![Skin cancer types [70]](https://www.researchgate.net/profile/V-Auxilia-Nancy/publication/373162909/figure/fig1/AS:11431281207469133@1701227447557/Skin-cancer-types-70_Q320.jpg)
![PRISMA Methodology-systematic review on algorithms and melanoma identification [59]](https://www.researchgate.net/profile/V-Auxilia-Nancy/publication/373162909/figure/fig3/AS:11431281207411728@1701227447860/PRISMA-Methodology-systematic-review-on-algorithms-and-melanoma-identification-59_Q320.jpg)
Content uploaded by V Auxilia Osvin Nancy
Author content
All content in this area was uploaded by V Auxilia Osvin Nancy on Jan 22, 2024
Content may be subject to copyright.
Vol.:(0123456789)
Multimedia Tools and Applications
https://doi.org/10.1007/s11042-023-16422-6
1 3
Comparative study andanalysis onskin cancer detection
using machine learninganddeep learning algorithms
V.AuxiliaOsvinNancy1· P.Prabhavathy1· MeenakshiS.Arya2· B.ShamreenAhamed1
Received: 5 August 2022 / Revised: 23 May 2023 / Accepted: 24 July 2023
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2023
Abstract
Exposure to UV rays due to global warming can lead to sunburn and skin damage, ulti-
mately resulting in skin cancer. Early prediction of this type of cancer is crucial. A detailed
review in this paper explores various algorithms, including machine learning (ML) tech-
niques as well as deep learning (DL) techniques. While deep learning strategies, particu-
larly CNNs, are commonly employed for skin cancer identification and classification, there
is also some usage of machine learning and hybrid approaches. These techniques have
proven to be effective classifiers of skin lesions, offering promising results for early detec-
tion. The paper analyzes various researchers’ reviews on skin cancer diagnosis to identify
a suitable methodology for improving diagnostic accuracy. A publicly available dataset of
dermoscopic images retrieved from the ISIC archive has been trained and evaluated. Per-
formance analysis is done, considering metrics such as test and validation accuracy. The
results indicate that the RF(random forest) algorithm outperforms other machine learning
algorithms in both scenarios, with accuracies of 58.57% without augmentation and 87.32%
with augmentation. MobileNetv2, ensemble of Dense Net and Inceptionv3 exhibit superior
performance. During training without augmentation, MobileNetv2 achieves an accuracy
of 88.81%, while the ensemble model achieves an accuracy of 88.80%. With augmenta-
tion techniques applied, the accuracies improved to 97.58% and 97.50%, respectively. Fur-
thermore, experiment with a customized convolutional neural network (CNN) model was
also conducted, varying the number of layers and applying various hyperparameter tuning
methodologies. Suitable architectures, including a CNN with 7 layers and batch normaliza-
tion, a CNN with 5 layers, and a CNN with 3 layers were identified. These models achieved
accuracies of 77.92%, 97.72%, and 98.02% on the raw data and augmentation datasets,
respectively. The experimental results suggest that these techniques hold promise for inte-
gration into clinical settings, and further research and validation are necessary. The results
highlight the effectiveness of transfer learning models, in achieving high accuracy rates.
The findings support the future adoption of these techniques in clinical practice, pending
further research and validation.
Keywords Melanoma· Benign· Machine learning· Deep learning· SVM· KNN·
Random forest· CNN· VGG16
Extended author information available on the last page of the article
Multimedia Tools and Applications
1 3
1 Introduction
Healthcare knowledge extraction is a highly difficult task due to issues, like noisy and
unbalanced datasets which come from clinical investigations, and are generally uncer-
tain and variable. Recently, a variety of machine learning algorithms have been reviewed
and implemented to assess their potential for usage in the medical industry. Typically,
the efficiency of the techniques is compared to the actual diagnostics of the medical
domain experts professionals that specializes in diagnosing particular diseases to pro-
vide an efficient methodological evaluation of classifiers using performance indicators.
Accuracy, sensitivity, and specificity are the three performance metrics that make up the
confusion matrix for any algorithm that is implemented [82].
Diseases, like COVID, pneumonia, heart problems, brain tumors, and skin cancer,
can be detected using ML and DL algorithms. Certain diseases require image input for
detection. In [6, 9], the lightweight hybrid deep learning approach (CNN + LSTM) is
implemented for cardiac vascular disease classification. The different types of spectrum
classification using ECG heart beats have been analyzed using CNN architectures [9].
Chest X-ray images were trained by hybrid approach (CNN + SVM, KNN, RF, Softmax)
to classify the normal based and pneumonia features [71]. Using PCG signal, AOCNet
was capable for classifying the heart sounds normal or abnormal [8]. In [7] CAD, cat-
egorization was implemented using SVM in digital for detection of the breast cancer.
Unchecked melanocyte development is the cause of melanoma [83]. Factors such as
symmetry, color, size, and shape play a crucial role in detecting skin cancer and distin-
guishing between benign and malignant lesions [83]. Many countries worldwide, espe-
cially the United States, have reported increasing death rates from skin cancer [70].
Recent cancer statistics reveal an estimated 1.9 million new cases of this type of cancer,
with an anticipated 608,570 deaths in the United States alone [17]. Early diagnosis is cru-
cial in reducing the mortality rate. Exposure to sunlight is strongly associated for develop-
ing both malignant and benign type of skin cancer [70]. Current classifications include
malignant melanoma and benign, including BCC and SCC as shown in Fig.1 [70].
Types of Malignant Skin cancer:
Melanoma (Fig. 1a): Melanoma, the most severe malignant tumor, begins in the
melanocyte cells of the epidermis. The tumor spreads quickly, has a high fatality rate
due to early metastasis, and is challenging for treatment. Although it only accounts
for 4% of all skin cancers, it results in death in 80% of cases. However, if detected
early, it has a 95% cure rate, highlighting the significant impact of early diagnosis on
improving survival prospects [69].
(a)MEL (b) SCC (c) BCC
Fig. 1 Skin cancer types [70]
Multimedia Tools and Applications
1 3
Squamous cell carcinoma (Fig.1b: The outermost layer of the epidermis, known as
squamous cells. Early detection makes it easily treatable, but if left untreated, it can
penetrate the skindeeper layers.
Basal cell carcinoma (Fig.1c): The basal cells, which reside in the innermost layer of
the epidermis, are where around 80% of cases begin. Since basal cell development is
slow, BCC is often highly treatable and causes minimal harm if detected and treated
promptly.
Melanoma, considered a severe type of skin cancer due to its rising incidence rate.
Timely detection is of utmost importance in maximizing curative outcomes and ensuring
optimal recovery, as failure to identify it promptly can lead to fatal consequences [65].
Determination of skinlesion classes is important for assessing the stage of the disease and
evaluating the probability of being at risk.
The following are the skin lesion classes.
1. BACC -Basal cell carcinoma
2. SEK- Seborrheic keratosis
3. ACK- Actinic keratosis
4. DEF-Dermato fibroma
5. MEL -Melanoma
6. NEVI -Nevus
7. VASCL-Vascular lesion
Dermatologists have historically utilized various approaches, such as the ABCD rule,
three checklists, seven checklists, and the Menzies rule, to detect and diagnose skin can-
cer. However, these dermoscopic imaging procedures are susceptible to errors and require
extensive expertise. Detecting skin lesions using traditional methodologies involves
employing common techniques like thresholding, clustering, edge-based analysis, and
regional approaches [29, 65] (Table1).
The aforementioned approaches were found to lack sufficient prediction accuracy. As
a result, the researchers are in need of an alternative classification and detection scheme
for cancer cells. The system flow is illustrated in Fig.2. The comparative analysis of skin
lesion using ABCD rule has been shown in Table. 2.
Pre-processing techniques are applied to mitigate picture artifacts and enhance lesion
visibility. The hair removal approach comprises two main steps: hair restoration and hair
detection, employing morphological operations, thresholding, and filtering methods like
Gaussian, median, and middle filters [66].
Color calibration is a crucial step in image enhancement to restore accurate lesion colors.
Additional image enhancement methods involve adjusting lighting conditions, enhancing
contrast, and improving edge sharpness. Illumination reflectance techniques, Retinex algo-
rithms [100], and filters such as bilateral filtering are commonly utilized to handle illumina-
tion variations. Contrast augmentation typically involves a combination of histogram equali-
zation, adaptive histogram equalization, and sharp masking. Edge detection is performed
using PCA (principle component analysis). Feature extraction can be achieved manually or
automatically using machine learning methods. The ABCD rule methodology relies on fea-
ture extraction to facilitate effective skin lesion classification by dermatologists [100].
When applying ML techniques, without any prior knowledge of the issue, learnt features
are automatically constructed from datasets. There are several options available, ranging
Multimedia Tools and Applications
1 3
Table 1 Criteria used by dermatologists for diagnosing melanoma
Approaches Description Methodology
ABCDE [29] The rule is based on morphological characteristics that encompass asymmetry, irregular edges,
non-homogeneous coloration, and a diameter equal to or greater than 6mm. It also considers
the concept of evolution, which involves assessing changes over time in parameters such as
shape, size, color, elevation, and the presence of symptoms like itching or bleeding.
1. Acquisition of Image
2. Preprocessing (PP)
3. Image segmentation
4. Selecting Feature using ABCDE rule
5. Classification
7 Point Checklist [43] Each of the minimal criteria, namely pigmentation, irregular streaks, and irregular patches, can
be assigned a score. Additionally, the major criterion of abnormal pigment networks can also
be assigned a score.
1. Image processing techniques
2. Pattern recognition
Menzies Method [73] There are a total of 11 features that are evaluated as either present or absent, including two
negative features and nine positive features.
Positive characteristics of melanoma include: Mul-
tiple brown dots scattered throughout a blue-white
veil, focal pseudopods, scar-like depigmentation,
and five to six different coloured black spots or
globules on the periphery.
Multimedia Tools and Applications
1 3
from traditional methods to DCNN. In this review article, multiple methods are investi-
gated to determine their performance in classifying skin lesions and detecting skin cancer.
Numerous studies have demonstrated that deep learning techniques exhibit superior effi-
ciency compared to skilled dermatologists when it comes to identifying and detecting skin
lesions. However, it is important to evaluate these methods that fall outside their training
scope. Automatic dermoscopic image classification encounters various difficulties, includ-
ing insufficient distinguish between the lesion and normal skin, similarities in appearance
between benign and malignant cases, and difference in individuals’ conditions of skin. In
relation to the latter point, skin tone is an important but often overlooked factor [54].
The available datasets for training machine learning algorithms primarily consist of
dermoscopic images of individuals with pale skin. However, there is a need to address
this gap and expand the datasets to ensure accurate performance in identifying skin
lesions in individuals with dark skin.
The utilization of algorithms offers several advantages. Firstly, it provides an opportu-
nity to avoid unnecessary biopsies and ensure timely diagnosis of melanomas. Addition-
ally, these techniques enable the diagnosis of skin conditions without physical contact. This
not only enhances patient comfort but also reduces the cost associated with non-melanoma
skin cancer diagnosis and treatment, which has been shown to be expensive.
There are multiple sections in this article. The review’s methodology was covered in
Section2 of the article. In Section3, several techniques (ML and DL), and hybrid AI
strategies are analysed. Review of ML/DL algorithms is the primary topic of Section4.
In Section5, an experimental analysis on deep learning models has been conducted.
Fig. 2 System architecture
Multimedia Tools and Applications
1 3
Table 2 Comparative analysis of skin lesion using ABCD rule
Ref Diagnosis Dataset Technique Proposed methodology Accuracy Drawbacks
[66] 2 class (Benign and Mela-
noma)
200 dermoscopic images ABCD rule 1. Preprocessing: Gabor
filters were used to pre-pro-
cess the image by detecting
the hair
2. Geodesic Active Contours:
To detect the lesion bounda-
ries
Sensitivity- 91.25%
Specificity-95.83%
Overall Accuracy-94%
Possible for misclassification
Bad contour detection
Inability to detect structure
Can’t work with large dataset.
[100] 2 class (Benign and Mela-
noma)
ISIC 2016, 2017, 2018 and
PH2
ABCD 1. Preprocessing: Unwanted
hair in the skin lesion were
removed using median
filters.
2. LSM (Lesion segmenta-
tion module): To divide the
area of interest (ROI)
3. BDM (Border Detection
Module): A segmented
image input into the BDM
helps to identify the lesion
area’s single pixel border.
4. Utilize CIEM to extract
the color regions from the
lesion (Color Information
Extraction Module).
Sensitivity-97.69%
Specificity-97.97%
Overall Accuracy-97.86%
Trained and Tested separately
rather than combining the
dataset.
In this article the proposed
system works on separate
dataset.
Multimedia Tools and Applications
1 3
Section6 addresses the research challenges. Section7 consists of discussions, conclu-
sions and future enhancements based on the review.
