You are on page 1of 7

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

net/publication/365186671

PROPOSING AN EFFICIENT MODEL TO DETECT MELANOMA BASED ON


DENSE-CNN

Article  in  International Journal on Soft Computing Artificial Intelligence and Applications · May 2022

CITATIONS READS

0 5

2 authors, including:

Adnan Saher Mohammed


Karabuk University
17 PUBLICATIONS   44 CITATIONS   

SEE PROFILE

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

FWM reduction View project

On the Energy Consumption in Optical Burst Switching (OBS) Networks View project

All content following this page was uploaded by Adnan Saher Mohammed on 07 November 2022.

The user has requested enhancement of the downloaded file.


International Journal of Soft Computing and Artificial Intelligence, ISSN: 2321-404X, Volume-10, Issue-1, May-2022
http://iraj.in
PROPOSING AN EFFICIENT MODEL TO DETECT MELANOMA
BASED ON DENSE-CNN
1
MOHAMMED Y. ALKHUZAIE, 2ADNAN SAHER MOHAMMED
1,2
Department of computer engineering, Faculty of engineering, Karabuk, University-Karabuk-Turkey
E-mail: 1who224466@gmail.com, 2adnanalajeeli@karabuk.edu.tr

Abstract - Deep learning algorithms provide the best judgment aid for medical experts, potentially improving the
classification of skin lesions and early detection of skin cancer. In manycases, deep learning techniques such as neural
networks have bypassed feature extraction tools. The aim of this study is to design a neural network that has the ability to
accurately and efficiently classify images of skin lesions. In this paper, a dense convolutional neural network (Dense-CNN)
model is designed for melanoma diagnosis. The proposed model was applied to two datasets: the HAM10000 dataset and the
DermQuest-DermIS dataset. These datasets are for images of skin lesions obtained through dermatoscopy. Our results were
compared with those of other studies. The proposed model achieved high diagnostic rates in terms of accuracy, sensitivity,
specificity, and precision when applied to the HAM10000 dataset, reaching 98.03%, 98.05%, 99.67%, and 98.11%,
respectively, while the results when applied to the DermQuest-DermIS dataset reached 97.85%, 97.87%,

Keywords - Melanoma, Dermoscopy, Convolutional Neural Networks, HAM10000, Derm Quest-DermIS.

I. INTRODUCTION which represent malignant and benign of skin lesions.


In the past years, several deep learning approaches
The primary cause of skin cancer is still a mystery have been proposed. While it was limited to
[1].However, other studies link the disease‟s onset to dermatoscopy, it did not prove its competence to
factors such as prolonged exposure to bright obtain accurate results when applied to various
sunshine, hormonal and genetic changes in datasets. To deal with the problem, we proposed
individuals, and hereditary factor. According to designing a pre-trained. Dense-CNN network,
certain research, people with white skin and colourful capable of handling various datasets for skin lesions
eyes are much more likely than people with darkskin to obtain very accurate diagnostic results.
[2], [3]. Four major medical diagnostic approaches
are used To effectively assess melanoma: the 7-point
checklist, pattern analysis , the Menzies method and
ABCD guidelines. All of these approaches are
considered traditional methods done through human
effort, and they are rather good at detecting skin
cancer, but there may be errors. A high level of
quality is needed to differentiate skin lesions and
obtain more accurate diagnostic results [4]. Because Figure 1: Some of images used in datasets
of the use of conventional feature extraction methods
with very good classifiers, including SVM and ANN, II. RELATED WORKS
the CAD technique is more efficient and has
improved results. Because of the changes in the types This section summarizes and describes relevant
of dermatological disorders, the handmade feature studies on picture categorization and differentiation
extraction methodology is impracticable for diagnosis that have been done. This is accomplished with a
[5]. Feature learning is necessary to overcome the particular emphasis on new studies that used deep
issue [6], and the usage of image analysis and AI learning approaches to diagnose and classify photos
vision for various medical image analysis has of skin lesions. The need for meticulous classification
increased dramatically in the last decade [7]. Deep demarcation of skin lesions has been highlighted, as it
learning and machine learning are two of the most can aid and support the clinician in the identification
essential approaches and methodologies for process and assessment, as well as define the kind of
diagnosing diseases from medical imaging in the lesion. Although all the results that we will observe in
medical field. The use of these technologies speeds these studies will appear rather good, they cannot be
up the diagnosis process while lowering human error. compared with the results of our current study. The
It can also increase the accuracy and tolerability of reason for this is that all previous studies have used a
dermatologists‟ and doctors‟ skin cancer assessments. maximum of one or two datasets, while we have used
In the field of diagnosing images of skin lesions, the eight datasets and applied our proposed model. In a
work of artificial intelligence networks is mainly study presented by Zhang, Ni et al. [2], they used a
focused on dividing images into two categories: CNN neural network with a whale optimization
malignant and benign. Figure (1) shows some images algorithm in their investigation. They employed two

