organized as follows. In the next

section, the most color features used in the literature

for melanoma inspection are reported. After that, we

present the way to deﬁne the color names using ma-

chine learning and use them to classify skin lesions.

Then, we introduce the conducted experiments and

the results evaluating the classiﬁcation process. So

the conclusions and future work are drawn in the last

section.

ICT4AWE 2022 - 8th International Conference on Information and Communication Technologies for Ageing Well and e-Health

148

2 RELATED WORK

Classifying skin lesions addresses the problem of

deﬁning skin lesions as malignant or benign. The

classiﬁcation process is led by some visual clues that

characterize a skin lesion. These features differentiat-

ing malignant and benign lesions should be quantiﬁed

and should have a high probability of being true clas-

siﬁed. More essentially, when detecting melanoma,

the decrease in false negatives (misclassiﬁed malig-

nant lesions) is critical.

Examining the shape, color and texture have been

the consideration of many researchers. The color

of a lesion is considered as a crucial criterion for

melanoma inspection. This is because the blue-black

rule has also proved to be a good practice in the di-

agnosis process (Argenziano et al., 2011) in addition

to the fact that the ABCD rule deﬁnes the most used

and accurate clue for dermatologists. The color fea-

tures have been examined on different color represen-

tations, such as Red, Green, Blue (RGB), HSV (Hue,

Saturation, Value) or the spherical coordinate LAB

average and variance responses for pixels. The color

inconsistency is quantiﬁed by calculating the mini-

mum, maximum, average and standard deviations for

each channel (Menzies et al., 2005; Maglogiannis and

Zaﬁropoulos, 2004a; Maglogiannis and Zaﬁropoulos,

2004b). In (Manousaki et al., 2006), the authors used

the color texture for determining the nature of skin le-

sions by measuring the lacunarity in the distribution

of colors. Furthermore, in (Yang et al., 2018), the au-

thors have tested the color SIFT (Abdel-Hakim and

Farag, 2006), which examines the color texture. An-

other interesting yet simple method is to examine the

color variegation in a lesion by calculating the vari-

ance of the local average color (Zhang et al., 2003).

Examining the uniformness of the tumor color, it can

be quantiﬁed by comparing the colors inside a le-

sion and the healthy skin colors as in (Claridge et al.,

2003). In (Yang et al., 2018), Yang et al. put forward

clinical skin lesion diagnosis using a representation

inspired by dermatologists, where the color is intro-

duced by two representations, deﬁning the different

color names and the continuous color values of le-

sions, which indicates for each pixel the probability

of belonging to a color bin.

3 PROPOSED METHOD

As a CAD system, many frameworks based on image

processing have been proposed and have proved their

efﬁciency in melanoma inspection. In the literature,

a large variety of classiﬁcation methods have been

adopted: KNN, SVM, ANN, etc. (Magalhaes et al.,

2021; Melbin and Raj, 2021a). In the last decades,

regarding the evolution of Convolutional Neural Net-

works (CNN), CAD systems have become more and

more oriented into the implementation of semantic

techniques called Deep Learning (DL) (Gonzalez-

Diaz, 2018; Saeed and Zeebaree, 2021). When us-

ing DL, the low-level features discriminating malig-

nant and benign lesions are automatically extracted.

This representation has shown a limitation in some

cases. Thus, it is important to extract hand-crafted

features that have been used and proved by dermatolo-

gists for the diagnosis process. Generally, these prim-

itives represent only the ABCD rule that describes the

color, the border and the texture to ﬁnd some dif-

ferential structures (Daghrir et al., 2020). The color

of the lesion is still a crucial criterion for diagnosing

melanomas. It represents a variety of colors. More-

over, dermatologists have proposed other important

rules for diagnosing melanomas like the blue-black

rule (Argenziano et al., 2011) and the ugly duck-

ling (Grob and Bonerandi, 1998). The blue-black

rule is deﬁned as the presence of the blue and black

color in a lesion surface (Daghrir et al., 2020). Thus,

extracting color features is extremely important in

melanoma inspection. For instance, some systems

represent the color by deﬁning the Color Name (CN)

features which are linguistic color labels representing

different colors in a skin lesion. As discussed above,

we can notice that the color is with a high sensitivity

in the whole melanoma CAD system. In this work, we

suggest a new color feature extraction method.First

we determine the different color names of skin le-

sions. Second,we search is a selection step the most

pertinent color names that ensure accurate melanoma

detection. The overall implementation of the extrac-

tion process is given in (see Fig. 2). To demonstrate

the effectiveness of our proposed method, we test the

selected features on a set of skin lesion images (see

Fig. 3).

