Skin Cancer Classification Using Image Processing

and Machine Learning

Arslan Javaid, Muhammad Sadiq Faraz Akram
Faculty of Engineering and Applied Sciences Faculty of Engineering and Applied Sciences Faculty of Engineering and Applied Sciences
Riphah International University Riphah International University Riphah International University
Islamabad, Pakistan Islamabad, Pakistan Islamabad, Pakistan
2021 International Bhurban Conference on Applied Sciences and Technologies (IBCAST) | 978-1-6654-0516-4/21/$31.00 ©2021 IEEE | DOI: 10.1109/IBCAST51254.2021.9393198

Abstract—one of the most rapidly spreading cancers among science behind skin cancer reveals that melanin is present in
various other types of cancers known to humans is skin cancer. human skin andmelanocytes are cells in the skin layer that
Melanoma is the worst and the most dangerous type of skin produce melanin. The amount and kinds of melanin produced
cancer that appears usually on the skin surface and then extends by melanocytes of different human bodies vary from person to
deeper into the layers of skin. However, if diagnosed at an early person.In addition to coloring our skin, it also protects from
stage; the survival rate of Melanoma patients is 96% with simple ultraviolet rays of the sun.
and economical treatments. The conventional method of
diagnosing Melanoma involves expert dermatologists, equipment, The factors contributing to skin cancer include prolonged
and Biopsies. To avoid the expensive diagnosis, and to assist exposure to direct sunlightultraviolet (UV) rays, the presence
dermatologists, the field of machine learning has proven to of many or unusual moles, skin types, and also if there is a
provide state of the art solutions for skin cancer detection at an history of melanoma that runs in family. The mortality rate due
earlier stage with high accuracy. In this paper, a method for skin to melanoma is usually very high butif diagnosed at an early
lesion classification and segmentation as benign or malignant is stage has a 99% probability of survival[1], [2]. In many cases,
proposed using image processing and machine learning. A novel it is a difficult task even for expert dermatologists to make
method of contrast stretching of dermoscopic images based on decisions whether a lesion is benign or malignant because of
the methods of mean values and standard deviation of pixels is the high resemblance of malignant with benign. Dermatologists
proposed. Then the OTSU thresholding algorithm is applied for use a few techniques such as ABCD rule (Atypical, Border,
image segmentation. After the segmentation, features including
Color, and Diameter) to get better classification accuracy, but
Gray level Co-occurrence Matrix (GLCM) features for texture
identification, the histogram of oriented gradients (HOG) object,
still, human expertise is required[3].
and color identification features are extracted from the Frequent use of biopsies is also not encouraged by
segmented images. Principal component analysis (PCA) dermatologists. According to International Skin Imaging
reduction of HOG features is performed for dimensionality Collaboration, the number of unnecessary culture tests which
reduction. Synthetic minority oversampling technique (SMOTE) are being performed vastly varies depending upon various
sampling is performed to deal with the class imbalance problem. parameters which include clinical setup,expertise of
The feature vector is then standardized and scaled. A novel
dermatologist, and the technology applied.For illustration,
approach of feature selection based on the wrapper method is
consider the cases of youngsters in which melanoma rates are
proposed before classification. Classifiers including Quadratic
Discriminant, SVM (Medium Gaussian), and Random Forest are
significantly low, 500000 culture tests a year have been
used for classification. The proposed approach is verified on the performed to analyze roughly 400 melanomas[4].
publicly accessible dataset of ISIC-ISBI 2016. Maximum Computer procedures and advancements in machine
accuracy is achieved using the Random Forest classifier. The learning not only aid the dermatologists in early detection of
classification accuracy of the proposed system with the Random melanoma but also avoid heavy expenses of melanoma
Forest classifier on ISIC-ISBI 2016 is 93.89%.The proposed detection and unnecessary biopsies. Novel automatic
approach of contrast stretching before the segmentation gives
melanoma detection systems save a lot of time, money, and
satisfactory results of segmentation. Further, the proposed
wrapper-based approach of feature selection in combination with
effort. Machine learning has proven to provide melanoma
the Random Forest classifier gives promising results as compared classification with improved and higher accuracies.
to other commonly used classifiers.
II. RELATED WORKS
Keywords—Skin lesion segmentation, contrast stretching,
features extraction, features reduction, features normalization,
Although the advancement in the dermatological equipment
features scaling, wrapper method, SMOTE sampling, skin cancer has increased the classification accuracy of melanoma, the
classification, random forest classifier. technological developments and improvements in the area of
machine learning and image processing have resulted in a
medical breakthrough in diagnosis, detection, and classification
I. INTRODUCTION of melanoma with much more accuracy and reliability.
With rapidly increasing global air pollution and damage to
the ozone layer an alarming number of the human population is The literature review reveals that different practices have
diagnosed to develop skin cancer as compared to any other been used to develop computer-aided automatic diagnostics
type of cancer combined. Melanoma has a very high death ratio systems for the classification of skin cancer which take
as compared to other types of skin cancer. The study of the