Proposing an Efficient Model to Detect Melanoma Based on Dense-CNN

70
International Journal of Soft Computing and Artificial Intelligence, ISSN: 2321-404X, Volume-10, Issue-1, May-2022
http://iraj.in
datasets for this method: DermQuest and DermIS, „pyogenic granulomas and hemorrage‟ , „akiec‟ =
which had over 22,000 dermatoscopy images of skin „Actinic Keratoses and interaepithelial carcinoma‟ ,
disorders. When this method was compared to „df‟ = „dermatofibroma‟.
datasets generated in other ways, it came out on top.
When compared to the other methods, the proposed
strategy outperformed them all. Albahar, Marwan Ali
[8]recommended developing a novel prediction
model and extracting features using the CNN-Novel
Regulaizer model. Skin lesions were split into two
categories using the ISIC data collection. The first is
benign carcinomas, while the second is malignant
Table 1 Show the distribution of images based on the type of
carcinomas. AUC-ROC prediction accuracy was dataset, and allocation of images on training-testing and
97.49%, a sensitivity of 94.3%, and a specificity of validation (before balanced).
93.6%. Lopez, Adria Romero et al. [9] they suggested
to develop VGGNet, a sort of convolutional neural 1) Data preparation: All datasets were subjected to
network, for classification and feature extraction standardization and scale data preparation. To acquire
using 1279 photos from the ISIC archive dataset. the very same set of input variables, RGB
dermatoscopy photos were standardized to [0, 1]. The
Based on transfer learning, network training, and normalizing procedure ensures that the image pixels
maintaining a learning model, three classification are equal and that the resolution is stable. Due to
algorithms were applied. The model was used in three discrepancies in the sliced skin lesions in the
different ways, with the third strategy yielding the diagnostic imaging scans, the information for the
best results. The sensitivity reached was 78.66% dermatological lesion isolation process was reduced
SertanKaymak et al. [10] trained the ISIC 2018 to 28 x 28 pixels using the np.array. Because of
dataset using the AlexNet classifier. To discover variances in the sizes of the supplied photographs, all
features and classify the photos, this diagnosis images collected by this method have indeed been
required visual identification of the skin mass. scaled.