2 Methodology
2.1 Searching strategy
An extensive literature review was conducted, focusing on predefined endpoints to deter-
mine the appropriate technique for skin cancer detection. To select relevant articles, spe-
cific keywords such as Melanoma, dermoscopic images, machine learning, deep learn-
ing, and melanoma classification were used in combination with “and” and “or” operators
in electronic database searches. The inclusion of studies and extraction of relevant data
followed the preferred reporting guidelines for systematic reviews. All relevant informa-
tion was obtained from reputable sources and meticulously analyzed for further analysis.
The completion of this survey will provide insightful, logical, and robust solutions to the
research questions at hand.
2.1.1 Structure ofstudy
To delineate the study, the initial phase involves assessing the system using a three-layered
approach.
• The first layer is the planning stage, where the overall framework and objectives of
the study are established. This includes defining the research questions, identifying the
scope of the study, and outlining the methodology to be employed.
• The second layer focuses on data selection and evaluation. In this stage, relevant datasets
are identified and collected for analysis. The quality and suitability of the data are carefully
assessed, considering factors such as sample size, data accuracy, and representativeness.
• The third layer encompasses result generation and conclusion. Here, the gathered data
is analysed through the use of suitable statistical and computational techniques to pro-
duce insightful findings and conclusions. The results obtained are rigorously evaluated
and interpreted, taking into account the research objectives and hypothesis. Finally, a
comprehensive conclusion is drawn based on the findings, highlighting the implications
and potential avenues for future research.
This three-layered approach provides a systematic and robust analysis.
2.1.2 Research problems
Framing the study questions is crucial to conduct a focused and relevant review on the
specific topic. The following inquiries have been formulated based on the most recent sys-
tematic review:
• What methods are employed in for skin cancer identification?
• What methods are utilized in DL for the skin cancer diagnosis?
• What are the key data points required for accurate identification of skin cancer?
• How can the most appropriate learning model for prediction be determined?
Multimedia Tools and Applications
1 3
By addressing these research questions, the study aims to explore and analyze the var-
ious methodologies used in both ML and DL approaches for skin cancer identification.
Additionally, the investigation aims to identify the essential data points required to enhance
the accuracy. Lastly, the study seeks to find the learning method suits to do effective pre-
diction in the context of diagnosis. Through a comprehensive survey of the existing litera-
ture, valuable insights and recommendations can be provided in response to these impor-
tant research questions.
2.1.3 Selection andevaluation criteria
A comprehensive search was conducted to gather relevant research articles for this review.
A total of 150 full-text articles were examined from reputable databases including Web
of Science, Scopus, and PubMed. After the abstract screening process, 115 studies were
selected for further analysis. The selection criteria focused on recent advancements in mel-
anoma identification using learning techniques, the visibility and impact of the contribu-
tions, as well as the publication venue (high-ranking conferences and journals) and cita-
tion count. The chosen studies represent the most significant and representative research
on learning-based melanoma detection, primarily published between 2015 and 2021, with
a focus on the years 2018 to 2021. The quality of the selected research papers has been
carefully evaluated to ensure their relevance and reliability. This evaluation includes con-
sidering factors such as journal reputation and citation metrics. The process adhered to the
PRISMA guidelines for systematic review and meta-analysis, as depicted in Fig.3 [59].
This diagram shows the study selection process and inclusion/exclusion criteria, ensuring
transparency and rigor in the review methodology.
2.2 Skin lesion –public databases
Computerized systems play a crucial role in diagnostics, especially in melanoma. In this
section, the dataset used for skin cancer detection approaches has been discussed [47]. The
datasets serve as a valuable resource for training and evaluating various algorithms and
techniques. The diagnostic models performance can be analyzed by dermoscopic images
that enable researchers to assess accurately.
2.2.1 ISIC archive
ISIC collection offers a wide range of datasets for cutaneous lesions. One of the promi-
nent datasets is the ISIC2016. It consists of 900 and 379 dermoscopic images of training
and testing subset. ISIC2017 collection, the dataset is divided into three categories: benign
nevi, benign SK, and melanoma, with a total of 2000 images. Among these, 150 images for
validation, and 600 are designated for test process [47].
Moving on to ISIC2018, the dataset includes seven classes of skin lesions along with
metadata, training and testing input images [22]. The categories in the ISIC2019 data-
set encompass 8 classes BACC, SEK, ACK, DEF, MEL, NEVI, VASCL, and SQCC.
The test data for ISIC2019 consists of 8239 images with outlier classes. [22]. To enable
comprehensive skin cancer detection, modern methods should be capable of accurately
identifying images from these diverse categories. Additionally, it has the metadata
information for analysis and research purposes.
Multimedia Tools and Applications
1 3
2.2.2 HAM10000
The human-versus-machine dataset for skin lesions, known as HAM10000, consists of
10,000 training photos [95]. This freely accessible dataset is an important resource for
study and advancement in the field of skin lesion classification. The dataset were con-
tributed by the Cliff Rosendahl’s research and the skin cancer center in Queensland,
Australia [94].
The images were then edited manually and stored with dimensions of 800 to 600 pixels.
These standardized images ensure consistency and compatibility for analysis and evalua-
tion purposes within the dataset.
Fig. 3 PRISMA Methodology-systematic review on algorithms and melanoma identification [59]
Multimedia Tools and Applications
1 3
2.2.3 PH2
The rise in melanoma cases has spurred the computer-aided diagnosis techniques develop-
ment in the dermoscopic images classification. The PH2 dataset was created and serves as
a benchmarking tool for research purposes. The dataset was collected at the Pedro Hispano
Hospital Dermatology Centre in Portugal. The PH2 dataset consists of 200, with 80 images
representing common nevi, and another 80 images depicting rare nevi, and 40 images
showcasing melanoma skin cancer [72].
Each image in the dataset is accompanied by medical annotations, including lesion
segmentation, clinical diagnosis, and assessment of various dermoscopic criteria such as
colors and streaks [1]. These annotations provide valuable information for training and
algorithms evaluation for the contribution of skin lesion classification techniques.
2.2.4 Med node
The Med node systems macroscopic images were developed and tested using a dataset
consisting of 100 nevus photos and 70 melanoma images. The dataset was obtained from
the department of dermatology at the University Medical Centre Groningen (UMCG) [36].
These images were utilized to evaluate and validate the performance of the Med node sys-
tems macroscopic image analysis techniques. By using this dataset, the researchers were
able to assess the effectiveness and accuracy of their proposed methods in differentiating
between nevi and melanoma, contributing to the advancement of skin cancer diagnosis and
classification.
Table 3 contains NEVI/AN (Atypical NEVI), COM NEVI (Common Nevi), MEL
(Melanoma), SEK (Seborrheic keratosis), BACC (Basal cell carcinoma), DEF (Dermato
fibroma), ACK (Actinic keratosis), VASCL (Vascular lesion), SQCC (Squamous cell
carcinoma).
3 AI techniques
Artificial intelligence (AI) is the machine that carries out tasks need the intelligence of
people, such as learning, problem-solving, planning, and interpreting natural language. AI
can be divided into two main branches, each with its own characteristics. The first branch is
transfer learning (TL), which involves leveraging knowledge from pre-trained networks to
improve performance on new tasks or domains. TL allows the transfer of learned features
or representations from one task to another, reducing the need for extensive training on
new datasets. It enables the adaptation and reusability of models, making AI systems more
efficient and effective in handling various tasks [38].
3.1 ML algorithms
The four crucial steps in the ML method are pre-processing, feature extraction, seg-
mentation, and classification to categorise skin lesions. Pre-processing involves clean-
ing up the input data by removing noise and artefacts. Segmentation involves identify-
ing and delineating the lesion region in the image. Feature extraction is a crucial step
where relevant information is extracted from the segmented region. In the context of
machine learning, domain-specific features are commonly used. These features capture
Multimedia Tools and Applications
1 3
Table 3 Publicly available skin lesion dataset summary
Dataset Skin lesion classes Total. no. of images
NEVI/AN COM NEVI MEL SEK BACC DEF ACK VASCL SQCC
ISIC2016 [48] 726 – 173 – – – – – – 899
ISIC2017 [49] 1372 – 374 254 – – – – – 2000
ISIC2019 [50] 12,875 – 4522 2624 3323 239 867 253 628 25,331
ISIC2020 [51] 27,124 5193 584 135 – – – – – 33,126
HAM10000 6705 – 1113 1099 514 115 327 142 – 10,015
PH280 80 40 – – – – – – 200
MED NODE 100 – 70 – – – – – – 170
Multimedia Tools and Applications
1 3
important characteristics of the lesion that can be utilized in machine learning algo-
rithms. The next stage involves training the machine learning model using a dataset
that includes both image features and corresponding labels. Several methods, including
LR (linear regression), LOR (logistic regression), KNN, SVM, and RF are frequently
used in dermatology for prediction. SVM is particularly useful for classifying data into
different groups by searching for hyperplanes. RF generates an ensemble of randomly
constructed decision trees to determine the most frequent outputs. KNN is utilized for
data classification and retrieval based on the nearest neighbors. Each of these strategies
will be detailed in subsection.
3.1.1 SVM
SVM is a powerful ML model, capable of performing regressions, outlier identification,
and linear and nonlinear classifications using kernel techniques. SVMs are particularly
effective when dealing with complex datasets, although they may not perform optimally on
extremely large datasets. One of the key features of SVMs is their ability to create hyper-
planes to solve classification problems [23]. These hyperplanes separate different classes
or categories in the data space, allowing for accurate classification and prediction. SVMs
have been widely used in various domains due to their flexibility and ability to handle both
linear and nonlinear relationships in the data.
3.1.2 DT (decision trees)
DT (Decision trees) are flexible methods that do not rely on assumptions about data
distribution or classifier design. They can be applied to both categorical and numerical
variables, making them suitable for various tasks such as classification, regression, and
handling multiple outputs. Decision trees provide precise and effective classifications,
even for large and complex datasets. RF (Random forest) builds upon decision trees
by creating an ensemble model. RF combines multiple decision trees, each with high
variance, to form a more robust model with improved generalization performance. By
leveraging the diversity of individual trees, random forests enhance the accuracy and
stability of predictions [24].
3.1.3 KNN
The KNN technique categorise raw data according to how closely it resembles nearby
labelled data. The classifier’s parameters, such as K value, indicates the number of closest
neighbors to take into account, are established. The new data’s label is chosen by a signifi-
cant majority vote among its K closest neighbours. Distance measures like the manhattan
distance are used to calculate the data points separation. To prevent overfitting or underfit-
ting, a value of K between 3 and 10 is commonly selected. This range strikes a balance
between including enough neighbors for accurate classification and mitigating the impact
of outliers or noise [11].
Multimedia Tools and Applications
1 3
3.1.4 ANN
The approach developed from discussions surrounding the brain aimed to tackle challenges
as intricate as those encountered in the real world. Initially, attempts were made to simulate
the human brain and its synaptic connections through neural networks. However, due to a
limited understanding of the brain’s complex operations, these attempts were not successful.
Instead, Artificial Neural Networks (ANNs) were constructed using simpler, organized func-
tional units known as neurons or nodes. These neurons are interconnected through simulated
connections and organized into layers. The layers, which make up the neural network’s cen-
tral structure, extract and process information that is pertinent to the current circumstance.
3.2 DL (deep learning)
A sophisticated algorithm called DL (Deep learning) uses a complicated architecture that
closely matches the structure of the human brain to be able to independently learn from a
dataset. This technique finds applications in medical image analysis, where it aids in pre-
diction, diagnosis, and detection tasks. Compared to traditional methods, current neural
networks used for image recognition exhibit superior performance in skin lesion classi-
fication [2]. The initial step involves in receiving the image and understanding its class
characteristics. However, the way humans recognize images significantly differs from that
of machines, which perceive images as pixels. In deep learning, two prominent techniques,
namely ANN (Artificial Neural Networks) and CNN (Convolutional Neural Networks), are
widely utilized. These neural networks are specifically designed for image detection and
computer vision challenges [28]. Unlike traditional machine learning approaches, CNN uti-
lizes a sequence of convolution, pooling, and non-linear layers to process and pre-process
the image as a two-dimensional vector, ultimately producing output through the Fully Con-
nected Layer (FCN) (http:// www. datam ind. cz/ cz/ Sluzby- Data- Scien ce/ umela- intel ingen
ce- AI- ML- machi ne- learn ing- neural- net). The overall process of classification using DL is
illustrated in Fig.4. The underlying methodology of deep learning surpasses that of tradi-
tional machine learning algorithms in terms of effectiveness and performance.
Fig. 4 DL (deep learning) – classification process
Multimedia Tools and Applications
1 3
3.2.1 CNN
In the detection of melanoma, the choice of classification technique plays a crucial role,
and CNN (Convolutional Neural Network) has proven to be the appropriate architecture
for image recognition [53]. CNN consists of hidden layers that include convolutional, non-
linear pooling, and FC layers [53]. The convolutional layers perform multiple convolutions
and are connected to the FC layers [53]. However, the layer that deeply connected is not
suitable for melanoma detection as it can lead to overfitting, especially when the lesion
images lack distinct features [52]. To overcome the overfitting problem, CNN is well-suited
for identifying and classifying images based on different classes [52]. The architecture of
CNN can be visualized in Fig.5.