3.1 Supervised Color Name Extraction

Skin lesions may contain many different colors re-

garding the healthy skin color, the kind of the skin dis-

ease, the stage of the disease, etc. Thus, it is important

to choose which colors are more likely to be present

in a skin lesion. Melanomas may mainly include six

colors. These colors can be presented with different

degrees. For an effective color name representation,

we extract the dominant six colors from a set of im-

ages using the K-means algorithm. Centroids for the

algorithm initially depend on the six considered col-

ors: light brown, dark brown, black, red, white, and

A Supervised Quantiﬁcation of the Color Names Characterizing the Visual Component Color in the ABCD Dermatological Criteria for a

Further Melanoma Inspection

149

Figure 2: Layout architecture of method of extracting color-names.

Figure 3: Layout architecture of evaluating extracted color names.

slate blue. K-means iteratively minimizes the intra-

class and maximizes the inter-class distances to create

a ﬁnal partition of image pixels into six groups. Each

input pixel is characterized by three intensities in the

RGB color space. When convergence is reached, six

centroids will be assigned for each color. For more

details about K-means clustering, please see (Melbin

and Raj, 2021b). Gathering all the extracted colors

will serve then to do clustering one more time to ﬁ-

nally extract the 24 major colors, which are displayed

in the pie chart shown in Fig. 4. We assume that for

each color four instances are adopted. After identify-

ing, the major 24 colors associated with skin lesions,

the process of preserving only those which guaran-

tee an effective representation of lesion color names

is done using feature selection, which is a trend in

a lot of machine learning systems. We use the in-

ﬁnite Feature Selection (inf-FS) (Roffo et al., 2015)

which is a feature selection method that performs a

ranking step in an unsupervised way, followed by a

cross-validation strategy to select only the most repre-

Figure 4: The most dominant 24 colors extracted from a set

of images.

sentative features. Ranking individually the relevancy

of features is done utilizing class labels: malignant or

benign. On the other hand, using a distance matrix,

we deﬁne how the 24 various color values frequently

occur in the lesion (see Fig 3). In the processed im-

age, every pixel value is assigned to the nearest color

ICT4AWE 2022 - 8th International Conference on Information and Communication Technologies for Ageing Well and e-Health

150

name regarding the distance of its intensity. Finally,

by counting the pixels of each color name and apply-

ing the inf-FS procedure, the suggested process gen-

erates a ranked features that refer to their character-

ization relevance. The six major colors found after

applying the feature selection, afﬁrm the efﬁciency of

the black-blue clue proposed by physicians in inspect-

ing the malignancy of skin lesions. The slight blue

and black colors are highly ranked (see ﬁgure3).

3.2 Classifying Skin Lesion using Color

Name Features

After extracting the most relevant color names, the

skin lesion classiﬁcation is implemented. The most

attractive feature of this process is to evaluate the im-

portance of using only the best-ranked color names.

Generally, using them would be more accurate in clas-

sifying skin lesions. However, counting the pixels for

the best color names using the distance matrix will

create an unfair distribution since some pixels have

different yet distant colors regarding the best color

names. These pixels will be assigned anyway with

the nearest color name, so it is crucial to disturb the

pixels partitioning by using alternative color names.

It would be more appropriate to use the six worst

color names for that purpose. For the classiﬁcation

task, we have use the K-Nearest Neighbor (KNN) al-

gorithm (Coomans and Massart, 1982). This method

demonstrates a high classiﬁcation accuracy especially

when using a low-dimensional representation.

