See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/361228242

A machine learning model for skin disease classification using convolution

neural network

Article · December 2021

CITATIONS READS

70 5,997

1 author:

Viswanatha Reddy Allugunti

VIT University
11 PUBLICATIONS 145 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Healthcare View project

All content following this page was uploaded by Viswanatha Reddy Allugunti on 11 June 2022.

The user has requested enhancement of the downloaded file.

 International Journal of Computing, Programming and Database Management 2022; 3(1): 141-147

E-ISSN: 2707-6644
P-ISSN: 2707-6636
IJCPDM 2022; 3(1): 141-147 A machine learning model for skin disease
Received: 09-10-2021
Accepted: 16-12-2021 classification using convolution neural network
Viswanatha Reddy Allugunti
Chief Solutions Architect, Viswanatha Reddy Allugunti
Arohak Inc, New Jersey, USA

Abstract
Melanoma is a skin disease that tends to be lethal. It occurs when melanocytes develop in an
uncontrolled manner. Melanoma goes under a few different names, including malignant melanoma.
The incidence of melanoma is at its highest level ever recorded in both Australia and New Zealand. It
is estimated that one in every 15 white New Zealanders will indeed be diagnosed with melanoma at
some point in their lives. Aggressive malignancy was the third most common kind of cancer in men
and women in 2012, respectively. Melanoma can develop at any age in adults, but it is highly unusual
in children and teenagers. It is hypothesized that the first step in developing melanoma is an
unregulated multiplication of melanocytic stem cells that have been genetically altered. The survival
rate can significantly increase if melanoma is identified in dermos copy images at an earlier stage. On
the other hand, the detection of melanomas is an incredibly challenging task. Consequently, the
detection and recognition of skin cancer are of tremendous assistance to the accuracy of pathologists.
In this research, a deep learning technique is shown for reliably diagnosing the type of melanoma
present at a preliminary phase. The proposed model makes a distinction among lesion maligna,
superficial spreading, and nodular melanoma. This permits the early diagnosis of the virus and the
quick isolation and therapy necessary to stop the transmission of infection further. Deep learning (DL)
and the standard non-parametric machine learning method are exemplified in the deep layer topologies
of the convolutional neural network (CNN), which are neural network algorithms. The effectiveness of
a CNN classifier was evaluated using data retrieved from the website https://dermnetnz.org/. The
outcomes of the experiments show that the proposed method is superior in terms of diagnostic accuracy
compared to the methodologies that are currently considered state of the art.

Keywords: Types of melanoma, skin disease, deep learning models, convolutional neural network,
machine learning classification

1. Introduction
Since the 1970s, skin cancer has held the title of the most prevalent disease globally. Over
the previous several decades, there has been an uptick in people diagnosed with non-
melanoma and melanoma skin cancers, respectively. Melanoma can be identified in that only
one in three cases of cancer, as stated by the World Health Organization (WHO), and
according to statistics provided by the Skin Cancer Foundation, one out of every five people
in the United States will develop skin cancer at some point during their lifetime. For the past
several centuries, the incidence of skin cancer has risen at a relatively constant rate,
particularly in the Western hemisphere. Countries such as the United States, Canada, and
Australia are just some of the places where this trend has been observed. Infectious diseases
of the skin typically have the potential to have a significant detrimental effect on the overall
health of people all over the world. According to one piece of a study released in 2017,
multiple studies have demonstrated that skin cancer is responsible for 1.79 percent of the
disease burden assessed in disability-adjusted life years on a global scale [1]. The incidence of
skin cancer accounts for around 7 percent of all newly diagnosed instances of cancer
worldwide [2], resulting in a loss of more than $8 billion for the Medicare program in the
United States in 2011. Clinical data suggest that there are such disparities in results based on
race in the case of skin cancer: Even though people with darker skin tones are approximately
20 to 30 times as likely to develop melanoma than those who have lighter skin tones, it has
been discovered that people with darker skin tones either have a higher or lower mortality
risk for specific types of melanoma, depending on their skin tone.
Corresponding Author:
Viswanatha Reddy Allugunti In order to administer the appropriate treatment, it is essential to identify a skin lesion
Chief Solutions Architect, correctly. It is possible to accomplish early diagnosis of melanoma in dermoscopy
Arohak Inc, New Jersey, USA photographs and pictures using this method, which improves the survival rate.