Convolution layers Convolutional layers in CNN have the ability to capture local pat-
terns, which leads to two significant advantages: they are translation-invariant, meaning the
learned patterns can be detected regardless of their position in the image, and they capture
spatial hierarchies of patterns. As the depth of the network increases, CNN becomes profi-
cient at detecting more complex visual concepts. In the convolutional process, these layers
apply a set of filters to the input image, performing convolutions and producing feature
maps that are then passed to subsequent layers for further processing. This hierarchical fea-
ture extraction allows CNN to effectively analyze and understand visual information.
Pooling layers CNN uses subsampling, also known as pooling, as a method to make fea-
ture maps less dimensional while preserving crucial data. It aids in reducing the network’s
computational burden and parameter count. Although these filters do not have trainable
parameters, they still perform convolution on the input feature maps. To downsample the
active maps by aggregating the values in each pooling window (e.g., taking the maximum
or average value) and selecting the most relevant features are possible in pooling. Down-
sampling decreases the spatial pixel values of the active maps while retaining important
attributes for subsequent layers.
Batch normalization layer These layers, known as normalization layers, apply a specific
function to normalize the input data. They are typically used in the forward propagation
stage of the network and do not have any trainable parameters. Their purpose is to ensure that
the input data is standardized and falls within a certain range, which can help improve the
training process. The usage of these normalization layers has been decreased nowadays, and
Fig. 5 CNN – basic architecture
Multimedia Tools and Applications
1 3
alternative techniques such as batch normalization have gained more popularity. Batch nor-
malization allows for normalization within mini-batches during training and has been shown
to be more effective in improving the training speed and stability of deep neural networks.
Regularization These layers, known as dropout layers, are used to combat overfitting in
neural networks. During each training session, a certain percentage of neurons in these
layers are randomly “dropped out” or ignored. This indicates that they don’t contribute to
the network’s forward or backward propagation because their outputs are momentarily set
to zero. By randomly dropping out neurons, the network becomes less reliant on specific
neurons or connections, and it compels the surviving neurons to acquire more reliable and
comprehensive data representations. A common regularisation method that helps avoid
overfitting is dropout by reducing the network’s tendency.
Fully connected layers Dense layer neurons, often referred to as the FC layer, are com-
bined with every neuron in the previous layer. This means that the activations from the pre-
vious layer serve as inputs to every neuron in the FC layer. The FC layer acts as a classifier
in a CNN architecture, taking the output feature maps from the previous convolutional or
pooling layers and transforming them into a form that is suitable for making predictions or
classifications. The last FC layer in the CNN is often referred to as the classifier because it
produces the final output of the network, to determine the predicted class.
3.2.2 Pre‑trained models
Pre-trained models are existing models developed by others to address similar problems. They
have already been trained on large datasets and learned to recognize visual patterns and features.
Using pre-trained models, training time and computational resources are reduced. Instead of build-
ing a model from initial level, we can adapt the trained model by fine-tuning or retraining only a
few layers to suit our specific problem. This approach benefits to speeds up the training process.
3.2.3 Transfer learning
TL is a powerful method that utilizes pretrained CNN models, such as those trained on Ima-
geNet, to achieve good results in a target domain. By reusing the learned weights and repre-
sentations, the model can adapt to a specific task with limited data. The approach involves
freezing the initial layers to preserve generic features and selectively fine-tuning specific lay-
ers to capture domain-specific characteristics. This enables efficient adaptation and improves
performance on the target task. To fine-tune a pre-trained model, you can follow these steps:
1. Load the pre-trained model and its weights.
2. Replace the last layers with new ones suitable for your task.
3. Freeze the initial layers to retain learned representations.
4. Train the modified model using your dataset, updating only the new layers.
5. Optionally, unfreeze and fine-tune some initial layers if your dataset is large and different.
6. Continue training to refine the model’s weights and capture dataset characteristics.
This approach allows you to leverage pre-trained weights and adapt the model to
your specific task, even with limited data. Experiment with different hyperparameters
for optimal results.
Multimedia Tools and Applications
1 3
3.2.4 Traditional learning vs deep learning
In traditional method, the initial steps involve data collection and preprocessing, where
the dataset is processed and relevant features are extracted [64]. Feature engineering
is a crucial part of this process, where domain expertise is applied to manually train
the dataset and identify important predictive features. This approach requires human
programming and can be time-consuming [103]. DL models, leverage predefined archi-
tectures and do not rely heavily on feature engineering. They can automatically learn
and make predictions without extensive manual intervention. This makes deep learning
more accurate and efficient compared to traditional machine learning methods [20]. Fig-
ure6 provides a visual representation of TL and DL.
4 Review onml anddl algorithms
In order to improve melanoma categorization, segmentation, and analysis, the litera-
ture review on skin lesion classification highlights the major application of ML and DL
approaches in dermoscopic imaging. DNN are trained on large datasets of benign and
malignant melanoma images with expert annotations. Several datasets, including ISIC
archive (2018, 2019, 2020) are commonly used in the research [99].
The reviewed articles cover various approaches, starting with those that primarily focus
on machine learning (ML) methods for skin cancer diagnosis. Subsequently, there are pub-
lications that specifically emphasize deep learning (DL) methods. Finally, there are publi-
cations that integrate methods with ML and DL to evaluate the models’ performance using
evaluation metrics. The performance indicators provide insights into accuracy and reliabil-
ity of the classification models [AUC].
Fig. 6 Traditional machine learning versus deep learning
Multimedia Tools and Applications
1 3
4.1 Machine learning strategies
In Yuan etal. [105], the author focused on classifying the benignity and malignancy of
skin lesions using texture data. The study employed a three-layer support vector machine
(SVM) mechanism, achieving an accuracy of 86.9%. The algorithm’s performance was
validated using 22 pairs of genuine clinical images.
The authors in [35] explored the use of color images for melanoma detection. The pro-
posed approach involved preprocessing steps such as background light correction and itera-
tive dilation-based noise removal. The generated image was then used as input for SVM
classification. Sequential minimal optimization was used to optimize SVM parameters.
The results indicated promising outcomes in accurately categorizing melanoma samples.
Babu et al. [13] focused efficient diagnosis of skin cancer using HOG features and
SVM. The ISIC 2018 dataset was preprocessed by converting it into a grayscale image
using median filters. HOG features were extracted from the preprocessed image and used
for SVM-based classification. The framework achieved a precision of 76%.
Adjed etal. [3] employed the LBP method for classification. LBP was used to extract
local texture information to create statistical characteristics that can differentiate between
melanoma and non-melanoma skin tissues. A database was prepared and trained based
on color and texture variables derived from the images. The SVM classification was per-
formed based on these variables, achieving an accuracy of 76.1%.
Fei etal. [33] utilized SVM on PSO (particle swarm optimization) to maintain accuracy.
The study utilized the PH2 dataset, which consisted of 200 images and 46 texture features
for full image analysis and classification.
These studies demonstrate various ML techniques, such as SVM, HOG features, LBP,
and PSO, for skin cancer classification, highlighting their potential in improving diagnostic
accuracy.
The comparative survey on detecting skin cancer using ML algorithms is shown in
Table4.
Preprocessing (PP), Feature Extraction (FE), Feature Selection (FS),Classification
(CLF), Segmentation (SEG), Support vector machine (SVM), K-nearest neighbor (KNN),
Random forest (RF), Decision Tree (DT), Naïve Bayes (NB),Binary Harris Hawks optimi-
zation algorithm (BHHO).
4.2 Deep learning strategies
CNN models have shown exceptional effectiveness in diagnosis in the realm of medical
imaging, classification, and segmentation tasks [39]. DeVries and Ramachandram [25]
developed a training framework using the ImageNet dataset, where a multi-scale CNN
architecture was employed. Their approach involved fine-tuning a pretrained v3 deep neu-
ral network using lesion images.
A technique for features extraction from skin lesions was put out by Mahbod etal. [67]
using pretrained models like AlexNet, ResNet-18, and VGG-16. An SVM classifier was
then trained using the retrieved characteristics. When tested against the 2017 ISIC dataset,
their model has a 97.55% accuracy rate.
Dorj etal. [27] utilized pretrained deep CNNs, particularly AlexNet, for categorizing
various types of skin lesion images.
Multimedia Tools and Applications
1 3
Table 4 Machine learning approaches on skin cancer detection
Ref Diagnosis Dataset Classifier Methodology Accuracy
[46] Two class Benign /Malignant HAM10000-2000 Samples SVM 1. PP-K-Means Clustering
2. SEG- LBP, HOG, BOVW
3. CLF-SVM
99.8%
[69] Two class Cancer/Non Cancer PH2SVM (Linear and Radial kernel
function)
1. PP-Gaussian Filter
2. SEG- Otsu ‘s thresholding
3. CLF-SVM (Linear and Radial Kernel
function)
92.30%
[31] Risk level -low, medium and high DermIS, DermQuest SVM, Neural classifiers 1. PP- Sharpening Filter, Hair removal using
Dull Razor
2. FE-
(i) Shape features extracted by ABCD rule
(ii) Texture features by GLCM, Coarseness
(iii) Color features extracted by Variance,
Entropy, Skewness
3. FS&CLF-SVM and Neural classifiers
Risk level
[87] Two way class Benign and Melanoma ISBI 2016 SVM,RF,KNN, NB 1. SEG-Dice coefficient and achieved
77.5%(i) CNN based U-Net algorithm
2. FE- LBP, EH, HOG &Gabor.
3. CLF- SVM, RF, KNN, NB
85.19%
[74] Benign and Melanoma ISIC SVM 1. PP-Median filtering (noise removal)
2. SEG&FE-DCT, HOG, Laws texture fea-
tures, Entropy Filter
3. CLF-SVM
84.86%
[76] Benign and Melanoma University Medical Center Gronin-
gen Department of Dermatology
SVM, DT, KNN, RF 1. SEG -Otsu’s thresholding, region growing,
compression based methods, K-Means
clustering.
2. FE-(i) Color Features by histogram; (ii)
Texture using Gray Level Co-occurrence
Matrix (iii) Border and asymmetry by
solidity
3. CLF-SVM, DT, KNN, RF
89.07%
Multimedia Tools and Applications
1 3
Table 4 (continued)
Ref Diagnosis Dataset Classifier Methodology Accuracy
[62] Melanoma and benign MedNode Ensembled Bagged Tree classifier 1. PP-Resize and De-noise by Gaussian Blur
2. SEG- Extraction of features from the lesion
3. FE-HSB + LBP
4.CLF- Ensembled Bagged Tree classifier
Accuracy = 95%,
Sensitivity = 94%,
Specificity = 97%,
AUC = 99%
[58] MEL, NEVI & SEK ISIC2017 KNN FE + KNN Accuracy = 68%,
Sensitivity = 80%,
Specificity =80%
[16] 7 class ISIC2019 LDA 1. PP-Explore the dataset(80% train and 20%
test)
2. CLF-SSTF statistical fractal signatures +
LDA classifier
4-classes: Accu-
racy = 87%Sensitiv-
ity = 63%,
Specificity = 89%,
Precision = 65%
7-classes: Accuracy = 88%,
Sensitivity = 41%,
Specificity = 92%,
Precision = 46%
[14] Two class
Melanoma, Non Melanoma
HAM10000 Linear SVM 1. PP-Increased brightness, addition of
contrast
2. FE&FS- BHHO-S algorithm
2.CLF-linear SVM
Accuracy = 88%,
Sensitivity = 89%,
Specificity = 89%
Precision = 86%
Multimedia Tools and Applications
1 3
In [97], the author introduced a methodology for enhancing contrast in skin lesion
images using a modified sigmoid transform based on the EfficientNetmodel. The segmen-
tation results, measured by the Dice Similarity Index (DSI), ranged from 68% to 81% using
Otsu’s thresholding.
Lastly, [57] discussed a common model for skin cancer diagnosis utilizing CNNs,
although further details about the specific approach and findings were not provided.
Overall, these studies demonstrate the effectiveness of CNN-based deep learning
approaches in skin lesion analysis and showcase their potential for improving diagnostic
accuracy in skin cancer detection and classification tasks.
The analysis of deep learning approaches is visualized in Table5.
4.3 Hybrid strategies
Transfer learning, which makes use of CNN, built on large datasets like ImageNet, is
the best option for skin cancer detection and classification [105]. By reusing the learned
weights and features from these models, transfer learning enables improved perfor-
mance and reduces training time. Fine-tuning the pre-trained models involves freezing
and unfreezing specific layers to adjust the hyperparameters and keep ready for target.
In the field of dermoscopic imaging, both techniques ML and DL are widely
employed. DL models, specifically deep neural networks (DNNs), are trained on anno-
tated datasets containing benign and malignant melanoma images. Various datasets
such as ISIC archive, HAM10000, PH2, Med Node, and DermIS are commonly used in
skin cancer research [99].