4 EXPERIMENTS AND RESULTS

The experiments are conducted using the ISIC2017

(International Skin Imaging Collaboration)

dataset (Codella et al., 2018). The images are

split into three sets, the training set which is used for

ranking the color names and for training the KNN,

and the validation and testing set which are used to

evaluate the performance of our color name extrac-

tion method. We have also used the SD-198 (Sun

et al., 2016), which contains 198 skin diseases rep-

resented by 6,584 images. The authors provide two

split strategies. We have used the ﬁfty split, which

contains 3,292 training and 3,292 testing images. The

SD-198 dataset was trained in (Yang et al., 2018) by

using different low-level features describing the three

visual components: texture, color, and border. We

compare our proposed color name extraction method

with the other color features used in the literature.

To demonstrate the performance of any classiﬁcation

task, the accuracy is basically introduced, which is

the fraction of correctly classiﬁed points and the total

number of points.

Accuracy =

T P + T N

T P + T N + FP + FN

(1)

where TP = True positive, FP = False positive, TN =

True negative and FN = False negative.

However, in some cases, accuracy does not really

indicate the relevancy of the system. For example

when evaluating a binary classiﬁer, one class having

positive labels can appear in the validation set more

than the other one having negative labels. Thus, to

overcome this class imbalance problem, sensitivity,

speciﬁcity and balanced accuracy are deﬁned to ef-

ﬁciently evaluate the classiﬁcation task. The sensitiv-

ity known also as recall, measures the proportion of

actual positives that are correctly identiﬁed.

Sensitivity =

T P

T P + FN

(2)

Nevertheless, speciﬁcity measures the proportion of

actual negatives that are correctly classiﬁed.

Speci f icity =

T N

T N + FP

(3)

As a consequence, balanced accuracy, which is the

arithmetic mean of both sensitivity and speciﬁcity

metrics will properly introduce the pertinence of ma-

lignancy inspection.

Balanced accuracy =

sensitivity + speci f icity

(4)

In the experiments, we ﬁrst ﬁnd the best color names

introducing the best representative features and we

use the validation data set to gain insights into what

the optimal hyperparameter K is. Therefore, we sug-

gest three scenarios to demonstrate the performance

of the extracted color names using the validation data.

The ﬁrst, called Scenario A, illustrates the use of the

six major and six worst color names. The second

named Scenario B, is deﬁned as the use of only the

best color names. On the other hand, we use all the

24 color names for scenario C.

Figure 5: Balanced accuracy using different scenarios with

various values of K on validation set.

A Supervised Quantiﬁcation of the Color Names Characterizing the Visual Component Color in the ABCD Dermatological Criteria for a

Further Melanoma Inspection

151

Table 1: Balanced accuracy using different scenarios applied on testing set.

Scenario A Scenario B Scenario C

Accuracy(%) 77.7 77 62.7

Sensitivity(%) 17.4 11.6 26.7

Speciﬁcity(%) 87.7 88.1 68.8

Balanced accuracy(%) 52.6 49.8 47.7

Generally, when the number of classes is odd, hy-

perparameter K must be even. Thus, by generating

the model on different even values of K and checking

their performance, we will understand which number

of neighbors should be considered in the testing pro-

cess. In Fig. 5, we can say that the classiﬁer does a

little better when using a small number of K with a

small number of color names. However, for the third

scenario, when the number of Ks is higher, balanced

accuracy is higher than when applying the other sce-

narios. The experiments show also that when using

the proposed color name combination the balanced

accuracy is higher than using the other scenarios.

Now, we have test the performance of the KNN

using the three scenarios as well, with K=3 (Table 1).

Based on balanced accuracy, it is obvious that sce-

nario A achieve the best performance. Finally, the

disturbing strategy reports the best classiﬁcation per-

formances so far referring to the other scenarios. It

is evident that adding further color names is generally

useful. With the three scenarios, balanced accuracy

is somewhat minimal, probably because we train the

classiﬁer with an unbalanced dataset having a limited

number of features.

We compare the extracted color names with other

several hand-crafted color features used in the state of

the art. Table 2 reports the prediction performance of

using various representations using KNN with k=3.