~ 141 ~
International Journal of Computing, Programming and Database Management http://www.computersciencejournals.com/ijcpdm

Dermatologists who have had considerable training in the 2. Literature Survey

many skin lesions that melanomas might cause are the most In order to address the major three aims in the field of skin
qualified to make an accurate diagnosis. Because of this, lesion image processing, the author of this study [7] utilised
diagnosing melanoma can be a challenging task because two different deep learning algorithms, namely the Lesion
there is no clear separation between skin lesions and the Feature Network (LFN) and the Lesion Indexing Network
skin itself, malignant and non-melanoma skin lesions appear (LIN). a) The division of the lesions into segments b) The
visually similar, and there are other factors to take into Removal of Dermoscopic Characteristics from the Lesion c)
account. Therefore, creating a trustworthy automatic The Categorization of the Lesion The author suggests a deep
detection method for skin tumors, such as a system that can learning system that is made up of two fully convolutional
automatically analyze skin lesions, will be greatly useful to residual networks (FCRN), which produce results for
pathologists. This is especially relevant in an era where segmentation and classification. In order to refine the
knowledge is scarce. findings of the coarse classification, the lesion index
According to the findings of this study, the classification calculation unit (LICU) is constructed by employing a
methods of K-nearest Neighbors, Support Vector Machines, distance heat map. A straightforward CNN is demonstrated
and Decision Trees all produced subpar results in terms of as a solution to the problem of extracting dermoscopic
precision and accuracy. After conducting further research features. In order to assess the provided deep learning
into the mathematics that underlies classification, it was framework, the author employed the ISIC 2017 dataset. The
found that employing Deep Learning models was the most author's experiment shows that the suggested (LIN)
sophisticated method for getting the desired outcomes (also performs better than existing machine learning algorithms
known as deep learning models). We experimented with for lesion segmentation and classification. However, for
many different mathematical models, both with and without dermoscopic feature extraction, the performs better than the
the application of Learning Algorithms. However, we proposed (LIN).
concluded that the depth and quality of activation that was The author developed a fully automatic method for the
made available by pre-trained models did not meet up. segmentation of skin lesions in the article [8]. This method
Consequently, we merged our mathematical expertise and makes optimal use of a trained 19-layer deep CNNs and
developed a model known as a Dense Convolutional therefore does not rely on prior knowledge of the data. The
Network, which offered an accuracy of more than 86.6 author came up with a series of strategies to assure effective
percent. and efficient learning despite having very little training data
We were able to accomplish efficient and reliable picture to work with. When using cross entropy as the loss function
categorization by using Deep Learning, a branch of artificial for picture segmentation, which is a normal process, there is
intelligence that is exceptionally robust and potent in its a severe imbalance between both the amount of foreground
capabilities. The structure and operation of this artificial and background pixels. As a result of this, an original loss
brain were extremely comparable to those of the human function based on Jaccard distance has also been designed to
brain, with neurons activating across the brain to transfer
minimise the requirement for sample re-weighting. The
information, categorize the data, draw conclusions, and
author made use of two datasets that are open to the public
produce consequences. Neural Networks are used in Deep
in order to evaluate the usefulness, efficiency, and potential
Learning, a form of machine learning. A Neural Network
for generalisation of the proposed framework. These
structure is a stack of layers collectively known as a Neural
databases are the ISBI 2016 database and the PH2 database.
Network structure. As their name suggests, neural networks
The authors of [9] show how to identify skin lesions using a
can perform functions comparable to those that neurons in
the brain can execute, such as recognizing patterns and single CNN that was trained from the ground up using
generating predictions. photographs utilising only pixels and sickness labels as
The project's overarching objective is to come up with a inputs. This method was used to train the CNN from start to
method of screening for skin cancer that is both reasonably finish. In order to train CNN, the author employed a dataset
quick and straightforward for the general public to use. In consisting of 129,450 clinical photos, which is two orders of
most instances, the sooner it is recognized, the better the magnitude greater than the prior datasets, which consisted of
probability that the individual will make a full recovery. 2,032 disorders. Utilizing biopsy-proven clinical pictures,
According to the American Academy of Dermatology 21 board-certified dermatologists evaluate the performance
Association, most dermatological malignancies are treatable of the system in two binary classification use cases:
if found at an early stage. The trained model is just a keratinocyte carcinomas versus benign seborrheic keratoses
preliminary step before doing a skin biopsy. Visit a and malignant melanomas versus benign nevi. Both use
dermatologist and get a skin biopsy performed if you want cases compare the system's accuracy in diagnosing the two
to know for sure if you have skin cancer or not. This is the types of skin lesions. In the first situation, it is necessary to
only method to get an accurate diagnosis. determine which malignancies are the most prevalent, and in
The organization of this document is broken down into its the second scenario, it is necessary to determine which skin
parts in the following paragraphs. In Chapter 2, "Literature cancer is the deadliest. According to the results of the
Survey," there was a comprehensive summary of all the author's experiment, CNN performed better than all of the
previous studies. In Section 3, you should describe the specialists who were tested in both tasks, proving that
design of the proposed system (System Architecture). The machine learning is capable of recognizing skin cancer at a
experimental setup and the dataset used in the experiment level of competence comparable to that of dermatologists.
are discussed in Section 4 (Results and Discussion). Section The author of this article [10] presents a novel method for
4 also examines the results of the experiment. Finally, in the melanoma recognition that is based on deep CNNs and
section under "Conclusion," both the conclusion reached contrasts it to earlier methods that were either based on low-
about the system and the work that will be done in the future level hand-crafted features or CNNs with shallower
to improve it were discussed. architectural designs. According to the authors' findings,