‹,(((

3URFHHGLQJVRIWK,QWHUQDWLRQDO%KXUEDQ&RQIHUHQFHRQ$SSOLHG6FLHQFHV 7HFKQRORJ\ ,%&$67 

,VODPDEDG3DNLVWDQ±-DQXDU\
Authorized licensed use limited to: Miami University Libraries. Downloaded on June 14,2021 at 20:07:58 UTC from IEEE Xplore. Restrictions apply.
dermoscopic images as input and give classification results as Dalila et al. [9] introduced a system for the Segmentation
benign or malignant at the output. and classification of melanoma and benign skin lesions based
on Ant-Colony based segmentation algorithm for the lesion.
HiamAlquran et al. [5] proposed a skin cancer detection They extracted features of shape, texture, and color. For
system based on OTSU Thresholding. After Image pre- classification, they used KNN and ANN classifiers and
processing and segmentation, they extracted features including achieved accuracies of classification of 85.22% and 93.60%
ABCD (Atypical, Border, Color, and Diameter), Total respectively.
dermoscopic score (TDS), Circulation, and texture features
using Gray Level Co-occurrence Matrix (GLCM). They used Sumithra R et al. [10]developed a skin cancer diagnosis
the SVM classifier with the RBF Kernel algorithm for system based on the region-growing segmentation techniques.
classification and achieved an accuracy of 92.1%. After Pre-Processing and segmentation of images, they
extracted color, texture, and RGB Histogram Features. In the
UzmaBano et al. [6] developed a system for detecting skin final phase of classification, they used three classifiers i.e.
cancer by image processing. After pre-processing, the resultant KNN, SVM, and combined SVM+KNN, and achieved
was subjected to maximum entropy thresholding for lesion classification accuracies of 86%, 87.5%, and 94% respectively.
segmentation. They extracted a few texture features using
GLCM. They used the SVM classifier and achieved a Most of the researchers have used their own data collected
classification accuracy of 95%. from hospitals. But their dataset is too limited for proper
training during the machine learning phase. Few researchers
M. Attique Khan et al. [7]developed an automatic skin have used the ISIC dataset which is a handsome amount of data
cancer detection and classification system based on the normal and also it is more practical.
distribution for image segmentation after some pre-processing.
They extracted shape features using a Histogram of Gradients The purpose of this research is to find improved and more
(HOG), texture features using Harlick Features, and color effective ways to detect skin cancer using digital image
features. For dimensionality reduction, they performed an processing and machine learning techniques. The final
entropy controlled feature selection. They utilized several objective is to assist the doctors in the detection of skin cancer
classifiers to check classification accuracy and achieved at an early stage by providing improved and reliable results.
maximum accuracy using a multi-class SVM classifier. They
The suggested system gives high accuracy of classification
achieved a classification accuracy of 97.5%, 97.75%, and
of the lesion as benign or malignant which will be very helpful
93.2% on PH2, ISIC (UDA, MSK-2), and Combined (ISBI
for the identification of patients efficiently and our results are
2016 -17) datasets respectively.
better than earlier work on the ISIC dataset.
Vijayalakshmi M M [8] developed an automatic skin
cancer detection system using image processing and artificial III. METHODOLOGY
intelligence. She utilized 1000-1500 images from publicly
available ISIC dataset. After initial Pre-processing, lesion In this paper, we propose a system that classifies
segmentation was performed using the OTSU segmentation dermoscopic images as benign or malignant by using image
method, Modified Otsu segmentation method, and watershed processing and machine learning.Fig.1 shows the block
segmentation method. She used multiple classifiers to evaluate diagram of the proposed system.
the performance including the Back Propagation Algorithm
(Neural Networks), SVM, and CNN. She achieved maximum
accuracy of 85% with the SVM classifier.