As we have a limited and unbalanced dataset, we use

the 5-Fold Cross Validation strategy to report how

well the representations are working. Thus, ﬁve accu-

racies as well as their mean using the colorSIFT, CH

(Color Histogram), and our proposed method (SCN)

are reported. It is obvious that all the color representa-

tions succeeded in accurately diagnosing melanoma.

Nevertheless, the best representation is the proposed

one, which reaches about 75% accuracy with only 12

features, unlike the other used representations.

Convolutional Neural Networks (CNNs) have

been widely used in the literature to automati-

cally characterize skin lesions, for their notable

performance. Therefore, in a previous research

work(Daghrir et al., 2020), we have used a CNN fol-

lowed by a fully-connected layer with a softmax ac-

tivation function to classify malignant and benign le-

sions on the ISIC2017 dataset. The extracted features

achieve an 85.5% accuracy, about 8% greater com-

pared to the achievement of our proposed features

(77.7% accuracy in Table 1). This result shows that

our proposed method can reﬂect the comprehensive

medical information relating to the black-blue feature.

As a result, when using other different features de-

scribing also the color, texture, and shape of the le-

sions, our suggested color name features will be more

relevant.

We compare also our proposed color name extrac-

tion using the SD-198 dataset. Thus, the whole pro-

cess was tested using 198 skin diseases. A set of the

most present 24 colors and their ranks in identifying

the type of the disease are extracted. In table 3, we

report the accuracy of using different color features

proposed and used in (Yang et al., 2018), as well as

the color-based features extracted using our proposed

method. It is shown that our method does not perform

well, it achieves only 5.58% accuracy. It fails in rec-

ognizing the different skin diseases, mainly because

of the variability and the speciﬁcity of the skin dis-

eases. In (Yang et al., 2018), the authors also have

compared the classiﬁcation performance using deep

features derived from ﬁne-tuned CNNs. Mainly, a

CNN achieves an accuracy of 53.35% in classifying

the 198 skin diseases, which is incomparable with the

use of our proposed color-based features.

Although, the use of more than 12 color names

(all the 24 extracted color names) slightly improves

the accuracy to 5.99% using KNN. This can some-

how prove the efﬁciency of our proposed method in

precisely identifying melanoma since it manifests the

presence of a very limited number of colors( gener-

ally 6 colors). However, 198 skin diseases might be

characterized by a huge number of color names.

Thus, the huge number of skin diseases in a

dataset, limits the performance of our method, as it is

shown in Table 3, 5.58% against 52.6% accuracy for

recognizing 198 skin diseases, compared to the use of

the ISIC2017.

5 CONCLUSION

In this paper, we have presented a new method of

color name extraction in a supervised way using fea-

ture selection to rank the extracted color names. The

application of the extracted color names has been

ICT4AWE 2022 - 8th International Conference on Information and Communication Technologies for Ageing Well and e-Health

152

Table 2: Accuracy using different validation Folds with different color representations on the ISIC2017 dataset.

Accuracy Fold1 Fold2 Fold3 Fold4 Fold5 Mean Dimension

SCN 0.78 0.76 0.77 0.78 0.67 0.75 12

CH 0.71 0.78 0.75 0.77 0.63 0.72 255

colorSIFT 0.69 0.71 0.75 0.71 0.62 0.69 10000

Table 3: Accuracy using different representations and clas-

siﬁers on the SD-198 dataset.

Accuracy

Features Dimension KNN SVM

CH 256 12.33 4.19

CN 21000 20.03 20.23

colorSIFT 21000 21.29 22.51

CN-L 21000 42.50 38.91

CCV-L 21000 42.80 40.13

SCN 12 5.58 4.73

proved utilizing a classiﬁer with three different par-

titions. These color names are mainly extracted to

classify skin lesions for more accurate inspection of

melanomas, which are considered as the most fa-

tal skin cancer. The proposed method has shown a

notable performance for diagnosing melanomas. A

comparison of different handcrafted features is pre-

sented as well, which proves the efﬁciency of our

color name features against the state-of-the-art color

representations. Accordingly, using only our pro-

posed color-based features shows a promising result

compared to automatically extracted features using

deep learning. However, our proposed representation

method shows a limitation when using a benchmark

dataset that contains several skin conditions. Thus,

these color names can be further used with other

hand-crafted features and more sophisticated machine

learning models to inspect melanomas to ameliorate

the diagnosis process. Fuzzy features of color names

could be also introduced in future work.

