Machine Learning-based Approaches for Skin

Cancer
Detection: An Overview of Different Algorithms
Kritika Dangol
School of Computer Science and Engineering
Lovely Professional University
Phagwara, Punjab, India
[email protected]

Abstract—Skin cancer is a highly prevalent disease that can be squamous cell carcinomas, despite the fact that melanoma is
effectively treated if detected early. Therefore, it is vital to the most dangerous. Melanoma can spread to other organs
recognize its warning signs and symptoms to identify potential outside of the skin if ignored or discovered too late, making it
indications. However, insufficient knowledge of preventive
difficult to treat and sometimes fatal. Fortunately, skin cancer
measures has increased skin cancer fatalities. Early detection is
crucial to prevent the disease's progression and spread, as it tends
may usually be cured with early detection and treatment.
to gradually affect other body parts. Given the high mortality
rates, increasing incidence, and expensive medical treatments Typically, a proficient dermatologist employs a series of
associated with skin cancer, timely symptom detection is critical. techniques to diagnose skin cancer, which begins with a
Malignant melanoma, in particular, is becoming more prevalent naked-eye examination of suspicious lesions, followed by
each year, and its detection can be challenging due to low magnified examination of lesions known as dermoscopy and
contrast or artefacts, making it appear similar to moles or scars. finally a biopsy. However, this approach is time-consuming,
Fortunately, technological advancements and research and the patient's condition may progress to advanced stages.
methodologies have made skin cancer detection easier.
Additionally, an accurate diagnosis is subjective, and it is
Nonetheless, certain types of skin cancer still pose challenges to
their detection, and further progress is required to overcome reliant on the expertise of the clinician. Studies have shown
these obstacles. Various machine learning methods, particularly that even the most adept dermatologists have an accuracy rate
deep learning, can significantly aid in the timely detection of skin of less than 80% in precisely diagnosing skin cancer.
cancer. This review paper provides a systematic evaluation of Furthermore, a scarcity of skilled dermatologists in the public
different machine-learning algorithms applied to skin cancer healthcare sector exists globally, exacerbating these
datasets and their respective accuracy. difficulties.
Keywords—Skin cancer, Machine learning, Deep learning, The potential for artificial intelligence (AI) in healthcare is
Malignant melanoma
enormous, particularly for dermatology. Machine learning
(ML), a subfield of AI, uses algorithms and statistical
I. INTRODUCTION frameworks to extract information from data and predict
outcomes for fresh samples. Although AI has previously
Skin cancer, also known as melanoma, is a disorder in which shown to be a useful tool for diagnosing skin cancer,
the epidermis, the top layer of skin, experiences uncontrollable dermatologists have not yet used it to the same extent as
growth and multiplication of malignant cells. Damage to DNA radiologists. The general population may now more easily use
and mutations are the cause of this unregulated cellular machine learning thanks to its growing accessibility and
proliferation. The uncontrolled cell division can result in the technological breakthroughs. The creation of systems that
growth of malignant tumors. New skin cells are typically analyze skin image data and correctly identify skin cancer at
produced when old ones are injured or die, but when this an early stage can be helped by deep convolutional neural
procedure is hampered, abnormal cells may begin to develop networks. Effective treatment and better results for melanoma
quickly. Skin cancer has risen to the top of the most prevalent patients depend on early identification. Although
cancers in recent years. The biggest organ in the body, the dermatologists are adept at detecting skin cancer, their
skin, makes skin cancer the most common kind of cancer in shortage makes the creation of computerized systems that can
people. Melanoma and non-melanoma skin cancer are the two do so necessary, saving lives and lessening the financial and
main forms of skin cancer, with the former being the most medical burdens placed on patients. In reaching this goal,
prevalent. Basal celluar carcinoma and squamous cell machine learning may be incredibly useful.
carcinoma are further subtypes of non-melanoma skin cancer.
The most common types of skin cancer include basal cell and A survey of the body of research on the subject is included in
Section II of this article. A method of study is provided in
1
Section III, and Section IV conducts a comparison analysis.
Section V covers the study's findings whilst Section VI wraps II. LITERATURE REVIEW
up the conclusions.
M. Krishna Monika et al. 2020 [1] introduced a method based quantitative measurements to assess the performance of the
on MSVM classification that combines two efficient feature proposed technique, which is trained and tested on the ph2
extraction techniques, namely ABCD and MSVM. Using dataset. With accuracy, sensitivity, specificity, and precision
high-resolution dermoscopic pictures gathered from the ISIC values of 98.61%, 98.33%, 98.93%, and 97.73%,
2019 challenge dataset, the system's goal is to correctly respectively, the suggested technique obtains excellent values
diagnose eight distinct forms of skin malignancies. The Dull for these metrics. When the suggested method's performance
Razor technique and the Gaussian filter are used by the
is compared to that of current approaches, it is discovered to
authors to improve the pictures' quality, while a Median filter
perform better in terms of classification accuracy.
is used to reduce noise. The preparation step is where these
methods are used. Furthermore, color-based k-means A hybrid method for the diagnosis of melanoma is introduced
clustering is used to segment the preprocessed pictures. The by Jinen Daghrir et al. 2020 [5] and may be used on troubling
accuracy of the method obtained is quite great, at about skin lesions. A convolutional neural network and two