CNN models have shown impressive results in classification, and segmentation in medi-
cal imaging applications [39]. Researchers have developed models using architectures like
DenseNet121, VGG-16, AlexNet, ResNet-18, and EfficientNet, among others. These mod-
els are -trained on ImageNet and then fine-tuned for skin cancer tasks [25, 67].
Traditional ML approaches, such as SVM, have also been utilized for skin cancer
diagnosis. SVM-based methods are used for texture analysis, color extraction, and clas-
sification [18].
Ensemble models, which combine multiple sources of information such as images,
handcrafted features, and metadata for classification. These models use deep learning back-
bones like EfficientNet and employ techniques like feature extraction, followed by clas-
sification using multi-layer perceptrons (MLPs) or artificial neural networks (ANNs) [88].
In addition, studies have investigated DL models for threshold prediction and contrast
enhancement in skin lesion analysis. Techniques such as deep CNNs and modified sigmoid
transforms have been employed for accurate segmentation and improved visualization of
skin lesions [15, 21, 98].
Overall, the research in skin lesion classification encompasses various approaches,
including transfer learning, deep learning, traditional machine learning, ensemble models,
and specialized techniques for specific tasks like threshold prediction and contrast enhance-
ment. In(Table6) the hybrid approaches for skin cancer detection has been visualized.
Experimental analysis using DL models was done to find the effective model for
accurate classification. The analysis compared different DL models, such as VGG-16,
ResNet-18, and EfficientNet, to identify with efficiency. The results confirmed the superior
performance of DL models in accurately categorizing skin lesions based on various charac-
teristics and features. This highlights the importance of utilizing DL techniques for reliable
and effective skin cancer diagnosis.
Multimedia Tools and Applications
1 3
Table 5 Deep learning approaches on skin cancer detection
Ref Diagnosis Dataset Methodology Accuracy (%)
[104] 7 Class HAM10000 TL- InceptionV3, ResNet50 and Denset201 Transfer-
ring a pre-trained network’s parameters to CNN
Accuracy = 96%
[37] 7 class HAM10000 Custom CNN
1. PP-Augmentation of dataset
2. CLF-Dense-I and Dense-II (Custom CNN)
Accuracy = 96.3%, Recall = 96%,
F1-score = 95.7%
[75] 2 class ISIC archive Custom shallow CNN
Altering the brightness and contrast of images
Accuracy = 98.7%, Precision = 98.9%
[81] 7 class HAM10000 DL approach (Pretrained) Accuracy = 99.7%
[93] 2 class ISIC2020 VGG16(New model based on VGG16)
90, 180, and 270 degree image rotation; centre crop-
ping; brightness modification; and mirroring.
Accuracy = 87%, Sensitivity = 85.2%, F1
score = 92.2%, AUC = 92.3%
[26] 7 class HAM10000 Custom CNN-AlexNet Accuracy = 87.8%, Specificity= = 96.2%,
Precision = 78.7%, Recall = 77.4%, F1
score = 77.8%
[89] 2 class HAM10000 VGG16, VGG19, DenseNet101& ResNet101 Accuracy = 84.8%
[80] 2 class HAM10000, ISIC2019, ISIC2020 Custom CNN’son Pretrained network (VGG16,
ResNet50, DenseNet169 and EfficientNetB3)
Accuracy =98%, Specificity = 99.9%, Preci-
sion = 96.1%, Recall = 98%, F1 score = 97%,
AUC = 70.9%
[5] 2 class ISIC archive Zooming, vertical/horizontal flipping, shearing, and
rescaling
TL use on MobileNetV2, DenseNet201, XceptionNet,
and ResNet50
Accuracy =86.6%, Precision = 86.5%,
Recall = 86%, F1score = 86.2%
[68] 8 class ISIC2019 DNN with specific Accuracy =95.6%, Specificity = 96.3%,
Precision =84.7%, Recall = 92.5%,
F1score = 88.4%
[56] 2 class 2742 dermoscopic images from
ISIC dataset
Region-based CNN with ResNet152
Region-based CNN and masks was employed for the
region of interest extraction, and ResNet152 applied
for classification.
90.4%
82%- Sensitivity 92.5%-Specificity
Multimedia Tools and Applications
1 3
Table 5 (continued)
Ref Diagnosis Dataset Methodology Accuracy (%)
[30] Multi-class classification PH2, HAM10000 Resnet50, VGG16
FEX-Resnet50, VGG16 and Deeplabv3 + modified
LSTM
PH2: Accuracy =94%, Sensitivity = 94%,
Specificity = 93%, Precision = 90%, F1
score = 92%;
HAM: Accuracy = 94%, Sensitivity = 94%,
Specificity = 94%, Precision = 34%,
F1score = 50%
[61] Multi-class classification HAM10000 Custom CNN Accuracy = 87%, Sensitivity = 86%, Preci-
sion = 87%, F1score = 86%
[85] Multi-class ISIC2019, HAM10000 Fine-tuned models (Ensemble) (ResNext, SeResNext,
ResNet, Xception and DenseNet)
Accuracy = 87%, Precision = 87%,
Recall = 93%, F1score = 89%, MCC = 87%
[45] Multi-class: ISIC2017, ISIC2018, ISIC2019 Custom CNN
Horizontal/vertical flipping,12.5% and rotation (30 to
30 degrees)
ISIC2017:Accuracy = 93%, Sensitivity = 93%,
Specificity = 91%, Precision = 94%,
F1score = 93%, AUC = 96%
ISIC2018:Accuracy = 89%, Sensitivity = 89%,
Specificity = 96%, Precision = 90%,
F1score = 89%, AUC = 99%;
ISIC2019:Accuracy = 90%, Sensitivity = 90%,
Specificity = 98%, Precision = 91%,
F1score = 90%, AUC = 99%
[44] 2 class ISIC archive, ISIC2016, Med-
Node, PH2
ModifiedPre-trained MobileNetV2 ISIC archive: Accuracy = 85%, Sensitiv-
ity = 85%, Specificity = 85%, Preci-
sion = 83%
ISIC2016:Accuracy = 83%, Sensitiv-
ity = 36%, Specificity = 95%, Preci-
sion = 64%MedNode:Accuracy = 75%,
Sensitivity = 76%, Specificity = 73%, Preci-
sion = 67%PH2:
Accuracy = 72%,
Sensitivity = 33%,
Specificity = 92%,
Precision = 67%
Multimedia Tools and Applications
1 3
Table 5 (continued)
Ref Diagnosis Dataset Methodology Accuracy (%)
[40] 2 class PH2
MedNode
DermIS and Quest
ISIC archive (2016,2017 and
2018)
Refined Residual Deep Convolutional Neural Network PH2:
Accuracy = 94.97%,
Sensitivity = 92.32%,
Specificity = 96.21%
MedNode: Accuracy = 92%, Sensitiv-
ity = 91.99%
Specificity = 91.99%
[41] 7 class ISIC2018 Transfer learning with AlexNet Accuracy = 98.7%,
Sensitivity = 95.6%,
Specificity = 99.2%
Precision = 95%
[42] 3 class MED NODE
DermIS and Quest
ISIC2017
DCNN (AlexNet, ResNet101, GoogleNet) Modified GoogleNet
MED NODE Accuracy = 99.29%
DermIS and Quest
Accuracy = 99.15%
ISIC2017
Accuracy = 98.14%
2 class- Benign and Malignant
3 class - Melanoma, Seborrheic Keratosis, and Nevus
7 class-NEVI/AN (Atypical NEVI), COM NEVI (Common Nevi), MEL (Melanoma), SEK (Seborrheic keratosis), BACC (Basal cell carcinoma), DEF (Dermato fibroma),
ACK (Actinic keratosis), VASCL (Vascular lesion)
8 class-SQCC (Squamous cell carcinoma), NEVI/AN (Atypical NEVI), COM NEVI (Common Nevi), MEL (Melanoma), SEK (Seborrheic keratosis), BACC (Basal cell car-
cinoma), DEF (Dermato fibroma), ACK (Actinic keratosis), VASCL (Vascular lesion)
Multimedia Tools and Applications
1 3
Table 6 Hybrid approaches on skin cancer detection
Ref Diagnosis Dataset Methodology Accuracy
[88] 7 Class ISIC2018,ISIC2019 FE- hand-extracted features + ANN and B4 to
486 B7 EfficientNet networks
ISIC2018: Accuracy = 91%, Sensitivity = 98%,
ISIC2019:Accuracy = 86%, Sensitivity = 98%
[18] 2 class HAM10000 FE- pre-trained DenseNet121 + XGBoost Accuracy = 91%;
Sensitivity = 93%;
Precision = 91%;
F1- score = 91%
[15] 2 class HAM10000, PH2FE-from ResNet and 503 and hand crafted, Effi-
cientNet + ANN
HAM10000:
Accuracy = 95%,
Sensitivity = 95%,
Specificity = 95%,
Precision = 95%,
F1score = 95%;
PH2:
Accuracy = 98%,
Sensitivity = 98%,
Specificity = 98%,
Precision = 96%,
F1 score = 97%
[34] 2 class ISIC2016, ISICI2017,
HAM10000, PH2FE-Pre-trained DenseNet121 network + MPL ISIC2016:Accuracy = 81%ISICI2017:Accuracy = 81%; HAM10000:
Accuracy = 81%;
PH2:
Accuracy = 98%
Multimedia Tools and Applications
1 3
4.4 Experimental analysis anddiscussion
4.4.1 ML algorithms withHAM10000 dataset
The proposed study employed various ML models, including DT, RF, SVM, KNN,
LOR, and LDA. These models were evaluated using different evaluation metrics to
compare their performance. The hyper-parameters used by these ML algorithms are
provided in Table7 [90].
The accuracy of the model was analyzed using two approaches: with data augmentation
and without data augmentation. The findings are in Table8.
Random Forest approach outperforms other ML algorithms. Table9 displays exper-
imental outcomes for machine learning, including accuracy, precision, recall, and
F1-score [90].
4.4.2 Fine‑tuned pre‑trained model withHAM10000 dataset forskin lesion
classification
VGG‑19 The VGG-19 is a CNN architecture with 19 layers. The network is capable of
classifying 1000 different categories. The input images used for training are 224 × 224
pixels in size. The VGG-19 architecture is known for its simplicity, achieved by stacking
multiple 3 × 3 convolutional layers and reducing volume size using maximum pooling. A
softmax layer and two fully linked layers with 4096 nodes each follow the network. Convo-
lutional and completely linked layers represented by the “19” in VGG-19.
In the context of skin lesion classification, the VGG-19 model was applied to the
HAM10000 database. The dermoscopic images undergo preprocessing steps such as resiz-
ing, frame removal, and cropping of the skin lesion. The preprocessed images are then
input into the VGG-19 model for classification. The model is trained and validated using
evaluation metrics derived from the training dataset.
Table 7 Hyper-tuning parameters-ML classifiers
ML classifier Hyper tuning parameter
Linear discriminant analysis solver = ‘svd’
Logistic regression random_state = 9
K-nearest neighbor n_neighbors = 5
Decision tree estimators = 100
Random forest n_estimators = 200 and random_state = 0
Support vector machine kernel = linear,c = 1,and random_state = 0
Table 8 Accuracy –ML
classifiers Dataset Accuracy (%)
HAM10000 LDA LR KNN DT RF SVM
Without augmentation 35.18 48.92 41.78 44.28 58.57 42.14
With augmentation 57.58 58.12 48.39 68.66 87.32 53.12
Multimedia Tools and Applications
1 3
VGG‑16 CNN architecture that extensively employs convolution layers continued by pooling
layers throughout its structure. It culminates in 2 FC layers and a softmax, creating a seam-
less connection. The number “16” in VGG-16 represents the total weighted layers present in
the architecture. With a staggering 138 million parameters, the VGG-16 network is consid-
ered large and complex, capable of capturing intricate patterns and features in the data [91].
MobileNet model The MobileNet network is composed of 28 layers. By adjusting the
width multiplier hyper-parameter, the number of parameters in the standard MobileNet
architecture, which is around 4.2 million, can be further reduced. In classification, the skin
lesion image undergoes pre-processing to remove duplicates, and then it is processed with
MobileNet for classification [19, 84].
Xception The Xception architecture, also known as extreme inception, incorporates
filtering techniques that consider the cross-channel or depth dimension as well as the
spatial dimension. This involves scanning a 2 × 2 pixel patch across all RGB channels
using convolution filters similar to an image input layer. The initial input is projected
onto smaller input regions using 1 × 1 convolutions, and different filters are applied to
modify these smaller 3D data blocks. Spatial correlation maps are generated for each
output separately, followed by an 11-depth convolution to capture cross-channel cor-
relation. This approach separates the learning of cross-channel and spatial informa-
tion by depth, reducing computational costs and memory requirements. The Xception
architecture consists of sixteen hidden levels distributed among fourteen modules [63].
An ensemble model is used for classification of skin lesion by combining multiple
models for improved categorization. Figure7 illustrates the ensemble model consisting
of Xception, ResNet152, and InceptionV3 architectures for skin cancer detection [4].