~ 142 ~
International Journal of Computing, Programming and Database Management http://www.computersciencejournals.com/ijcpdm

their method, which makes use of substantially deeper finished layers. Only the layers that were removed and
networks, is able to learn discriminative traits, which results restored are complete. According to an alternative model,
in better identification accuracy. In order to make full use of unknown photos could be identified using GoogleNet and
very deep networks, the author proposed a variety of SVM multi-class systems. The authors attempted to use the
methodologies that can be used to ensure good training and VGG19 in the same way as they used GoogleNet, but they
learning even with a limited amount of training data. The were unable to test the same equipment or use it themselves.
following procedures make up the method: a) When a It requires very stringent hardware criteria, which most
network goes deeper, employ residual learning to address researchers working in different nations simply are unable
degradation and overfitting issues so that the network may to meet.
continue to function properly. The performance gains that They developed a melanoma detection and segmentation
were made possible by raising the network depth will be approach in a work that was published by Saleh Albahli and
preserved as a result of this action. b) Construct a fully colleagues. [Citation needed] Using morphological
convolutional residual network (FCRN) for accurate skin procedures on dermoscopic photos allows for the removal of
lesion segmentation, and then enhance it through the artefacts like hairs, gel bubbles, and clinical marks, as well
application of a multi-scale contextual information as the sharpening of image regions. They used the YOLOv4
integration technique. c) If you want to make a two-stage object detector, which was modified for melanoma
framework, combine the deep residual networks and the detection, to differentiate between strongly correlated
FCRN that was provided. This approach enables the infected regions and non-infected regions in order to
classification network to extract more representative and discover infected regions. Using the ISIC2018 and
specific features from segmented findings rather of the ISIC2016 datasets, the results of the presented system are
complete dermoscopy images, hence reducing the quantity analysed and compared to the results of melanoma
of training data that is required to be collected. The author identification and segmentation techniques.
used the ISBI 2016 dataset for the purpose of evaluation. This model includes a feature discriminating network and a
Rashmi Patil and her associates [11] with the gradient descent lesion classification network, according to Lisheng Wei et
Similarity Measure presented CNN approach for Text al. Additionally, it incorporates a lightweight skin malignant
Processing (SMTP). This research resulted in the recognition model with feature discrimination dependent on
development of two different methods for identifying the fine-grained classification standard. They devised a
stages of melanoma cancer. In the first approach, melanoma discriminating model in order to recognise lesions in
is divided into two phases, which are referred to as stages 1 dermoscopy images. In order to develop a dermoscopy
and 2. Melanoma is divided into three distinct phases, which image lesion classification branch network, it makes use of
are referred to as stages 1, 2, and 3. The suggested SMTP a pre-trained lightweight network for the purpose of feature
loss function causes exceptionally little loss in compared to extraction. Through collaborative training of each branch
all of the previous loss functions, and this has shown to be network, the presented model is able to conduct
very effective in terms of enhancing sensitivity, specificity, simultaneous classification of lesion type and similarity of
and accuracy. lesion features. This enables the model to extract more
Konstantin Korotkov and colleagues [12] described discriminative lesion features. Our framework produces
approaches to address the issue of matching skin lesions as model performance comparable to, and in some
part of their research. The process known as "whole body circumstances better than, the existing multi-CNN fusion
photography" entails taking clinical photographs of patients approach or the scheme based on local depth feature Fisher
in various seated positions while they are in an institutional Vector coding, but having fewer model parameters. Our
setting. Clinical specialists in total body skin evaluation framework's fewer parameters enable this.
make use of photographs in order to document the current Loretta Ichim and her colleagues proposed a unique
condition of a patient's skin, as well as study the progression approach that is built on several neural networks that are
of a variety of cutaneous conditions. It plays an important connected on two classification stages. Transferring the
role in the early detection of melanoma, a potentially fatal system's data from one database to another is made easier by
form of skin cancer. Images of skin lesions are shown by using this procedure. The use of several classifiers was
circles that are a perfect fit within optimally steady extreme suggested as a method that might effectively identify
areas. After the lesion has been identified, those results will melanomas in images of lesions at a reduced cost. When
be added to the matching algorithm. selecting classifiers, the following considerations were taken
M. A. Kassem and colleagues introduced a method that into account:
makes use of transference learning and a pre-trained Google a) Classifiers are to be arranged on two levels of
Network. This method was created for the ISIC 2019 hierarchy, one of which is subjective and the other of
Challenges Data Set. Even if there are imbalances between which is objective.
the classes, the suggested method is still capable of b) Classifiers for the first stage that make use of a variety
accurately classifying all eight different types of lesions. of lesion characteristics.
The suggested methodologies' precision, sensitivity, c) Relevant learning found that any classification
specificities, and precision were measured, and their dependent on characteristics was subjective.
respective values were 94.92 percent, 79.8 percent, 97
percent, and 80.36 percent, respectively. The performance In addition to this, the implementation of a conditional GAN
of the method that was provided increased when the number for categorising accidents in accordance with the ABCD law
of images in each class was decreased in order to eliminate was an important addition to this. The accuracy of the
the problem of imbalance between classes. When weight is classifier has been brought up to par. The final classifier was
modified for each layer of the architecture, the performing connected with a likelihood that was derived from the TDS
metrics are significantly higher than those for merely the score after it had been turned into a probability. The