96.25%. conventional machine learning classifiers which are trained
on variables including the texture, colour, and boundaries of
A method for classifying and segmenting skin lesions as skin lesions make up the three prediction techniques used in
benign or malignant was presented by Arslan Javaid et al. the proposed method. Then, to improve their aggregate
2021 [2]using both image processing and machine performance, these techniques are incorporated utilising
learning.Through the use of mean and standard deviation majority voting. According to the testing data, combining all
approaches, subsequently followed by segmentation utilising three procedures yields the best degree of accuracy.The study
the OTSU thresholding algorithm, this work presents a discovered that the KNN classifier had the lowest
unique way for improving the contrast of dermoscopic performance when only taking into account the five closest
pictures. The Grey level Co-occurrence Matrix (GLCM) neighbours. This finding suggests that because of its
features and the histogram of oriented gradients (HOG) sensitivity to outliers, the KNN classifier is ineffective in
approach are then used to extract texture, colour, and object identifying malignant skin lesions. The SVM classifier, on the
contrary , succeeded better than KNN because of its
data from the segmented pictures. Synthetic Minority
effectiveness and flexibility. CNN was thought to be a more
Oversampling Technique (SMOTE) sampling is utilised to
potent and robust method for diagnosing melanoma skin
address the issue of class imbalance while Principal cancer while obtaining respectable performance. KNN, SVM,
Component Analysis (PCA) is employed to minimise the and CNN each had accuracy rates of 57.3%, 71.8%, and
dimensionality of the HOG features. The generated feature 85.5%; when all three algorithms were combined, the
vector is scaled and standardised before categorization and accuracy rate was 88.4%.
wrapper-based feature selection. Several classifiers are used
for classification, namely Quadratic Discriminant, SVM An totally automated technique for identifying dermatological
(Medium Gaussian), alongside Random Forest. On the illnesses using lesion photos was proposed by Vijayalakshmi
publicly accessible dataset of ISIC-ISBI 2016, the suggested M M et al. 2019 [6]. Instead of the more common medical
technique outperformed other widely used classifiers, personnel-based detection method, this strategy uses
achieving a classification accuracy of 93.89% applying the machine-based detection. Three stages make up the model:
Random Forest classifier. The findings indicate that the data gathering and augmentation, model design, and
prediction.
suggested method of extending contrast before segmentation
The research included a number of pre-processing
produces excellent results, and the suggested feature selection
procedures, including segmentation and the removal of hair,
method in combination using the random forest classifier shadows, and glare. In order to create an enhanced structure
offers potential for precise classification. and achieve an accuracy rate of 85%, the authors used a
variety of AI techniques, such as Convolutional Neural
An automated system for classifying skin lesions has been
Networks and Support Vector Machines, that were paired
suggested by Khalid M. Hosny et al. 2018 [3]. This paper
with image processing software. This methodology's capacity
suggests a way for classifying skin lesions using a deep for fast detection, which enables technicians to improve their
learning network that has already been trained with transfer diagnostic abilities, is a significant component. This method's
learning methods. In particular, a softmax function is used to distinctiveness is in its automation and speed, makes it an
identify three distinct kinds of tumors: melanoma, common effective tool for the early detection of dermatological
nevus, and atypical nevus. This replaces the last layer of conditions.
AlexNet. The suggested model is further tuned, and data
augmentation is employed to enhance performance. The Rehan Ashraf et al. 2020 [7] suggested a smart system built
accuracy, sensitivity, specificity, and precision are utilised as around the Region of Interest (ROI) to recognise and
discriminate between melanoma and nevus malignancy by
2
applying the transfer learning technique. They extracted ROIs melanoma. In comparison with the original and non-
from the photos using an enhanced k-mean technique. Given augmented techniques, they transferred the first low-level
that the system is trained using pictures containing only feature layers into the AlexNet model and assessed the ROI
melanoma cells, this ROI-based technique helps in detecting using augmentation, which produced the best results.
discriminative characteristics. The authors presented a Additionally, using ROI pictures from the DermIS and
transfer learning-based effective technique using the AlexNet DermQuest datasets, the
model to effectively categorise and diagnose skin cancer
scientists applied a Convolutional Neural Network (CNN) 95.6%, respectively. These results demonstrate how well the
based transfer learning model with data augmentation. technique recognises and segments skin lesions.
DermIS and DermQuest accuracy rates for the proposed
system were 97.9% and 97.4%, respectively.
III. RESEARCH METHODOLOGY
A method for artificially detecting skin cancer that
incorporates machine learning and image processing An overview of the techniques for conducting the research for
approaches was released by Mahamudul Hasan et al. 2019 analysis of different machine learning algorithms is given in
[8]. They suggested a Convolutional Neural Network-based this section.
method for melanoma classification by segmenting
dermoscopic pictures and extracting attributes of damaged Table I. presents a graphical representation of the methods used for executing
the study.
skin cells. The resultant approach may help both people and
medical professionals diagnose and categorise skin cancer as
benign or malignant. The group used a deep learning method
and a convolutional neural network classifier to divide the
retrieved characteristics into groups. On publicly accessible
data sets, their method produced an accuracy of 89.5%, and a
training accuracy of 93.7%.