Depending on how the dataset was partitioned, three 2) Preprocessing: It is a technique for removing
models were developed and trained. The second noise and blemishes from images, visually
model was able to distinguish between benign and superfluous data. Additional preprocessing
malignant melanoma with greater precision, with an procedures could include a spatial reduction and
accuracy of 84.4%, a sensitivity of 84.7%, and a locating regions of the image for more processing.
specificity of 83.8%. specificity. Light and noise impacts are common in skin surface
images when taken with expert image sensors, so
III. MATERIALS AND METHODS they must be removed [11], [12].These effects
include non uniform lighting and light-reflective
A. Datasets surfaces on the surface of the skin. To minimize the
Two datasets were used, which to our knowledge is impact of these variables on CNN classification, first
the most data used in providing images of skin lighting is used. On the input images, the adjustment
lesions in medical and research centers. Most of these step is carried out. At this stage, it detects lighting
types of data are freely available on the Internet, and impacts as sharp changes in the absorption and
some of them require certain subscriptions for the valuation streams for the color space. As a result,
purpose of obtaining them. The datasets used in this with a particular set of contours, the lighting impacts
research contained seven categories of skin lesions, are overlooked. This is done without trying to destroy
and some of these data contained 21 sub-categories. the initial image‟s actual factual edges.
Each dataset contains a number of images obtained
by the dermatoscopy method. 3) Data augmentation: To acquire successful
outcomes from processing vast volumes of data,
The common factor in dividing this data is that it neural networks require great practice [13]. The
contains two main categories: benign and melanoma. arrangement of images used in learning may be
HAM10000 datasets have selected all the image data inadequate or imbalanced. Reproduction is carried
in them, which contained 10015 images, while out in the dataset in order to enhance the rate of
DermIS-DermQuest dataset contained 6600 images. incomplete evidence used in learning. Photo
All images in the eight datasets were distributed as augmentation is a breeze with Keras‟ built-in
75To facilitate the classification process, each libraries. Additional data for the training dataset is
category of skin lesions is encoded with a special generated via the data replication method. It‟s done in
symbol, as shown here: „nv‟= „melanoma nevi‟ , a variety of ways, including rotating, scaling,
„mel‟ = „melanoma‟, „bkl‟ = „benign keratosis-like translating to the main image, and mixing these
lesions‟ , „bcc‟ = „maligna melanoma‟ , „vac‟= techniques[13], [14]. In this study, we use a rotation

Proposing an Efficient Model to Detect Melanoma Based on Dense-CNN

71
International Journal of Soft Computing and Artificial Intelligence, ISSN: 2321-404X, Volume-10, Issue-1, May-2022
http://iraj.in
angle variety that is random within the spectrum of 2) Implementation: This section will detail the steps
[0, particular angle. Each of the trained dermatoscopy for creating the proposed architecture, including a
images in each of the eight datasets was rotated four mention of the libraries on which the architecture is
times at 0°, 90°, 180°, and 270° angles. Figure (2) based. We will support building architecture with
illustrates the effect of data augmentation. By diagrams, tables, and illustrations with an explanation
improving the data, we boosted the pictures of the of the functions used in the proposed model. In order
basic training set in all of the datasets utilized here. to classify melanoma as malignant or benign, the
The increment in the quantity of photos in the model of Dense-convolutional neural network (CNN)
datasets is detailed in table (2). To train the model, architecture was established. The architecture consists
we divided the enhanced dataset into a training of four basic layers: the Conv2D layer, MaxPool2D
dataset and a validation dataset with a ratio of 0.8:0.2, layer, the flattening layer, and the dense layer.
respectively. Function layers consist of two layers, namely the
Relu layer and the softmax layer. After processing the
images that consist of (preprocessing, segmentation,
feature extraction, and ckassification), as shown in
figure (3). Building and organizing the basic layers
was the first phase, followed by testing the series of
training steps, activation functions, and ultimately
introducing the optimizer to the model. To test how
data augmentation affects prediction accuracy,
Figure 2: Example of melanoma Images after data Adam‟s optimizer was utilized. When we compared
augmentation. the effectiveness of the algorithm to earlier work on
the same model, we discovered that Adam‟s
optimizer is a better data optimizer than others. The
activation functions will return a maximum value of
(0, x). It will just return x if the relationship between
x and 0 is higher than 0. This function returns 0 if it is
not called. A hyperparameter is the magnitude of the
TABLE 2 Information on comprehensive data enhancement zero padding. When filters are applied to an input
matrix, the resulting volume‟s spatial size shrinks.
B. Proposed System The size of the output produced will shrink when
more convolutional layers are applied. However, in
1) Convolutional neural network: A sort of order to obtain low-level characteristics from the
convolutional neural network (CNN) that is supposed margins of images, sometimes it is preferable to keep
to work similarly to human vision. They‟re mostly as much data from the source input matrix as
utilized in deep learning image recognition possible. As a result, the output volume emphasizes is
technologies that aggregate diverging or comparable frequently kept the same as the input volume. Zero
images into representations that may be used to padding is a feature that allows you to customize the
classify items in images and videos, including faces, output volume‟s spatial size. To achieve this, it pads
individuals, or individuals with images in medical the input space with zeros all around the boundary
imaging [13]. [15]. The proposed model for the diagnosis of skin
A CNN is made up of several layers, each of which cancer is shown in figure (4).
receives the output from the previous layer. The first
layer‟s job is to recognize unprocessed features in
images, including horizontal, vertical, and diagonally-
bordered images. The following layer picks out more
complicated elements like angles or shapes formed by
collecting edges. The procedure of detecting high-
level characteristics such as objects, faces, and other
features starts as you advance from one layer to the
next. In a CNN design, there are numerous
convolution layers, a series of pooling layers, and
then contrasting normalizing layers. That there‟s at
least one fully connected layer at the finish.CNN is
much less compact than a feed-forward network and
therefore easier to train. On elevated images,
nevertheless, the greatest result of CNNs is
marginally lower and much more computationally
expensive than the feed line [11],[13],[14].