The detailed model described above, which includes architectures such as VGG-19,
MobileNet, and Xception, is combined with the Skin Cancer dataset for the purpose of iden-
tifying a suitable pre-trained model for detection. By training and evaluating these pre-trained
models on the skin cancer dataset, their performance in accurately detecting and classifying
skin cancer can be assessed. This process helps in determining which model performs best
for skin cancer detection based on the specific characteristics and requirements of the dataset.
4.4.3 Performance metrics
In deep learning, common metrics like accuracy and loss are to evaluate the performance
of a model.
The model’s accuracy is gauged by how well it can diagnose the classes of data sam-
ples. Based on the values of TP, TN, FP, and FN, it is calculated.
Table 9 Evaluation Metrics-ML
model Evaluation (%) ML Model
RF DT LOR LDA SVM KNN
Accuracy 87 68 58 57 53 48
Precision 94 75 56 62 54 54
Recall 94 74 55 58 50 50
F1 Score 94 74 55 54 50 50
Multimedia Tools and Applications
1 3
An indicator, the model matches the training data is the training loss. It is determined by
adding the mistakes for each training set example. The training loss helps assess how the
model is learning and adjusting its weights and biases during the training process.
Similar to this, the validation loss is a statistic used to assess how well the model per-
forms on a different validation set. Additionally, the errors for each sample in the validation
set are added up to determine the validation loss.
Implications of loss Both the training loss and the validation loss give information about
how well the model generalises to new data and can show whether the training data are
being overfitted or underfitted by the model. Monitoring these data allows for performance
improvements in the model architecture, hyperparameters, and training procedure.
Confusion matrix Summary of a classification model’s performance by identifying the
count of instances of true positives (TP), false negatives (FN), false positives (FP), and true
negatives (TN). From the confusion matrix, important metrics can be derived.
The true positive rate (TPR), also known as recall and sensitivity measures, indicates
the percentage of the positive class that is correctly classified, and the formula has been
shown in Eq. (1) and followed by false negative rate (FNR),Specificity, false positive rate
(FPR) in Eqs. (2, 3, 4) respectively.
False negative rate, the proportion of the positive class got classified incorrectly.
The percentage of negative classes that are accurately identified allows one to determine
the specificity of a model.
(1)
Sensitivity∕TPR∕Recall =TP∕(TP +FN)
(2)
FNR =FN∕(TP +FN )
Fig. 7 Ensemble model (Xception, ResNet152, and inceptionv3)
Multimedia Tools and Applications
1 3
False positive rate, the proportion of negative classes got classified incorrectly.
The performance metrics for the classification also includes ROC and AUC curves.
ROC based on the false positive rate or specificity and true positive rate or sensitivity at
various thresholds of classification scores. General format for confusion matrix are shown
in(Fig.8).
4.4.4 Performance offine‑tuned pre‑trained model withHAM10000 dataset
Hyperparameters play a crucial role in fine-tuning pre-trained models. These parameters
are adjustable settings that influence the learning process and overall performance of the
model [101]. In Table10, relevant hyperparameters for the procedure are listed, specifying
the values chosen for each model.
Fine-tuning pre-trained models involves adjusting hyperparameters that are carefully
selected and tuned to achieve optimal performance for the specific task. The identified values
for these hyperparameters can significantly impact the model’s training process and final results.
Optimizer Optimizers are an essential component in deep learning for modifying a mod-
el’s parameters during training. Their main purpose is to optimize the model’s performance
by adjusting the weights to minimize the loss function [10]. In deep learning, various opti-
mizers can be used, each with its own algorithm and update rules. Some commonly used
optimizers include:
1. Adam: Adam has the advantages of RMSprop and momentum methods. It provides
effective optimisation by adjusting the learning rate based on gradients [96].
(3)
Specificity =TN∕(TN +FP)
(4)
FPR =FP∕(TN +FP)=1−Specificity
Fig. 8 Confusion matrix
Multimedia Tools and Applications
1 3
2. Momentum: Momentum optimizer uses a moving average of past gradients to accelerate
the learning process. It accumulates a velocity term that directs the update direction,
helping the optimizer to overcome local optima [96].
3. RMSprop (Root Mean Square Propagation): The learning rate is divided by a moving
average of the root mean square of previous gradients in RMSprop. It helps to stabilize
and normalize the learning process, especially in cases where the gradients vary sig-
nificantly [96].
These optimizers offer different strategies for updating the model’s weights during the
training process, allowing for efficient convergence and better generalization. The choice
of optimizer depends on the specific problem, data, and network architecture and it can
impact the training speed and final performance of the model.
Adam (adaptive moment learning rate) optimizer Adam combines the advantages of
two other optimization methods: momentum and RMSprop. Based on estimations for the
first and second moments of the gradients, it adjusts the learning rate for each parameter.
The optimizer converges more quickly and successfully manages various types of data and
model architectures because to its adjustable learning rate.
The update equation for Adam can be expressed as follows:
where:
θ(t) => parameter at iteration t,
α => learning rate,
m(t)=> estimate of the mean,
v(t)=>estimate of the variance,
ε=>a tiny constant to prevent division by zero.
Changing the learning rate in accordance with the gradient’s first and second moments,
Adam optimizes the weight updates, making it a widely used optimizer in deep learning.
The model’s total parameters (TtP), trainable (TrP) and non-trainable parameters (NTrP)
are shown in the Table11.
The HAM10000 dataset was used to train the fine-tuned pre-trained model, and the
accuracy and loss were calculated and displayed in the Table12.
𝜽
(t+1)=𝜽(t)−𝜶∗m(t)∕
(
sqrt(v(t)) +
𝜺
Table 10 Hyper tuning parameters for fine-tuned pretrained model [102]
Fine _tuned Pretrained Model Size of the input image Learning rate Optimizer Epoch Batch size
VGG19 (224,224,3) 0.0001 Adam 35 64
VGG16 (224,224,3) 0.0001 Adam 35 64
DenseNet (224,224,3) 0.0001 Adam 30 32
ResNet152 (224,224,3) 0.0001 Adam 35 64
InceptionV3 (224,224,3) 0.0001 Adam 35 64
InceptionResNetV2 (224,224,3) 0.0001 Adam 25 64
Xception (224,224,3) 0.0001 Adam 35 64
MobileNetV2 (224,224,3) 0.01 Adam 15 10
Multimedia Tools and Applications
1 3
Based on the findings presented, it can be concluded that MobileNetV2 and the ensem-
ble of DenseNet and InceptionV3 models have shown superior performance compared to
other models in terms of effectiveness. These models have demonstrated higher accuracy
or better classification results in the given context of skin lesion classification or detection.
The use of MobileNetV2 and the ensemble of DenseNet and InceptionV3 can be recom-
mended as effective choices for skin lesion analysis tasks.
4.4.5 CNN (convolutional neural network) MODEL withHAM10000 dataset
The CNN architecture consists of several key components:
1. Convolution Layer: This layer uses filters or kernels to extract features from the input
image through a process called convolution. Each filter detects specific patterns or fea-
tures in the image, resulting in feature maps [78].
Table 11 Trainable parameters for fine-tuned pre-trained model
Fine _tuned pretrained model Total parameters (TtP) Trainable param-
eters (TrP)
Non train-
able parameters
(NTrP)
VGG19 20,290,631 20,290,631 0
VGG16 14,980,935 7,345,671 7,635,264
DenseNet 19,309,127 8,097,479 11,211,648
ResNet152 59,423,623 59,272,199 1,51,424
InceptionV3 22,855,463 22,821,031 34,432
InceptionResNetV2 55,127,271 24,352,647 30,774,674
Xception 21,914,159 21,859,631 54,528
MobileNetV2 3,236,039 3,214,151 21,888
Table 12 Performance on fine-tuned pre-trained model
Fine _tuned pre-trained model Training Validation Testing
Accuracy (%) Loss Accuracy (%) Loss Accuracy (%) Loss
VGG19 86.93 0.3617 84.38 0.492 82.87 0.520
VGG16 88.92 0.3140 79.46 0.602 79.82 0.650
DenseNet 98.00 0.0580 85.86 0.579 83.93 0.691
ResNet152 95.83 0.1080 80.21 0.870 81.77 0.880
InceptionV3 98.04 0.0599 85.42 0.655 83.97 0.690
InceptionResNetV2 93.32 0.1898 81.75 0.617 82.53 0.640
Xception 98.60 0.0392 86.72 0.676 84.57 0.800
MobileNetV2 94.22 0.1936 88.81 0.331 87.71 0.350
Ensemble (DenseNet, InceptionV3) –88.80 0.399 88.52 0.411
Ensemble (Xception,ResNet152,
InceptionV3)
– 87.53 0.443 85.92 0.487
Multimedia Tools and Applications
1 3
2. Kernel or Filter: The kernel is an iterative matrix that performs dot product operations
with subregions of the input data. It helps extract image features and is typically of size
3 × 3 or 5 × 5. Different kernels are applied to capture local features in the image [78].
3. Padding: Padding is used to preserve data at the borders of the activation maps during
convolution. It ensures that the spatial dimensions are maintained and helps in building
deeper networks. Zero-padding is commonly used for its simplicity and effectiveness [78].
4. Stride: The stride determines how many pixels the kernel moves with each operation. It
affects the output layer dimensions and feature extraction. A stride of 1 means the kernel
moves one pixel at a time. Increasing the stride value reduces the output dimensions and
restricts feature extraction [78].
5. ReLU Layer: The Rectified Linear Unit (ReLU) layer applies an element-wise activation
function to introduce non-linearity. It replaces negative activations with zero, enhancing
the features of the image [60].
6. Fully Connected Layer: This layer takes the output of the convolutional process and
performs classification based on the extracted features. It calculates class scores for each
image and assigns the predicted class based on the highest probability score [60].
In addition to these layers, CNNs may also include other components such as dropout
layers and activation functions to further enhance performance. The CNN architecture pro-
cesses the pre-processed image and classifies it into different categories. The classification
process involves both pre-processing and classification steps. The overall workings of the
CNN model for skin lesion classification are depicted in Fig.9 [60].
The experimental analysis involved adjusting the number of convolution layers and
fine-tuning hyperparameters to evaluate the CNN’s performance. The model’s hyperpa-
rameters were tuned to optimize its performance. The model was trained for a total of
25 epochs, as indicated in Table13. Initially, the model was trained with 50 epochs and
a learning rate of 0.001. After 16 epochs, the validation accuracy started to decrease.
At epoch 20, the training process was stopped, and the learning rate was reduced as it
reached the patience value. The model then continued training with the reduced learn-
ing rate and achieved its best validation accuracy at epoch 22. However, the model’s
Fig. 9 Skin lesion classification using CNN Model [78]
Multimedia Tools and Applications
1 3
accuracy gradually declined, and early stopping was enabled when no further improve-
ment was observed. The model weights were restored to the best epoch, and the training
process concluded at epoch 25. Table13 provides details of the hyperparameters used
for hyper-tuning the CNN model.
After training the model with the chosen hyperparameters, the accuracy and loss values
were determined and visualized in Table14. This table presents the accuracy and loss met-
rics obtained during the training process, providing insights into the model’s performance
on the dataset.
The results indicate that when using the HAM10000 raw dataset, the CNN model’s
validation accuracy does not exceed 80%. Additionally, the models with 7 and 5 layers
exhibit overfitting, as evidenced by the significant difference between the training loss and
validation loss. However, the CNN model with 3 layers performs better in terms of avoid-
ing overfitting. It achieves a more balanced performance between training and validation,
resulting in improved accuracy on the raw dataset.
In the context of skin cancer classification, dealing with imbalanced datasets is cru-
cial for accurate model training. Skin cancer datasets often exhibit class imbalance,
where certain types of skin lesions are more prevalent than others. This class imbalance
can lead to biased models with lower accuracy on minority classes.
To solve this problem, particular methods can be used to balance the dataset
and enhance the model’s performance. One approach is data augmentation, which
involves generating new synthetic samples by applying various transformations to
the existing data.
Impact of data augmentation Techniques for image augmentation, such as rotation,
scaling, flipping, and brightness or contrast adjustments, can be used to classify skin can-
cer. These modifications add more diverse samples and aid in balancing the distribution of
classes. One effective method is random rotations, which introduce variations in the ori-
entation of the images and help the model generalize better to different angles and view-
points [12].
Two common approaches are used to fit the input size of the augmented data:
Table 13 Hyper tuning and trainable parameters of CNN models
CNN model Total param-
eters
Trainable
parameters
Non -train-
able param-
eters
Hyper tuning parameters
Epochs Batch size Lr-rate Optimizer
CNN(7 lay-
ered)
1,275,079 1,273,671 1408 25 128 0.001 Adam
CNN(5
layered) with
Batch Nor-
malization
6,19,207 6,17,799 1408 25 128 0.001 Adam
CNN(5 lay-
ered) without
Batch Nor-
malization
4,11,400 4,11,400 0 25 64 0.001 Adam
CNN(3 lay-
ered)
2,124,839 2,124,839 0 30 64 0.001 Adam
Multimedia Tools and Applications
1 3
Table 14 Performance on CNN models
CNN model Size of input image Training Validation Testing
Accuracy (%) Loss Accuracy (%) Loss Accuracy (%) Loss
CNN(7 layered) with Batch Normalization (28,28,3) 99.95 0.0082 77.92 0.7192 76.94 0.8204
CNN(5 layered) with Batch Normalization (28,28,3) 98.25 0.0807 77.22 0.8191 76.55 0.8508
CNN(5 layered) without Batch Normalization (28,28,3) 91.22 0.2572 73.80 0.8804 75.88 0.870
CNN(3 layered) (64,64,3)79.85 0.5263 77.48 0.6034 76.54 0.6466
Multimedia Tools and Applications
1 3
(i) Scaling: The resolution of the images is multiplied by a scaling factor to resize them.