Fig. 1. Block Diagram of system suggested for skin lesion segmentation & classification

3URFHHGLQJVRIWK,QWHUQDWLRQDO%KXUEDQ&RQIHUHQFHRQ$SSOLHG6FLHQFHV 7HFKQRORJ\ ,%&$67  

,VODPDEDG3DNLVWDQ±-DQXDU\
Authorized licensed use limited to: Miami University Libraries. Downloaded on June 14,2021 at 20:07:58 UTC from IEEE Xplore. Restrictions apply.
 grayscale Pre-Processed images. OTSU thresholding aims to
A. Image Database find the threshold and classify pixels into two classes so that
The database is downloaded from publicly available images within-class variance is minimum (between-class variance is
on the ISIC website. The database included the ISBI-2016 maximum).
challenge which has RGB dermoscopic images along with their
labels and segmentation ground truths. Background triangles that exist in a few images on four
sides are then removed by creating a mask and adding it to the
binary image. Later the image is inverted to make the skin
B. Pre-Processing
lesion white and the background to be black. Flood fill
As ISIC-ISBI dataset images have different artifacts, so operation on 4-connected pixels in the background is
pre-processing is done on them to make them more performed to remove holes. A morphological opening is then
meaningful. Pre-processing included image resizing, noise done to remove small objects having pixels fewer than 2000.
removal, contrast stretching, RGB to Gray conversion, and hair The lesion border is smoothened using opening operation first
removal. We resized all images to 767x1022. For noise with a disk-shaped structuring element of radius 20 followed
removal median filter of 3-by-3-by-3 is used. For contrast by a closing operation with the same structuring element.
enhancement, a new method is proposed. Contrast
enhancement significantly improves the results of D. Features extraction and reduction
segmentation in the next phase. Proposed contrast stretching is
based on the mean and standard deviation of pixels intensities After the segmentation, features of texture, shape, and color
of images. Minimum and maximum intensity values i.e. "Low are extracted from segmented skin lesions. To extract the
in" and "High in" of input images values of input images are texture features, segmented lesions mapped on original images
calculated using formulas defined as in (1) and (2). are converted from RGB to grayscale first. Then two-
dimensional wavelet decomposition of level 3 with mother
Low in = Avg ı*N(1) wavelet as Daubechies (db4) is performed on grayscale images.
Low pass approximation coefficients matrix (CA) of level 3
High in=Avg + ı*N(2)
wavelet decomposition result in new images of reduced size.
Where, Afterward, texture features are obtained via gray level co-
occurrence matrix of approximation coefficients matrix. A total
Avg = Average of 13 texture features are extracted as shown in Table 1.
ı 6WDQGDUG'HYLDWLRQ A total of 36 color features are extracted using three
N = 0.4 different color spaces i.e. RGB, HSV, and LAB color space as
shown in Table 1. For extracting shape features, Histograms of
Intensity values are then mapped in the output image from Gradient features (HOG) are extracted. The size of all
0 to 255. Contrast stretching is performed on R, G, and B segmented lesion images is reduced first to 96 X 128 to reduce
channels separately which are later concatenated to contrast the dimensionality of images. Then Histogram of Gradients
stretched RGB images. RGB to gray conversion is performed "HOG" feature vector is determined. The size of the "HOG"
next. For removing hairs bottom-hat filtering is performed, feature vector for a single image is 1x3456 because each image
then those pixels which have been filtered are replaced by is of dimension 96 X 128 and the block size is chosen to be [8
neighboring pixels. 8]. For reducing the dimensionality of a large data array of
HOG vector, PCA is used. The size of the HOG feature vector
C. Segmentation using PCA is reduced from 1x3456 to 1x100 for a single
After the Pre-Processing, image segmentation is performed image. PCA increases understandability and reduces
to extract the region of interest i.e. skin lesion. We used the information loss.
OTSU thresholding algorithm for the segmentation of