REFERENCES

Abbasi, N. R., Shaw, H. M., Rigel, D. S., Friedman, R. J.,

McCarthy, W. H., Osman, I., Kopf, A. W., and Polsky,

D. (2004). Early diagnosis of cutaneous melanoma:

revisiting the abcd criteria. Jama, 292(22):2771–

2776.

Abdel-Hakim, A. E. and Farag, A. A. (2006). Csift: A sift

descriptor with color invariant characteristics. In 2006

IEEE computer society conference on computer vision

and pattern recognition (CVPR’06), volume 2, pages

1978–1983. Ieee.

Argenziano, G., Longo, C., Cameron, A., Cavicchini, S.,

Gourhant, J.-Y., Lallas, A., McColl, I., Rosendahl, C.,

Thomas, L., Tiodorovic-Zivkovic, D., et al. (2011).

Blue-black rule: a simple dermoscopic clue to recog-

nize pigmented nodular melanoma. British Journal of

Dermatology, 165(6):1251–1255.

Claridge, E., Cotton, S., Hall, P., and Moncrieff, M. (2003).

From colour to tissue histology: physics-based inter-

pretation of images of pigmented skin lesions. Medi-

cal Image Analysis, 7(4):489–502.

Codella, N. C., Gutman, D., Celebi, M. E., Helba, B.,

Marchetti, M. A., Dusza, S. W., Kalloo, A., Liopy-

ris, K., Mishra, N., Kittler, H., et al. (2018). Skin

lesion analysis toward melanoma detection: A chal-

lenge at the 2017 international symposium on biomed-

ical imaging (isbi), hosted by the international skin

imaging collaboration (isic). In 2018 IEEE 15th in-

ternational symposium on biomedical imaging (ISBI

2018), pages 168–172. IEEE.

Coomans, D. and Massart, D. L. (1982). Alternative k-

nearest neighbour rules in supervised pattern recogni-

tion: Part 1. k-nearest neighbour classiﬁcation by us-

ing alternative voting rules. Analytica Chimica Acta,

136:15–27.

Daghrir, J., Tlig, L., Bouchouicha, M., and Sayadi, M.

(2020). Melanoma skin cancer detection using deep

learning and classical machine learning techniques: A

hybrid approach. In 2020 5th International Confer-

ence on Advanced Technologies for Signal and Image

Processing (ATSIP), pages 1–5. IEEE.

Dimitriou, F., Krattinger, R., Ramelyte, E., Barysch, M. J.,

Micaletto, S., Dummer, R., and Goldinger, S. M.

(2018). The world of melanoma: epidemiologic, ge-

netic, and anatomic differences of melanoma across

the globe. Current oncology reports, 20(11):1–9.

Gonzalez-Diaz, I. (2018). Dermaknet: Incorporating the

knowledge of dermatologists to convolutional neural

networks for skin lesion diagnosis. IEEE journal of

biomedical and health informatics, 23(2):547–559.

Grob, J. and Bonerandi, J. (1998). The ‘ugly duckling’sign:

identiﬁcation of the common characteristics of nevi

in an individual as a basis for melanoma screening.

Archives of dermatology, 134(1):103–104.

Heinzerling, L. and Eigentler, T. K. (2021). Skin cancer

in childhood and adolescents: Treatment and implica-

tions for the long-term follow-up. In Late Treatment

Effects and Cancer Survivor Care in the Young, pages

349–355. Springer.

Khazaei, Z., Ghorat, F., Jarrahi, A., Adineh, H., Sohrabi-

vafa, M., and Goodarzi, E. (2019). Global incidence

and mortality of skin cancer by histological subtype

and its relationship with the human development in-

dex (hdi); an ecology study in 2018. World Cancer

Res J, 6(2):e13.

Magalhaes, C., Tavares, J. M. R., Mendes, J., and Vardasca,

R. (2021). Comparison of machine learning strategies

for infrared thermography of skin cancer. Biomedical

Signal Processing and Control, 69:102872.