This method allows for adjustments in the overall size of the images while preserving
their spatial content [12].
(ii) Cropping: This approach involves extracting sub-regions from the images. The cropping
can be done by selecting either the center portion or a random area of the image. This
technique helps maintain the important features of the image while reducing its size [77].
By applying data augmentation techniques, the training process benefits from a larger
and more diverse dataset. This helps address the issue of overfitting. The augmented data-
set provides the model with a wider range of examples, enabling it to learn more robust and
generalizable representations of the skin lesion classes.
Performance of fine‑tuned pre‑trained and CNN model after augmentation Tables15
and 16 shows the impact of augmentation in the model, the accuracy of the model was
improved when compared to the raw dataset.
The table illustrates the improvement in model performance in terms of accuracy when
using data augmentation techniques on an imbalanced dataset. These results highlight
the significance of augmenting the dataset to address class imbalances. The experimental
analysis demonstrates that CNN models with different layer sequences perform well when
trained on a balanced dataset.
Among the models evaluated, the fine-tuned pre-trained model MobileNetV2 shows
efficient performance compared to other models. Additionally, when using a custom CNN
model, CNN architectures with 3 or 5 layers yield favorable results compared to other CNN
configurations. These findings are consistent with previous studies conducted using similar
models, validating the effectiveness of the proposed approach.
Comparing results with literature Table 17 and Fig. 10 show the comparative results
with previous experiments.
Advantages of DLTechniques Deep learning models, including pre-trained models, trans-
fer learning models, and CNN-based models, offer several advantages in image recognition
tasks. They can automatically extract pertinent characteristics from images. Transfer learn-
ing allows leveraging pre-trained models for specific tasks, reducing the need for extensive
training. Deep learning models achieve high accuracy and can process large amounts of
data efficiently, making them suitable for image recognition. They also enable end-to-end
learning, simplifying the workflow. Overall, deep learning has revolutionized image recog-
nition with its efficient and effective techniques.
5 Challenges
5.1 Extensive background
Challenges in Employing Neural Networks for Skin Cancer Diagnosis
1. Analysis and Comprehension of Dermoscopic Images: One of the major difficulties in
using neural networks for skin cancer diagnosis is the task of analyzing and comprehend-
ing the characteristics of dermoscopic images. This process requires significant time
and effort.
Multimedia Tools and Applications
1 3
Table 15 Performance of fine-tuned pre-trained model after augmentation
Fine _tuned pretrained model with augmentation Training Validation Testing
Accuracy (%) Loss Accuracy (%) Loss Accuracy (%) Loss
VGG19 87.90 0.3285 87.76 0.4084 87.76 0.4083
VGG16 87.31 0.3448 86.51 0.4548 86.50 0.4548
DenseNet 99.87 0.0450 96.82 0.0220 96.80 0.0225
ResNet152 99.90 0.0120 96.71 0.0240 96.71 0.0241
InceptionV3 99.90 0.0110 97.01 0.0230 97.00 0.0229
InceptionResNetV2 97.32 0.0298 91.75 0.1170 91.55 0.1260
Xception 99.60 0.0192 96.72 0.0667 95.27 0.0567
MobileNetV2 97.99 0.0671 97.58 0.0551 97.28 0.0591
Ensemble (DenseNet, InceptionV3) –97.50 0.0561 97.02 0.0610
Ensemble (Xception,ResNet152, InceptionV3) – 96.53 0.0340 95.92 0.0420
Multimedia Tools and Applications
1 3
Table 16 Performance of CNN Model after augmentation
CNN model with augmentation Size of the input
image
Training Validation Testing
Accuracy (%) Loss Accuracy (%) Loss Accuracy (%) Loss
CNN(7 layered) with Batch Normalization (28,28,3) 99.98 0.0047 91.38 0.4199 92.08 0.3631
CNN(5 layered) with Batch Normalization (28,28,3) 99.56 0.0385 89.62 0.3734 89.38 0.3930
CNN(5 layered) without Batch Normalization (28,28,3) 100 0.0002 97.72 0.1313 98.04 0.1100
CNN(3 layered) (28,28,3) 96.83 0.0895 98.02 0.240 97.35 0.280
Multimedia Tools and Applications
1 3
2. Variation in Lesion Size: The variation in lesion size poses another challenge in skin
cancer identification. The accuracy of diagnosis can vary depending on the size of the
lesion. Smaller lesions, particularly those measuring 1mm or 2mm, make diagnosis
more challenging and prone to errors.
3. Unbalanced Datasets: Unbalanced datasets pose challenges in training neural networks
for skin cancer diagnosis. These datasets often have a highly uneven distribution of pho-
tos for different types of skin cancers. While there may be numerous images of common
carcinomas, there are only a few images of rare forms, making it difficult to generalize
visual characteristics from dermoscopic images.
4. Analysis of Genetic and Environmental Factors: When combined with environmental
hazards like excessive UV exposure, these genetic factors significantly increase the like-
lihood of cancer. Incorporating these factors into deep learning algorithms can improve
their performance [92].
6 Discussion, conclusion andfuture enhancement
To identify the course of therapy and increase the likelihood that a patient will survive,
early detection and proper diagnosis of skin lesions are essential, especially insituations of
malignant lesions. However, manual diagnosis by dermatologists can be time-consuming
and challenging. To simplify and expedite this process, Computer-Aided Diagnosis (CAD)
technologies have been employed.
Table 17 Comparative of results
with previous experiments HAM 10000 Dataset Model Accuracy
Balanced Fine- tuned MobileNetV2 97.58
Not Balanced 88.81
Not Balanced Indraswari, R.et al. [44] 85.00
Balanced Chaturvedi, S. et. al [19] 95.84
Balanced Customized CNN 98.02
Not Balanced 74.17
Balanced Shetty, B.et. al [90] 95.18
Balanced Nugroho, A. et. al [79] 80.00
Fig. 10 MobileNetV2 and CNN with previous study
Multimedia Tools and Applications
1 3
This review aims to provide an overview of the application of deep learning (DL)
and machine learning (ML) in dermatology. Academic works focusing on ML and DL
methods for identifying and classifying skin lesions were selected from publications
between 2012 and 2022. The search was conducted in databases such as arXiv and
Science Direct, resulting in the inclusion of approximately 70 articles. These articles
utilized photos from open datasets and reported the model performance in terms of
classification accuracy.
It is worth noting that no publications were included before 2016, as the inclusion cri-
teria required the use of public datasets. Furthermore, over 50% of the articles on DL and
ML/DL were published from 2021 onwards, indicating a recent surge in research activity
in this field. Additionally, within the selected articles, around 70% were published within
the last two years, highlighting the increasing interest and advancements in DL and ML for
skin lesion classification.
6.1 Dataset analysis
According to the study mentioned in this review article, the datasets most frequently used
for training and testing classification models in the cited studies were HAM10000 and ISIC
archive. These datasets have been widely utilized in various research papers to develop
and evaluate skin lesion classification models. The availability and quality of these datasets
Fig. 11 Dataset analysis. Others–DermQuest, DermIs, 7 point checklist, DermNZ, and IDS datasets
Fig. 12 ML model analysis
Multimedia Tools and Applications
1 3
make them popular choices for training deep learning and machine learning algorithms in
the field of dermatology. The analysis is visualized in Fig.11.
6.2 Algorithm analysis
According to the study presented in the review article, pre-trained convolutional neural
networks (CNNs) were found to be the dominant deep learning (DL) and machine learn-
ing with DL (ML/DL) technique for skin lesion classification (Fig.12). These pre-trained
CNN models have shown excellent performance in extracting meaningful features from
skin lesion images and achieving high accuracy in classification tasks.
On the other hand, random forest (RF) models emerged as the most popular machine
learning (ML) classifiers in the studies (Fig.13). RF models have been widely utilized in
skin lesion classification tasks, showcasing their effectiveness in handling structured data
and providing good predictive performance.
However, among the various solutions examined, DL-based approaches, particularly
deep CNNs, were found to be prevalent in addressing the categorization and diagnosis of
skin lesions. The ability of deep CNNs to automatically learn intricate patterns and repre-
sentations from images contributes to their potential for increasing the efficiency and reli-
ability of skin lesion classification and diagnosis tasks.
The review conducted in this study explored different ML and DL algorithms for the diag-
nosis and classification of skin lesions. The study focused on ML techniques and deep learning
strategies that combined CNN with other pre-trained models to categorize skin lesion images.
Every algorithm has benefits and limitations, and selecting the appropriate categori-
zation technique is crucial for achieving optimal results. However, the review findings
suggest that CNN-based approaches, particularly when integrated with pre-trained mod-
els, tend to outperform ML methods in terms of accuracy and effectiveness. DL algo-
rithms, including CNN, show promise in delivering more robust and reliable results.
In the future, further research and exploration of different DL algorithms will be con-
ducted to enhance the efficiency and performance of the models. Specific focus can be
given to key stages such as segmentation and classification, where advancements in
deep learning techniques can lead to improved outcomes in skin lesion diagnosis and
classification.
Fig. 13 DL model analysis
Multimedia Tools and Applications
1 3
References
1. Addi - automatic computer-based diagnosis system for dermoscopy images (n.d.) Retrieved January
7, 2023, from https:// www. fc. up. pt/ addi/
2. Adegun AA, Viriri S (2020) FCN-based DenseNet framework for automated detection and classifica-
tion of skin lesions in dermoscopy images. IEEE Access 8:150377–150396. https:// doi. org/ 10. 1109/
access. 2020. 30166 51
3. Adjed F, Faye I, Ababsa F, Gardezi SJ, Dass SC (2016) Classification of skin cancer images using
local binary pattern and SVM classifier. AIP Conf Proc 10(1063/1):4968145
4. Admisysrg, Admisysrg, 23, P. R. L. A., 30, admisysrg Author O, Author, & 30, celebrities A. (2020).
Skin lesion classification based on deep ensemble Convolutional Neural Network. ISYSRG. https://
isysrg. com/ 2020/ 06/ 17/ deep- ensem ble- learn ing- for- skin- lesio ns- class ifica tion- with- convo lutio nal-
neural- netwo rk/
5. Agarwal K, Singh T (2022) Classification of skin cancer images using convolutional neural networks.
SSRN Electronic J. https:// doi. org/ 10. 2139/ ssrn. 40550 37
6. Al-Issa Y, Alqudah AM (2022) A lightweight hybrid deep learning system for cardiac valvular dis-
ease classification. Sci Rep 12(1). https:// doi. org/ 10. 1038/ s41598- 022- 18,293-7
7. Alqudah AM, Algharib HM, Algharib AM, Algharib HM (2019) Computer aided diagnosis system
for automatic two stages classification of Breast Mass in digital mammogram images. Biomed Eng
Appl Basis Commun 31(01):1950007. https:// doi. org/ 10. 4015/ s1016 23721 95000 78
8. Alqudah AM, Alquran H, Qasmieh IA (2020) Classification of heart sound short records using
Bispectrum analysis approach images and deep learning. Netw Model Anal Health Inform Bioinform
9(1). https:// doi. org/ 10. 1007/ s13721- 020- 00272-5
9. Alqudah AM, Qazan S, Al-Ebbini L, Alquran H, Qasmieh IA (2021) ECG heartbeat arrhythmias
classification: a comparison study between different types of spectrum representation and convolu-
tional neural networks architectures. J Ambient Intell Humaniz Comput 13(10):4877–4907. https://
doi. org/ 10. 1007/ s12652- 021- 03247-0
10. Alqudah A, Alqudah AM, Alquran H, Al-Zoubi HR, Al-Qodah M, Al-Khassaweneh MA (2021) Rec-
ognition of handwritten Arabic and Hindi numerals using convolutional neural networks. Appl Sci
11(4):1573. https:// doi. org/ 10. 3390/ app11 041573
11. Altman NS (1992) An introduction to kernel and nearest-neighbor nonparametric regression. Am Stat
46(3):175. https:// doi. org/ 10. 2307/ 26852 09
12. Auxilia Osvin Nancy V, Arya MS, Shamreen Ahamed B (2022) Automated seven-level skin can-
cer staging diagnosis in dermoscopic images using Deep Learning. Machine Learning and Artificial
Intelligence in Healthcare Systems, 93–109. https:// doi. org/ 10. 1201/ 97810 03265 436-4
13. Babu GNK, Peter VJ (2021) Skin cancer detection using support vector machine with histogram of
oriented gradients features. ICTACT J Soft Comput 11(02):2229–6956 (Online). https:// doi. org/ 10.