TABLE 1. EXTRACTED TEXTURE AND COLOR FEATURES

Texture Skewness Mean Contrast Energy Homogeneity Standard

features Root Mean Square Variance Smoothness Kurtosis Correlation Entropy

Inverse Difference Movement
Mean R Mean G Mean Blue Variance R Variance G Variance Blue

Color Kurtosis R Kurtosis G Kurtosis Skewness R Skewness G Skewness Blue

Features Mean hue Mean Saturation Mean Variance Hue Variance Variance Value
Kurtosis H Kurtosis S Kurtosis V Skewness H Skewness S Skewness V

Mean L Mean A Mean B Variance L Variance A Variance B

3URFHHGLQJVRIWK,QWHUQDWLRQDO%KXUEDQ&RQIHUHQFHRQ$SSOLHG6FLHQFHV 7HFKQRORJ\ ,%&$67  

,VODPDEDG3DNLVWDQ±-DQXDU\
Authorized licensed use limited to: Miami University Libraries. Downloaded on June 14,2021 at 20:07:58 UTC from IEEE Xplore. Restrictions apply.
 Kurtosis L Kurtosis A Kurtosis B Skewness L Skewness A Skewness B
13 GLCM texture features, 36 color features, and 100 classification accuracy an objective function. "LA"
shape features are all then concatenated to make a single [15]algorithm determined 95 features and gave the most
feature vector of dimension 1x149 for a single image. promising results in terms of sensitivity, specificity, precision,
FPR, and accuracy as compared to other algorithms.
E. SMOTE Sampling
Data of ISIC- ISBI has a class imbalance problem. Almost H. Classification
80% of data is "benign" and 20% of data is "malignant". So In the final step, we classified data as benign or malignant.
any classifier trained on this data will be more biased or Numbers of classifiers including Quadratic Discriminant, SVM
skewed to have more training on benign data. As a result, it (Medium Gaussian), and Random Forest are then used for
would be poor training, poor training accuracy, and poor training and their performance is evaluated on test data.
classification accuracy on test data. Hence data needs to be Classifier models are later evaluated on test data. SVM divides
balanced first. For this purpose Synthetic Minority Over the data into distinguishable classes to determine hyperplane
Sampling (SMOTE) technique is implemented on the final with maximum margin. Proposed features selection based on
feature vector [11]. SMOTE technique resulted in data being wrapper methods gives remarkable and promising results when
50% benign and 50% malignant. Random Forest classification is performed. 500 decision trees
are used in the Random Forest classifier and classification is
All of the training and test data of the ISIC-ISBI data set is done based on the majority votes of decisions given by
used to extract the feature vector. Next, the SMOTE technique different trees. The Random Forest technique generates
is used for data balancing. After that feature vector or decision trees on samples of data and predictions from each
predictors which are a total of 149 predictors of all images tree undergo the process of voting for classification. More
along with their given labels are randomized. Next data is votes from different trees mean more likely it belongs to that
broken down in such a way that 80% is utilized for purpose of particular class.
training and 20% is utilized for testing. Now the ratio of benign
and malignant images in training data is 50:50. Similarly, the
ratio of benign and malignant images in test data is 50:50. IV. RESULTS
Fig.2 shows the results of contrast stretching. Results of
F. Features Standardization segmentation compared with ground truths are shown in Fig.3.
Features vector is standardized using the method of "Linear Classification accuracies achieved on the ISIC-ISBI2016
scaling to unit variance" [12]. Average and standard deviation dataset using SVM (Medium Gaussian), Quadratic
of each column of the feature vector is computed. Then the Discriminant, and Random Forest classifiers are 85.50%,
average of each column is subtracted from the original feature 88.17%, 90.84%, and 93.89% respectively. The confusion
vector's respective column to get zero mean feature vector. matrix using Random Forest classifier for ISIC-ISBI 2016 is
Afterward zero mean feature vector's each column is divided shown in Fig.4. Among 114 malignant guesses, 94.7% are
by the standard deviation of that column to get the feature accurate and 5.3% are erroneous. Among 148 benign guesses,
vector of unit variance. 93.2% are accurate and 6.8% are erroneous. Among 144
benign subjects, 95.8% are properly foreseen as benign and
G. Features Scaling and selection 4.2% are foreseen as malignant. From 118 malignant subjects,
Features are scaled in such a way that in the feature vector 91.5% are properly categorized as malignant and 8.5% are
each feature is normalized in a range from 0 to 7 by using the categorized as benign. 93.9% of the complete guesses are
Min-Max Mapping Algorithm [13]. correct and 6.1% are erroneous.
In this paper, a novel approach of using wrapper methods
as feature selection methods for skin cancer classification is
proposed. The wrapper method aims at finding the subset of
features that are most strongly relevant and also weakly
relevant features that improve performance, using the training
algorithm itself as a part of the evaluation function. For the
problem of finding a subset of features that are most relevant as
well as to improve accuracy, the wrapper method was proposed
by George H. John et al [14]. Feature selection is used to
improve the performance of the machine learning models,
increase the accuracy of the classification, and to aid the
analysis of the results. Normalized features are used to Fig. 2. Proposed contrast stretching results
determine the most prominent features out of 149 features in
hand." FEATURESELECT" is free open source software under
MIT license [15]. A total of 95 most optimum features are
selected by using optimization algorithms as a feature selection
method and decision tree as a base classifier. We considered