For the categorization of skin lesions using the HAM10000

dataset, Ioannis Kousis et al. 2022 [9] published a research in
which they looked into 11 alternative CNN architectures. The
study used methods including data augmentation, transfer
learning, and fine-tuning to solve the problems of unbalanced
data and picture similarity among various skin lesions.
DenseNet169 outperformed previous state-of-the-art
methods, reaching accuracy of 92.25%, sensitivity of 93.59%,
and an F1-score of 93.27% among the 11 designs. In addition,
the scientists created an Android app that lets users manually
trim photos of skin lesions taken with their mobile device's
camera using a lightweight variant of DenseNet169. The
programme goes on to categorise the lesion as benign or
malignant and offers details on the appropriate amount of sun
exposure time depending on the user's skin's phototype, the
level of sunscreen, and the level of UV radiation that is now
present.

Deep learning was used by Marriam Nawaz et al. 2021 [10]

to create an automated method for segmenting early-stage
cutaneous melanoma. Their approach helps dermatologists
identify this fatal condition by using fuzzy k-means clustering
(FKM) and quicker region-based convolutional neural
networks (RCNN). Images are initially pre-processed in order
to improve visual information and remove noise and lighting
concerns. The afflicted skin region is then segmented using
FKM and faster RCNN to create a fixed-length feature vector
with a range of sizes and limits. The method outperformed
state-of-the-art techniques when it was evaluated on three
common datasets, ISBI-2016, ISIC-2017, and PH2. Its
average accuracy was 95.40%, followed by 93.1% and
3
 This section provides a comparison of several machine
learning models used for skin cancer detection. The accuracy
of various models and the techniques employed are presented
in Table II.

Table II. displays a comparison of various techniques that have been

discussed in the literature survey.

REFERENCE
ALGORITHM/
PAPER DATASET ACCURACY
MODEL
ISIC 2019 Challenge MSVM 96.25%
[1]
Dataset
Random
Forest 92.47%
[2] ISIC-ISBI 2016
classifier

AlexNet 98.61%
[3] PH2 dataset
SVM 97.8%
[4] ISIC
CNN 94.25%
[5] ISIC
SVM 85.27%
[6] ISIC dataset
97.9% ,
DermIS and CNN
[7] 97.4%
DermQuest
CNN 89.5%
[8] Random dataset
CNN 92.25%
[9] HAM10000
95.40%,
RCNN along
ISIC-2016, ISIC- 93.1%,
[10] FKM
2017, PH2 95.6%

The investigation discovered that the author, who used the

pre-defined model AlexNet, obtained the maximum accuracy
of 98.61%. SVM had an accuracy of 93.89%, CNN models
had an overall accuracy of 95.89%, and the Random Forest
Classifier model had the lowest accuracy at 92.47%.