21917/ Ijsc. 2021. 0329
14. Bansal P, Vanjani A, Mehta A, Kavitha JC, Kumar S (2022) Improving the classification accuracy of
melanoma detection by performing feature selection using Binary Harris Hawks optimization algo-
rithm. Soft Comput 26(17):8163–8181. https:// doi. org/ 10. 1007/ s00500- 022- 07234-1
15. Bansal P, Garg R, Soni P (2022) Detection of melanoma in dermoscopic images by integrating fea-
tures extracted using handcrafted and deep learning models. Comput Ind Eng 168(108):060. https://
doi. org/ 10. 1016/j. cie. 2022. 108060
16. Camacho-Gutiérrez JA, Solorza-Calderón S, Álvarez-Borrego J (2022) Multi-class skin lesion clas-
sification using prism- and segmentation-based fractal signatures. Expert Syst Appl 197(116):671.
https:// doi. org/ 10. 1016/j. eswa. 2022. 116671
17. Cancer Facts and Figs.2021. [online]. Available: https:// www. cancer. org/ resea rch/ cancer- facts- stati
stics/ all- cancer- facts- figur es/ cancer- facts- figur es- 2021. html. Accessed 28 Apr 2021
18. Car vajal DC, Delgado BM, Ibarra DG, Ariza LC (2022) Skin cancer classification in dermatological images
based on a dense hybrid algorithm. In: 2022 IEEE XXIX international conference on electronics, electri-
cal engineering and computing (INTERCON). https:// doi. org/ 10. 1109/ inter con55 795. 2022. 98701 29
19. Chaturvedi SS, Gupta K, Prasad PS (2020) Skin lesion analyser: An efficient seven-way multi-class
skin cancer classification using MobileNet. Adv Intell Syst Comput 165–176. https:// doi. org/ 10. 1007/
978- 981- 15- 3383-9_ 15
20. Chen X, Yuan H, Li. (2019) Research on a real-time monitoring method for the Wear State of a tool based on a
convolutional bidirectional LSTM model. Symmetry 11(10):1233. https:// doi. org/ 10. 3390/ sym11 101233
21. Codella NC, Nguyen Q-B, Pankanti S, Gutman DA, Helba B, Halpern AC, Smith JR (2017) Deep
learning ensembles for melanoma recognition in dermoscopy images. IBM J Res Dev 61(4/5). https://
doi. org/ 10. 1147/ jrd. 2017. 27082 99
Multimedia Tools and Applications
1 3
22. Codella NC, Gutman D, Celebi ME, Helba B, Marchetti MA, Dusza SW, Kalloo A, Liopyris K,
Mishra N, Kittler H, Halpern A (2018) Skin lesion analysis toward melanoma detection: a challenge
at the 2017 International Symposium on Biomedical Imaging (ISBI), hosted by the International Skin
Imaging Collaboration (ISIC). In: 2018 IEEE 15th international symposium on biomedical imaging
(ISBI 2018). https:// doi. org/ 10. 1109/ isbi. 2018. 83635 47
23. Cristianini N, Shawe-Taylor J (2000) An introduction to support vector machines and other kernel-
based learning methods. Apple Books. Retrieved January 7, 2023, from https:// books. apple. com/ us/
book/ an- intro ducti on- to- suppo rt- vector- machi nes- and/ id811 960631? l= vi
24. de Ville B (2013) Decision trees. Wiley Interdiscip Rev Comput Stat 5(6):448–455. https:// doi. org/
10. 1002/ wics. 1278
25. DeVries T, Ramachandram D Skin lesion classification using deep multi-scale convolutional neu-
ral networks. arXiv 2017, arXiv:1703.01402. Available online: http:// arxiv. org/ abs/ 1703. 01402.
Accessed 6 Sept 2021
26. Diwan T, Shukla R, Ghuse E, Tembhurne JV (2022) Model hybridization & learning rate anneal-
ing for skin cancer detection. Multimed Tools Appl 82(2):2369–2392. https:// doi. org/ 10. 1007/
s11042- 022- 12,633-5
27. Dorj U-O, Lee K-K, Choi J-Y, Lee M (2018) The skin cancer classification using deep convolutional
neural network. Multimed Tools Appl 77(8):9909–9924. https:// doi. org/ 10. 1007/ s11042- 018- 5714-1
28. dshahid380 (2019) Convolutional Neural Network. Medium. Retrieved January 7, 2023, from https://
towar dsdat ascie nce. com/ covol ution al- neural- netwo rk- cb088 3dd65 29
29. Duarte AF, Sousa-Pinto B, Azevedo LF, Barros AM, Puig S, Malvehy J, Haneke E, Correia O (2021)
Clinical ABCDE rule for early melanoma detection. Eur J Dermatol 31(6):771–778. https:// doi. org/
10. 1684/ ejd. 2021. 4171
30. Elashiri MA, Rajesh A, Nath Pandey S, Kumar Shukla S, Urooj S, Lay-Ekuakille A (2022) Ensemble of
weighted deep concatenated features for the skin disease classification model using modified Long short
term memory. Biomed Signal Process Control 76(103):729. https:// doi. org/ 10. 1016/j. bspc. 2022. 103729
31. Farooq MA, Azhar MA, Raza RH (2016) Automatic lesion detection system (ALDS) for skin cancer
classification using SVM and neural classifiers. In: 2016 IEEE 16th international conference on bioin-
formatics and bioengineering (BIBE). https:// doi. org/ 10. 1109/ bibe. 2016. 53
32. Feature extraction and expression classification using histogram of oriented gradients (hog) and sup-
port vector machine (SVM). (2020) Strad Res 7(9):10.37896/sr7.9/003
33. Fei D-Y, Almasiri O, Rafig A (2020) Skin cancer detection using support vector machine learn-
ing classification based on particle swarm optimization capabilities. Trans Mach Learn Artif Intell
8(4):01–13. https:// doi. org/ 10. 14738/ tmlai. 84. 8415
34. Gajera HK, Nayak DR, Zaveri MA (2023) A comprehensive analysis of dermoscopy images for mela-
noma detection via deep CNN features. Biomed Signal Process Control 79(104):186. https:// doi. org/
10. 1016/j. bspc. 2022. 104186
35. Gautam D, Ahmed M (2015) Melanoma detection and classification using SVM based decision support sys-
tem. In: 2015 annual IEEE India conference (INDICON). https:// doi. org/ 10. 1109/ indic on. 2015. 74434 47
36. Giotis I, Molders N, Land S, Biehl M, Jonkman MF, Petkov N (2015) Med-Node: A computer-
assisted melanoma diagnosis system using non-dermoscopic images. Expert Syst Appl 42(19):6578–
6585. https:// doi. org/ 10. 1016/j. eswa. 2015. 04. 034
37. Girdhar N, Sinha A, Gupta S (2022) DenseNet-II: an improved deep convolutional neural network for
melanoma cancer detection. Soft Comput. https:// doi. org/ 10. 1007/ s00500- 022- 07406-z
38. Grignaffini F, Barbuto F, Piazzo L, Troiano M, Simeoni P, Mangini F, Pellacani G, Cantisani C,
Frezza F (2022) Machine learning approaches for skin cancer classification from Dermoscopic
Images: a systematic review. Algorithms 15(11):438. https:// doi. org/ 10. 3390/ a1511 0438
39. Harley AW (2015) An interactive node-link visualization of Convolutional Neural Networks.
Advances in Visual Computing, 867–877. https:// doi. org/ 10. 1007/ 978-3- 319- 27,857-5_ 77
40. Hosny KM, Kassem MA (2022) Refined residual deep convolutional network for skin lesion classifi-
cation. J Digit Imaging 35(2):258–280. https:// doi. org/ 10. 1007/ s10278- 021- 00552-0
41. Hosny KM, Kassem MA, Fouad MM (2020) Classification of skin lesions into seven classes
using transfer learning with AlexNet. J Digit Imaging 33(5):1325–1334. https:// doi. org/ 10. 1007/
s10278- 020- 00371-9
42. Hosny KM, Kassem MA, Foaud MM (2020) Skin melanoma classification using ROI and data aug-
mentation with deep convolutional Neural Networks. Multimed Tools Appl 79(33-34):24029–24055.
https:// doi. org/ 10. 1007/ s11042- 020- 09067-2
43. Hussaindeen A, Iqbal S, Ambegoda TD (2022) Multi-label prototype based interpretable machine
learning for melanoma detection. Int J Adv Signal Image Sci 8(1):40–53. https:// doi. org/ 10. 29284/
ijasis. 8.1. 2022. 40- 53
Multimedia Tools and Applications
1 3
44. Indraswari R, Rokhana R, Herulambang W (2022) Melanoma image classification based on Mobile-
NetV2 Network. Proc Comput Sci 197:198–207. https:// doi. org/ 10. 1016/j. procs. 2021. 12. 132
45. Iqbal I, Younus M, Walayat K, Kakar MU, Ma J (2021) Automated multi-class classification of skin
lesions through deep convolutional neural network with Dermoscopic Images. Comput Med Imaging
Graph 88(101):843. https:// doi. org/ 10. 1016/j. compm edimag. 2020. 101843
46. Iqtidar K, Iqtidar A, Ali W, Aziz S, Khan MU (2020) Image pattern analysis towards classification of
skin cancer through dermoscopic images. In: 2020 first international conference of smart systems and
emerging technologies (SMARTTECH). https:// doi. org/ 10. 1109/ smart- tech4 9988. 2020. 00055
47. ISIC Archive. Available online: https:// isic- archi ve. com/
48. Isic Challenge (n.d.) Retrieved January 7, 2023, from https:// chall enge. isica rchive. com/ data/# 2016
49. Isic Challenge. (n.d.) Retrieved January 7, 2023, from https:// chall enge. isic- archi ve. com/ data/# 2017
50. Isic Challenge. (n.d.) Retrieved January 7, 2023, from https:// chall enge. isic- archi ve. com/ data/# 2019
51. Isic Challenge. (n.d.) Retrieved January 7, 2023, from https:// chall enge. isic- archi ve. com/ data/# 2020
52. Jafari MH, Karimi N, Nasr-Esfahani E, Samavi S, Soroushmehr SMR, Ward K, Najarian K (2016)
Skin lesion segmentation in clinical images using Deep Learning. In: 2016 23rd international confer-
ence on pattern recognition (ICPR). https:// doi. org/ 10. 1109/ icpr. 2016. 78996 56
53. Jafari MH, Nasr-Esfahani E, Karimi N, Soroushmehr SM, Samavi S, Najarian K (2017) Extraction
of skin lesions from non-dermoscopic images for surgical excision of melanoma. Int J Comput Assist
Radiol Surg 12(6):1021–1030. https:// doi. org/ 10. 1007/ s11548- 017- 1567-8
54. Jaisakthi SM, Devikirubha B (2021) Role of deep learning techniques in detecting skin cancer: a
Review. Handbook of Deep Learning in Biomedical Engineering and Health Informatics, pp 253–
279. https:// doi. org/ 10. 1201/ 97810 03144 694- 10
55. Jaworek-Korjakowska J, Kleczek P, Gorgon M (2019) Melanoma thickness prediction based on
convolutional neural network with VGG-19 model transfer learning. In: 2019 IEEE/CVF confer-
ence on computer vision and pattern recognition workshops (CVPRW). https:// doi. org/ 10. 1109/
cvprw. 2019. 00333
56. Jojoa Acosta MF, Caballero Tovar LY, Garcia-Zapirain MB, Percybrooks WS (2021) Melanoma
diagnosis using deep learning techniques on dermatoscopic images. BMC Med Imaging 21(1).
https:// doi. org/ 10. 1186/ s12880- 020- 00534-8
57. Kalouche S (2016) Vision-based classification of skin cancer using deep learning. Available online:
https:// www. seman ticsc holar. org/ paper/ Vision- Based- Class ifica tion- of- Skin- Cancer- using- Kalou
che/ b57ba 90975 6462d 812dc 20fca 157b3 972bc 1f533. Accessed 10 Jan 2021
58. Kanca E, Ayas S (2022) Learning hand-crafted features for k-NN based skin disease classification.
In: 2022 international congress on human-computer interaction, optimization and robotic applica-
tions (HORA). https:// doi. org/ 10. 1109/ hora5 5278. 2022. 97998 34
59. Kassem MA, Hosny KM, Damaševičius R, Eltoukhy MM (2021) Machine learning and deep
learning methods for skin lesion classification and diagnosis: a systematic review. Diagnostics
11(8):1390. https:// doi. org/ 10. 3390/ diagn ostic s1108 1390
60. Kaur M, Kumar V, Yadav V, Singh D, Kumar N, Das NN (2021) Metaheuristic-based deep covid-
19 screening model from chest X-ray images. J Healthc Eng 2021:1–9. https:// doi. org/ 10. 1155/
2021/ 88298 29
61. Khan MA, Zhang Y-D, Sharif M, Akram T (2021) Pixels to classes: intelligent learning frame-
work for multiclass skin lesion localization and classification. Comput Electr Eng 90(106):956.
https:// doi. org/ 10. 1016/j. compe leceng. 2020. 106956
62. Kumar S; Kumar A Extended feature space-based automatic melanoma detection system. arXiv
2022, arXiv:2209.04588
63. Kumar N, Gupta M, Gupta D, Tiwari S (2021) Novel deep transfer learning model for COVID-19
patient detection using X-ray chest images. J Ambient Intell Humaniz Comput. https:// doi. org/ 10.