3URFHHGLQJVRIWK,QWHUQDWLRQDO%KXUEDQ&RQIHUHQFHRQ$SSOLHG6FLHQFHV 7HFKQRORJ\ ,%&$67  

,VODPDEDG3DNLVWDQ±-DQXDU\
Authorized licensed use limited to: Miami University Libraries. Downloaded on June 14,2021 at 20:07:58 UTC from IEEE Xplore. Restrictions apply.
 V. CONCLUSION
In this work, a novel method of skin cancer classification
using machine learning and image processing is implemented.
In the first step, a novel method of contrast stretching based
on the mean and standard deviation of pixels for dermoscopic
images enhancement is proposed. Then OTSU thresholding is
performed for segmentation.
Shape, color, and texture features are extracted in the
second step, and shape features are then reduced using the
PCA. The class imbalance problem of the ISIC dataset is
overcome using the SMOTE sampling technique.
In the third step, features are standardized and scaled and
then a novel approach of features selection based on wrapper
methods for selecting the most optimum features is proposed.
The recommended system is tested on an openly available
dataset i.e. ISIC-ISBI 2016 and it is concluded that the
proposed wrapper method for feature selection in combination
with the Random Forest classifier gives promising results as
compared with other classifiers.

TABLE 2 CLASSIFICATION RESULTS USING DIFFERENT

CLASSIFIERS

Classification
Accuracy (%)
Algorithm
SVM 88.17
Quadratic
Fig.
Fig 3.
3 Comparison of segmentation results to ground truth a) 90.84
Discriminant
Actual image b) Segmented skin lesion c) Segmented lesion masked
on actual image d) Ground Truth Random Forest 93.89