~ 143 ~
International Journal of Computing, Programming and Database Management http://www.computersciencejournals.com/ijcpdm

suggested two-tier architecture had the advantage of spreading melanoma. Additionally, the model correctly
allowing for quicker switching between databases (or identified uneven borders from melanoma skin lesions,
devices), with the need to learn only the second stage (final which is an extremely significant finding. A dataset
perceptron PF). The proposedsystem increases the consisting of 2475 dermoscopic images was developed so
effectiveness of the many alternative comparison that algorithms could be trained on and evaluated using the
approaches. As a consequence of this, the goal neural data. Regarding the precision, the f1 score, the recall, and
network achieves a score of 97.5 percent ACC and 94.47 the accuracy. The results of many experiments indicate that
percent, 97 percent ACC, and 93.67 percent F1 VGG19 performs better than ResNet50, VGG16, and
correspondingly. MobileNet.
A customised CNN that has the capability to automatically A feed-forward artificial neural network having one or more
learn features and predict the class of numbers from a big layers that are entirely coupled together is known as an
data collection is demonstrated, as well as a visualisation of MLP network. At a minimum, an MLP network will include
the intermediate results. three layers, which are referred to as "the input layer," "the
According to D. T. Mane et al., customised CNN layers that hidden layer," and "the output layer".
show the dynamics of the network were developed. There
are 80,000 different Marathi numerals; however, only 3. Implementation Details
70,000 are used for instruction, while the remaining 10,000 A. Convolutional Neural Network (CNN)
are put to use in examinations. When the CNN's In Image Processing, the most commonly used and efficient
performance was checked using K-fold cross validation, it algorithm of Deep Learning is CNN. These are the most
attained an accuracy rate of 94.93 percent on average for the cutting-edge approaches to deep learning [2]. Due to its
data sets that were used for testing. The photos can have parameter-sharing and stronger adaptation capabilities than
high-level properties extracted from them using CNN that a deep learning model, CNN has several applications;
has been customised. By combining the capabilities of the include biological text analysis, object detection and
central processing unit (CPU) and the graphics processing recognition, and malware classification. Researchers exploit
unit (GPU). The overall accuracy of the system can be CNN's features in several ways. CNNs were inspired by the
summed up as 94.93 percent if we look at it in its whole. cortical organisation for animal vision, which detects light
Rashmi Patil et al. state that melanomas can manifest in a in small patches that overlap one another across the visual
variety of different ways and might not display any of the field. As we delve deeper into the network, the layers will
typical warning signs. [Citation needed] It is extremely extract higher-level properties that are increasingly accurate.
important to detect melanomas in their early stages. This It is common knowledge that CNN models have the ability
research uses transfer learning to create a computer-aided to create a correlation between the internal representation of
melanoma assessment tool. Several existing frameworks pixel values and how those values are shown in the form of
show that computer vision may be used to find medical a two-dimensional matrix. It works particularly effectively
images. This paper also describes a computer-aided with information that can be linked to a specific location.
melanoma identification method. The model correctly The architecture of the CNN is shown in figure 2.
categorised lentigo maligna, nodular, and superficial