1007/ s12652- 021- 03306-6
64. Kumar N, Gupta M, Sharma D, Ofori I (2022) Technical job recommendation system using apis
and web crawling. Comput Intell Neurosci 2022:1–11. https:// doi. org/ 10. 1155/ 2022/ 77975 48
65. Lopes J, Rodrigues CM, Gaspar MM, Reis CP (2022) How to treat melanoma? the current status
of innovative nanotechnological strategies and the role of minimally invasive approaches like PTT
and PDT. Pharmaceutics 14(9):1817. https:// doi. org/ 10. 3390/ pharm aceut ics14 091817
66. Maglogiannis I, Delibasis K (2015) Hair removal on dermoscopy images. In: 2015 37th annual
international conference of the IEEE engineering in medicine and biology society (EMBC).
https:// doi. org/ 10. 1109/ embc. 2015. 73190 13
67. Mahbod A, Schaefer G, Wang C, Ecker R, Ellinge I (2019) Skin lesion classification using hybrid
deep neural networks. In: ICASSP 2019 - 2019 IEEE international conference on acoustics, speech
and signal processing (ICASSP). https:// doi. org/ 10. 1109/ icassp. 2019. 86833 52
Multimedia Tools and Applications
1 3
68. Malibari AA, Alzahrani JS, Eltahir MM, Malik V, Obayya M, Duhayyim MA, Lira Neto AV, de
Albuquerque VH (2022) Optimal deep neural Network-driven computer aided diagnosis model
for skin cancer. Comput Electr Eng 103(108):318. https:// doi. org/ 10. 1016/j. compe leceng. 2022.
108318
69. Mane S, Shinde S (2018) A method for melanoma skin cancer detection USING Dermoscopy
Images. In: 2018 fourth international conference on computing communication control and auto-
mation (ICCUBEA). https:// doi. org/ 10. 1109/ iccub ea. 2018. 86978 04
70. Marks R (1995) An overview of skin cancers. Cancer 75(S2):607–612
71. Masad IS, Alqudah A, Alqudah AM, Almashaqbeh S (2021) A hybrid deep learning approach
towards building an intelligent system for pneumonia detection in chest X-ray images. Int J Electr
Comput Eng (IJECE) 11(6):5530–5540. https:// doi. org/ 10. 11591/ ijece. v11i6
72. Mendonca T, Ferreira PM, Marques JS, Marcal AR, Rozeira J (2013) Ph2 - A dermoscopic image
database for research and benchmarking. In: 2013 35th annual international conference of the IEEE
engineering in medicine and biology society (EMBC). https:// doi. org/ 10. 1109/ embc. 2013. 66107 79
73. Menzies method. Dermoscopedia.org. (n.d.). Retrieved December 26, 2022, from https:// dermo
scope dia. org/w/ index. php? title= Menzi es_ Metho d& oldid= 9988
74. Moazen H, Jamzad M (2020) Automatic skin cancer (melanoma) detection by processing der-
matoscopic images. In: 2020 international conference on machine vision and image processing
(MVIP). https:// doi. org/ 10. 1109/ mvip4 9855. 2020. 91169 18
75. Montaha S, Azam S, Rafid AK, Islam S, Ghosh P, Jonkman M (2022) A shallow deep learning
approach to classify skin cancer using down-scaling method to minimize time and space complex-
ity. PLoS ONE 17(8). https:// doi. org/ 10. 1371/ journ al. pone. 02698 26
76. Munia TT, Alam MN, Neubert J, Fazel-Rezai R (2017) Automatic diagnosis of melanoma using linear
and nonlinear features from digital image. In: 2017 39th annual international conference of the IEEE
engineering in medicine and biology society (EMBC). https:// doi. org/ 10. 1109/ embc. 2017. 80378 02
77. Nancy VA, Arya MS, Nitin N (2022) Impact of data augmentation on skin lesion classification
using deep learning. In: 2022 5th international conference on information and computer technolo-
gies (ICICT). https:// doi. org/ 10. 1109/ icict 55905. 2022. 00020
78. Narayan Das N, Kumar N, Kaur M, Kumar V, Singh D (2022) Automated deep transfer learn-
ing-based approach for detection of COVID-19 infection in chest x-rays. IRBM 43(2):114–119.
https:// doi. org/ 10. 1016/j. irbm. 2020. 07. 001
79. Nugroho AA, Slamet I, Sugiyanto. (2019) Skins cancer identification system of Haml0000 skin
cancer dataset using Convolutional Neural Network. In: International conference on science and
applied science (icsas) 2019. https:// doi. org/ 10. 1063/1. 51416 52
80. Ozturk S, Cukur T (2022) Deep clustering via center-oriented margin free-triplet loss for skin
lesion detection in highly imbalanced datasets. IEEE J Biomed Health Inform 26(9):4679–4690.
https:// doi. org/ 10. 1109/ jbhi. 2022. 31872 15
81. Patil SM, Rajguru BS, Mahadik RS, Pawar OP (2022) Melanoma skin cancer disease detection
using convolutional neural network. In: 2022 3rd international conference for emerging technol-
ogy (INCET). https:// doi. org/ 10. 1109/ incet 54531. 2022. 98253 81
82. Performance evaluation of different machine learning classification algorithms for diseases diag-
nosis. (2021). Int J E-Health Med Commun, 12(6). https:// doi. org/ 10. 4018/ ijehmc. 20211 101oa 09
83. Premier Surgical Staff. What is the difference between melanoma and non-melanoma skin cancer?
PSS. Available online: https:// www. premi ersur gical. com/ 01/ whats- the- diffe rence- betwe en- melan
oma- and- non- melan oma- skin- cancer/
84. Pujara A (2022) Image classification with MobileNet. Medium. Retrieved December 12, 2022,
from https:// medium. com/ analy tics- vidhya/ image- class ifica tion- with- mobil enet- cc6fb b2cd4 70
85. Rahman Z, Hossain MS, Islam MR, Hasan MM, Hridhee RA (2021) An approach for multiclass
skin lesion classification based on ensemble learning. Inform Med Unlocked 25(100):659. https://
doi. org/ 10. 1016/j. imu. 2021. 100659
86. Sarker IH (2021) Machine learning: algorithms, real-world applications and research directions.
SN Comput Sci 2(3). https:// doi. org/ 10. 1007/ s42979- 021- 00592-x
87. Seeja RD, Suresh A (2019) Deep learning based skin lesion segmentation and classification of
melanoma using support vector machine (SVM). Asian Pac J Cancer Prev 20(5):1555–1561.
https:// doi. org/ 10. 31557/ apjcp. 2019. 20.5. 1555
88. Sharafudeen M, S VC (2022) Detecting skin lesions fusing handcrafted features in image network
ensembles. Multimed Tools Appl 82(2):3155–3175. https:// doi. org/ 10. 1007/ s11042- 022- 13,046-0
89. Sharma P, Gautam A, Nayak R, Balabantaray BK (2022) Melanoma detection using advanced
deep neural network. In: 2022 4th international conference on energy, power and environment
(ICEPE). https:// doi. org/ 10. 1109/ icepe 55035. 2022. 97981 23
Multimedia Tools and Applications
1 3
90. Shetty B, Fernandes R, Rodrigues AP, Chengoden R, Bhattacharya S, Lakshmanna K (2022) Skin
lesion classification of dermoscopic images using machine learning and convolutional neural net-
work. Sci Rep 12(1). https:// doi. org/ 10. 1038/ s41598- 022- 22,644-9
91. Step by step VGG16 implementation in Keras for beginners. (n.d.). Retrieved December 12, 2022,
from https:// towar dsdat ascie nce. com/ step- by- step- vgg16- imple menta tion- in- keras- for- begin ners-
a833c 686ae 6c
92. Sturm RA (2002) Skin Colour and Skin Cancer -MC1R, the Genetic Link. Melanoma Res.
12:405–416
93. Tabrizchi H, Parvizpour S, Razmara J (2022) An improved VGG model for skin cancer detection.
Neural Process Lett. https:// doi. org/ 10. 1007/ s11063- 022- 10927-1
94. Tschandl P (2021) The HAM10000 dataset, a large collection of multi-source dermatoscopic
images of common pigmented skin lesions. Harvard Dataverse. Retrieved January 7, 2023, from
https:// datav erse. harva rd. edu/ datas et. xhtml? persi stent Id= doi% 3A10. 7910% 2FDVN% 2FDBW 86T
95. Tschandl P, Rosendahl C, Kittler H (2018) The HAM10000 dataset, a large collection of multi-
source dermatoscopic images of common pigmented skin lesions. Sci Data 5(1). https:// doi. org/
10. 1038/ sdata. 2018. 161
96. Types of optimizers in deep learning every AI engineer should know. upGrad blog. (2022)
Retrieved January 7, 2023, from https:// www. upgrad. com/ blog/ types- of- optim izers- in- deep- learn
ing/#: ~: text= In% 20deep% 20lea rning% 2C% 20opt imize rs% 20are ,train ing% 20a% 20neu ral% 20net
work% 20mod el
97. Venugopal V, Joseph J, Vipin Das M, Kumar Nath M (2022) An efficientnet-based modified sigmoid
transform for enhancing dermatological macro-images of melanoma and Nevi skin lesions. Comput
Methods Programs Biomed 222(106):935. https:// doi. org/ 10. 1016/j. cmpb. 2022. 106935
98. Venugopal V, Joseph J, Das MV, Nath MK (2022) DTP-net: A convolutional neural network model
to predict threshold for localizing the lesions on dermatological macro-images. Comput Biol Med
148(105):852. https:// doi. org/ 10. 1016/j. compb iomed. 2022. 105852
99. Vipin V, Nath MK, Sreejith V, Giji NF, Ramesh A, Meera M (2021) Detection of melanoma using
deep learning techniques: a review. In: 2021 international conference on communication, control and
information sciences (ICCISc). https:// doi. org/ 10. 1109/ iccis c52257. 2021. 94848 61
100. Vocaturo E, Zumpano E, Veltri P (2018) Image pre-processing in computer vision systems for mela-
noma detection. In: 2018 IEEE international conference on bioinformatics and biomedicine (BIBM).
https:// doi. org/ 10. 1109/ bibm. 2018. 86215 07
101. What is hyperparametertuning? Anyscale.(n.d.) Retrieved January 7, 2023, from https:// www. anysc
ale. com/ blog/ what- is- hyper param eter- tuning
102. What is learning rate in machine learning. Deepchecks (2022) Retrieved January 7, 2023, from
https:// deepc hecks. com/ gloss ary/ learn ing- rate- in- machi ne- learn ing/#: ~: text= The% 20lea rning%
20rate% 2C% 20den oted% 20by,netwo rk% 20con cerni ng% 20the% 20loss% 20gra dient% 3E
103. Woodie A (2017) Machine learning, deep learning, and ai: What’s the difference? Datanami.
Retrieved January 7, 2023, from https:// www. datan ami. com/ 2017/ 05/ 10/ machi ne- learn ing- deep- learn
ing- ai- whats- diffe rence/
104. Wu Y, Lariba AC, Chen H, Zhao H (2022) Skin lesion classification based on deep convolutional
neural network. In: 2022 IEEE 4th international conference on power, intelligent computing and sys-
tems (ICPICS). https:// doi. org/ 10. 1109/ icpic s55264. 2022. 98737 56
105. Yuan X, Yang Z, Zouridakis G, Mullani N (2006) SVM-based texture classification and application to
early melanoma detection. In: 2006 international conference of the IEEE engineering in medicine and
biology society. https:// doi. org/ 10. 1109/ iembs. 2006. 260056
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and applicable
law.
Multimedia Tools and Applications
1 3
Authors and Aliations
V.AuxiliaOsvinNancy1· P.Prabhavathy1· MeenakshiS.Arya2· B.ShamreenAhamed1
* V. Auxilia Osvin Nancy
auxilianancy@gmail.com
P. Prabhavathy
prabhavp1@srmist.edu.in
Meenakshi S. Arya
raina.arya@gmail.com
B. Shamreen Ahamed
shamu1502@gmail.com
1 Department ofComputer Science andEngineering, College ofEngineering andTechnology,
SRM Institute ofScience andTechnology, Vadapalani campus, No.1, Jawaharlal Nehru Road,
Vadapalani, Chennai, TamilNadu, India
2 Institute ofTransportation, Iowa State University, Ames, IA, USA
A preview of this full-text is provided by Springer Nature.
Content available from Multimedia Tools and Applications
This content is subject to copyright. Terms and conditions apply.