A Supervised Quantiﬁcation of the Color Names Characterizing the Visual Component Color in the ABCD Dermatological Criteria for a

Further Melanoma Inspection

153

Maglogiannis, I. and Doukas, C. N. (2009a). Overview

of advanced computer vision systems for skin lesions

characterization. IEEE transactions on information

technology in biomedicine, 13(5):721–733.

Maglogiannis, I. and Doukas, C. N. (2009b). Overview

of advanced computer vision systems for skin lesions

characterization. IEEE transactions on information

technology in biomedicine, 13(5):721–733.

Maglogiannis, I. G. and Zaﬁropoulos, E. P. (2004a). Char-

acterization of digital medical images utilizing sup-

port vector machines. BMC Medical Informatics and

Decision Making, 4(1):1–9.

Maglogiannis, I. G. and Zaﬁropoulos, E. P. (2004b). Char-

acterization of digital medical images utilizing sup-

port vector machines. BMC Medical Informatics and

Decision Making, 4(1):1–9.

Manousaki, A. G., Manios, A. G., Tsompanaki, E. I., and

Tosca, A. D. (2006). Use of color texture in determin-

ing the nature of melanocytic skin lesions—a qualita-

tive and quantitative approach. Computers in biology

and medicine, 36(4):419–427.

Melbin, K. and Raj, Y. J. V. (2021a). Integration of modi-

ﬁed abcd features and support vector machine for skin

lesion types classiﬁcation. Multimedia Tools and Ap-

plications, 80(6):8909–8929.

Melbin, K. and Raj, Y. J. V. (2021b). Integration of modi-

ﬁed abcd features and support vector machine for skin

lesion types classiﬁcation. Multimedia Tools and Ap-

plications, 80(6):8909–8929.

Menzies, S. W., Bischof, L., Talbot, H., Gutenev, A.,

Avramidis, M., Wong, L., Lo, S. K., Mackellar, G.,

Skladnev, V., McCarthy, W., et al. (2005). The per-

formance of solarscan: an automated dermoscopy im-

age analysis instrument for the diagnosis of primary

melanoma. Archives of dermatology, 141(11):1388–

1396.

Roffo, G., Melzi, S., and Cristani, M. (2015). Inﬁnite fea-

ture selection. In Proceedings of the IEEE Interna-

tional Conference on Computer Vision, pages 4202–

4210.

Saeed, J. and Zeebaree, S. (2021). Skin lesion classiﬁcation

based on deep convolutional neural networks archi-

tectures. Journal of Applied Science and Technology

Trends, 2(01):41–51.

Spalding, J. (1993). The doctor with an inherited defect

of colour vision: effect on clinical skills. The British

Journal of General Practice, 43(366):32.

Stolz, W., H

olzel, D., Riemann, A., Abmayr, W., Prze-

tak, C., Bilek, P., Landthaler, M., and Braun-Falco, O.

(1991). Multivariate analysis of criteria given by der-

matoscopy for the recognition of melanocytic lesions.

In Book of Abstracts, Fiftieth Meeting of the American

Academy of Dermatology, Dallas, Tex: Dec, pages 7–

12.

Sun, X., Yang, J., Sun, M., and Wang, K. (2016). A

benchmark for automatic visual classiﬁcation of clin-

ical skin disease images. In European Conference on

Computer Vision, pages 206–222. Springer.

Yang, J., Sun, X., Liang, J., and Rosin, P. L. (2018). Clinical

skin lesion diagnosis using representations inspired by

dermatologist criteria. In Proceedings of the IEEE

Conference on Computer Vision and Pattern Recog-

nition, pages 1258–1266.

Zhang, Z., Moss, R. H., and Stoecker, W. V. (2003). Neu-

ral networks skin tumor diagnostic system. In Inter-

national Conference on Neural Networks and Signal

Processing, 2003. Proceedings of the 2003, volume 1,

pages 191–192. IEEE.

ICT4AWE 2022 - 8th International Conference on Information and Communication Technologies for Ageing Well and e-Health

154