Fig 1: CNN architecture

B. System Architecture 4. In this experiment, the CNN classification approach is

The following are the stages that must be completed in order applied to both the newly constructed training file
for our suggested system to be implemented. collection and the freshly created test input images,
1. As input, the system will take a dataset of picture data. which were both made from scratch.
2. To increase the quality of the image and eliminate hairs 5. Melanomas are detected by using the CNN algorithm,
from it, which was before is carried out before to which determines whether or not melanoma is present
printing. in the input test data set.
3. A training file is formed as a consequence of the
extraction of a number of features from the input image Finally, in order to determine the overall performances of
dataset. the proposed technique and provide recommendations, a
graphic evaluation is carried out at the conclusion

~ 144 ~
International Journal of Computing, Programming and Database Management http://www.computersciencejournals.com/ijcpdm

Fig 2: Proposed system architecture

4. Experimental Results despite the fact that the model incorrectly anticipated
A. Experimental Dataset them to be normal.
The dataset was contributed via the website  A false positive, sometimes known as an FP, is an
https://dermnetnz.org/. In order to prevent ourselves from incorrectly positive prediction made when the observed
over fitting the data, we made use of a data augmentation assault is, in fact, normal.
method. It's a method for generating new training data  The percentage of false positives that correctly identify
artificially out of existing data, using the old data. The attacks is the value that is being measured by the true
photographs have had a zoom effect added to them as an negative (TN).
enhancement. It does so in a haphazard manner, either
randomly zooming in on the image and adding new image When evaluating the performance of suggested task,
pixels or interpolating existing ones. accuracy, recall, f1 score, and precision are all taken into
consideration. The origin of measurements is found in
Table 1: Dataset Description

(1)

B. Experimental Setup (2)

The research was conducted on a personal computer running
Windows 10 that had a Core i7 processor, 8 gigabytes of
dynamic random access memory (DDRAM), a GeForce (3)
MX150 NVIDIA graphics card, and a 500 gigabyte solid
state drive. The Anaconda IDE was utilised so that a
programme could be executed. (4)

C. Evaluation Parameters D. Results Analysis

After the models had been trained, we put them through
their paces using the held-over test set. After that, the Table 2: Performance Analysis of ML, and DL Algorithms
confusion matrix was utilised in the computation of the
Classifier Precision Recall F1-Score Accuracy
performance metrics. In order to signal both intended and DT 70.00 69.00 59.00 67.00
actual classifications, elements of the confusion matrix are RF 71.00 71.00 65.00 71.35
utilised. The process of classification results in two classes: GBT 68.00 73.00 70.00 73.44
correct and incorrect. In order to calculate the prediction CNN 91.07 87.68 89.32 88.83
model, we conducted an analysis of the following four
fundamental case studies: The accuracy graphs with regard to the number of epochs
 The term "true positive" refers to the percentage of true are displayed in Figure 3, which is a comparison of the
positives that are identified with high degree of CNN accuracy and the CNN validation accuracy with
precision (TP). respect to the number of epochs. The loss comparison
 Incorrect forecasts are referred to as having a false graphs of various CNN classification algorithms are shown
negative (FN). It finds situations that are malevolent, in Figure 4.