V. OUTCOME

The study found that the highest accuracy of 98.61% was

Due to its accuracy and capacity for handling vast volumes of achieved by the author who utilized the pre-defined model
data and its ease of pattern recognition, machine learning has AlexNet. Models using CNN had an overall accuracy of
become essential for the early identification of skin cancer. 95.89%, SVM had an accuracy of 93.89%, while the least
The papers were first and foremost evaluated for their
accurate model was the Random Forest Classifier with an
applicability to the study's topic. Three crucial considerations
were taken into account when selecting the literature survey accuracy of 92.47%. The application of transfer learning has
data for this study. The published paper must first be up-to- significantly improved the accuracy of the model.
date after which it must be relevant to the subjects, and
finally, the algorithms must be compatible with the study.
The accuracy attained by the model or algorithm is then taken VI. CONCLUSION
into account, together with the dataset and comparative table
that is designed. A summary is given that compares and This paper unfolds the use of various Machine Learning
explains the outcomes of the algorithms. algorithms in skin cancer detection and also gives an idea of
how various researchers used different Machine Learning
IV. COMPARATIVE STUDY
techniques to detect cancer. The principal objective of this
paper is to mention an overview of how machine Learning
4
techniques like CNN, SVM, Random Forest Classifier are [5] Jinen Daghrir 1,2, Lotfi Tlig 2, Moez Bouchouicha 3, Mounir
Sayadi 2. “Melanoma skin cancer detection using deep learning
used to detect skin cancer and also evaluate their
and classical machine learning techniques: A hybrid approach”
performances based on accuracy and other performance DOI: 10.1109/ATSIP49331.2020.9231544
metrics. [6] Vijayalakshmi M M. “Melanoma Skin Cancer Detection using
Image Processing and Machine Learning”
[7] Rehan Ashraf; Sitara Afzal; Attiq Ur Rehman; Sarah Gul; Junaid
Baber; Maheen Bakhtyar; Irfan Mehmood. “Region-of-Interest
REFERENCES Based Transfer Learning Assisted Framework for Skin Cancer
Detection”
[1] M. Krishna Monika, N. Arun Vignesh, Ch. Usha Kumari, DOI: 10.1109/ACCESS.2020.3014701
M.N.V.S.S. Kumar, E. Laxmi Lydia. “Skin cancer detection and [8] Mahamudul Hasan, Surajit Das Barman, Samia Islam, Ahmed
classification using machine learning.” Wasif Reza.“Skin Cancer Detection Using Convolutional Neural
DOI: 10.1016/j.matpr.2020.07.366 Network”
[2] A. Javaid, M. Sadiq and F. Akram, “Skin Cancer Classification DOI: 10.1145/3330482.3330525
Using Image Processing and Machine Learning,” 2021 [9] Ioannis Kousis, Isidoros Perikos, Ioannis Hatzilygeroudis, and
International Bhurban Conference on Applied Sciences and Maria Virvou. “Deep Learning Methods for Accurate Skin Cancer
Technologies (IBCAST), Islamabad, Pakistan, 2021, pp. 439-444, Recognition and Mobile Application”
DOI:10.1109/IBCAST51254.2021.9393198. DOI: 10.3390/electronics11091294
[3] Khalid M. Hosny, Mohamed A. Kassem, and Mohamed M. [10] Marriam Nawaz1, Zahid Mehmood2, Tahira Nazir 1, Rizwan Ali
Foaud. “Skin Cancer Classification using Deep Learning and Naqvi 3, Amjad Rehman 4, Munwar Iqbal 2. “Skin cancer
Transfer Learning” detection from dermoscopic images using deep learning and fuzzy
DOI:10.1109/CIBEC.2018.8641762 k-means clustering.”
[4] M. Vidya, Maya V. Karki. “Skin Cancer Detection using Machine DOI:10.1002/jemt.23908
Learning Techniques”
DOI: 10.1109/CONECCT50063.2020.9198489

5
6