Classification accuracies using the SVM, Quadratic

Discriminant, and Random Forest are compared in Table 2.
REFERENCES

[1] D. Schadendorf et al., “Melanoma,” Lancet, vol. 392, no. 10151, pp.
971–984, 2018.
[2] S. Gupta and H. Tsao, “Epidemiology of melanoma,” Pathol.
Epidemiol. Cancer, pp. 591–611, 2016.
[3] F. Nachbar et al., “The ABCD rule of dermatoscopy: high
prospective value in the diagnosis of doubtful melanocytic skin
lesions,” J. Am. Acad. Dermatol., vol. 30, no. 4, pp. 551–559, 1994.
[4] “The International Skin Imaging Collaboration (ISIC).”
https://www.isic-
archive.com/#!/topWithHeader/tightContentTop/about/isicArchive
(accessed Oct. 04, 2020).
[5] H. Alquran et al., “The melanoma skin cancer detection and
classification using support vector machine,” 2017 IEEE Jordan
Conf. Appl. Electr. Eng. Comput. Technol. AEECT 2017, vol. 2018-
Janua, pp. 1–5, 2017.
[6] U. B. Ansari and M. E. Student, “Skin Cancer Detection Using
Image Processing Tanuja Sarode 2,” Int. Res. J. Eng. Technol., vol.
Fig. 4. Confusion Matrix for ISIC-ISBI-2016 4, no. 4, pp. 2395–56, 2017, [Online]. Available:
The https://www.irjet.net/archives/V4/i4/IRJET-V4I4702.pdf.
proposed system gives remarkable results using the Random [7] M. A. Khan et al., “An implementation of normal distribution based
Forest classifier in terms of accuracy, sensitivity, precision, and segmentation and entropy controlled features selection for skin
AUC-ROC. lesion detection and classification,” BMC Cancer, vol. 18, no. 1, pp.
1–20, 2018.
[8] V. M. M, “Melanoma Skin Cancer Detection using Image
Processing and Machine Learning,” Int. J. Trend Sci. Res. Dev., vol.

3URFHHGLQJVRIWK,QWHUQDWLRQDO%KXUEDQ&RQIHUHQFHRQ$SSOLHG6FLHQFHV 7HFKQRORJ\ ,%&$67  

,VODPDEDG3DNLVWDQ±-DQXDU\
Authorized licensed use limited to: Miami University Libraries. Downloaded on June 14,2021 at 20:07:58 UTC from IEEE Xplore. Restrictions apply.
 Volume-3, no. Issue-4, pp. 780–784, 2019.
[9] F. Dalila, A. Zohra, K. Reda, and C. Hocine, “Segmentation and
classification of melanoma and benign skin lesions,” Optik (Stuttg).,
vol. 140, pp. 749–761, 2017.
[10] R. Sumithra, M. Suhil, and D. S. Guru, “Segmentation and
classification of skin lesions for disease diagnosis,” Procedia
Comput. Sci., vol. 45, no. C, pp. 76–85, 2015.
[11] N. V Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer,
“SMOTE: synthetic minority over-sampling technique,” J. Artif.
Intell. Res., vol. 16, pp. 321–357, 2002.
[12] S. Aksoy and R. M. Haralick, “Feature normalization and
likelihood-based similarity measures for image retrieval,” Pattern
Recognit. Lett., vol. 22, no. 5, pp. 563–582, 2001.
[13] X. H. Cao and Z. Obradovic, “A robust data scaling algorithm for
gene expression classification,” 2015 IEEE 15th Int. Conf.
Bioinforma. Bioeng. BIBE 2015, no. 1, 2015.
[14] G. John, R. Kohavi, and K. Pfleger, “IrreleJohn, G., Kohavi, R., &
Pfleger, K. (1994). Irrelevant Features and the Subset Selection
Problem. Icml, 121–129. Retrieved from http://machine-
learning.martinsewell.com/feature-
selection/JohnKohaviPfleger1994.pdfvant Features and the Subset
Selectio,” Icml, pp. 121–129, 1994, [Online]. Available:
http://machine-learning.martinsewell.com/feature-
selection/JohnKohaviPfleger1994.pdf.
[15] Y. Masoudi-Sobhanzadeh, H. Motieghader, and A. Masoudi-Nejad,
“FeatureSelect: A software for feature selection based on machine
learning approaches,” BMC Bioinformatics, vol. 20, no. 1, pp. 1–17,
2019.

3URFHHGLQJVRIWK,QWHUQDWLRQDO%KXUEDQ&RQIHUHQFHRQ$SSOLHG6FLHQFHV 7HFKQRORJ\ ,%&$67  

,VODPDEDG3DNLVWDQ±-DQXDU\
Authorized licensed use limited to: Miami University Libraries. Downloaded on June 14,2021 at 20:07:58 UTC from IEEE Xplore. Restrictions apply.