~ 145 ~
International Journal of Computing, Programming and Database Management http://www.computersciencejournals.com/ijcpdm

Fig 3: Accuracy and Validation Accuracy of CNN classifier

Fig 4: Loss Comparison graphs of CNN classifiers

5. Conclusions 2. Bi Lei, Jinman Kim, Euijoon Ahn, Feng D. Automatic

Within the scope of this study, a Convolutional Neural Skin Lesion Analysis using Large-scale Dermoscopy
Network (CNN) model for the diagnosis of skin cancer was Images and Deep Residual Networks. ArXiv abs.
created, constructed, and evaluated using a well-known 2017;1703:04197.
melanoma dataset. Our proposed method, which is a two- 3. Yuan Y, Lo YC. Improving Dermoscopic Image
stage learning platform, has great-predicted accuracy at each Segmentation with Enhanced Convolutional-De
stage, as demonstrated by its overall accuracy of 88.83 convolutional Networks. IEEE J Biomed Health
percent. This is true not only for classification algorithms Inform. 2019 Mar;23(2):519-526. DOI:
such as DT, RF, GBT. 10.1109/JBHI.2017.2787487. Epub 2017 Dec 25.
The strategy that has been suggested is based on CNN, and PMID: 29990146.
it is possible to think of it as an effective method of multi- 4. Feng-ying Xie, Jiawen Yang, Liu J, Zhi-guo Jiang,
class categorization. In terms of melanoma classification Yushan Zheng, Yukun Wang. Skin lesion segmentation
accuracy, the modular and hierarchical structure of our CNN using high-resolution convolutional neural network.
classifier not only beats state-of-the-art machine learning Computer methods and programs in biomedicine.
techniques, but it significantly minimises the amount of 2020;186:105241.
computational effort that is required. The fact that this 5. Yao Y, Luo Z, Li S, Fang T, Quan L. MVSNet: Depth
strategy is only tested on a single dataset is one of the Inference for Unstructured Multi-view Stereo. In:
method's drawbacks. Ferrari V., Hebert M., Sminchisescu C., Weiss Y. (eds)
Computer Vision – ECCV 2018. ECCV 2018. Lecture
6. References Notes in Computer Science, 2018, 11212. Springer,
1. Zhang Ce, Xin Pan, Huapeng Li, Gardiner A, Sargent I, Cham. https://doi.org/10.1007/978-3-030-01237-3_47
Jonathon S Hare, et al. A hybrid MLP-CNN classifier 6. Li Y, Shen L. Skin Lesion Analysis towards Melanoma
for very fine resolution remotely sensed image Detection Using Deep Learning Network. Sensors
classification. Isprs Journal of Photogrammetry and (Basel). 2018 Feb 11;18(2):556. DOI:
Remote Sensing. 2017;140:133-144.

~ 146 ~
International Journal of Computing, Programming and Database Management http://www.computersciencejournals.com/ijcpdm

10.3390/s18020556. PMID: 29439500; PMCID:

PMC5855504.
7. Rashmi Patil, Sreepathi Bellary. Machine learning
approach in melanoma cancer stage detection, Journal
of King Saud University - Computer and Information
Sciences, 2020, ISSN:1319-1578,
https://doi.org/10.1016/j.jksuci.2020.09.002.
8. Konstantin Korotkov, Josep Quintana, Ricard Campos,
Am´erica Jes´us-Silva, Pablo Iglesias, Susana Puig, et
al. An Improved Skin Lesion Matching Scheme in
Total Body Photography, IEEE Journal of Biomedical
and Health Informatics. 2019 Mar;23(2):586-598.
9. Kassem MA, Hosny KM, Fouad MM. Skin Lesions
Classification into Eight Classes for ISIC 2019 Using
Deep Convolutional Neural Network and Transfer
Learning, in IEEE Access. 2020;8:114822-114832.
DOI: 10.1109/ACCESS.2020.3003890.
10. Ichim L, Popescu D. Melanoma Detection Using an
Objective System Based on Multiple Connected Neural
Networks, in IEEE Access. 2020;8:179189-179202.
DOI: 10.1109/ACCESS.2020.3028248.
11. Patil R, Bellary S. Transfer learning based system for
melanoma type detection. Revue d'Intelligence
Artificielle. 2021;35:(2):123-130.
https://doi.org/10.18280/ria.350203
12. Ravi V, Pradeepkumar D, Deb K. Financial time series
prediction using hybrids of chaos theory, multilayer
perceptron and multi-objective evolutionary algorithms.
Swarm Evolut Comput. 2017;36:136-149.

~ 147 ~

View publication stats