Proposing an Efficient Model to Detect Melanoma Based on Dense-CNN

72
International Journal of Soft Computing and Artificial Intelligence, ISSN: 2321-404X, Volume-10, Issue-1, May-2022
http://iraj.in

Figure 4: proposed model for detection of melanoma

IV. RESULTS Figure 5: Shows the testing-training accuracy for (A): HAM
10000 dataset, (B): DermIS-DermQuest dataset.

In this section, the measurement of accounts using the


confusion matrix application, which has four
indicators, is presented. True positive (TP), false
positive (FP), false negative (FN), and true negative
(TN) values are assigned to the indices. By applying
the mentioned equations, we were able to derive the
final values of the classification. Table No. (3) shows
all the equations that were used with a simplified
definition. The results of all datasets were represented
by four plots (confusion matrix, training test Figure 6: Shows the testing-training loss for (A): HAM 10000
accuracy, loss accuracy, and ROC). The results of dataset, (B): DermIS-DermQuest dataset
applying the proposed model and its implementation
were presented on eight datasets, as shown in table
(4). We divided the dataset into 75% for training, 5%
for validation, and 20% for testing. The results for
accuracy,specificity, sensitivity, accuracy, and
receiver operational characteristics (ROC) were
presented with a 7 × 7 noise matrix for each dataset.

After testing the Dense-CNN model on the eight


datasets, we got the following: Figure (6) represents
the plotted results for testing-training accuracy.
Figure (7) represents the plotted results for testing-
training loss. Figure (8) represents the confusion
matrix for datasets used.

TABLE 3 The performance measures that were employed in


this work, together with their description

Fig. 7. Figure 7. Figure 7. Shows the confusion matrix for the


datasets used. (A): HAM 10000, (B): DermIS-DermQuest

V. DISCUSSION

Firstly, we review the results obtained by applying


TABLE 4 Shows the results of accuracy(Acc), sensitivity(Sen),
specificity(Spe), precision (Pre) and AUC for all datasets used
the model to two datasets (HAM10000 and
according to apply proposed model DermISDermQuest). Secondly, we compare our
Proposing an Efficient Model to Detect Melanoma Based on Dense-CNN

73
International Journal of Soft Computing and Artificial Intelligence, ISSN: 2321-404X, Volume-10, Issue-1, May-2022
http://iraj.in
findings to those of related studies that dealt with a
skin cancer diagnosis by applying them to a variety of
models on a variety of datasets. Here we are trying to
focus mainly on the studies that dealt with the
implementation of neural networks in all their
models. We believe that a comparison of the CNN
results used in previous studies and our own, Dense-
CNN, will give more reliable results. Through the
results, it is clear to us that our model has achieved
results of accuracy, specificity, sensitivity, precision,
and AUC much better than the results of previous
studies. The experimental results of the proposed TABLE 6 The comparison of results with studies that used the
DERMQUEST-DERMIS dataset
model showed that the accuracy of the model was
98.03% when our model was applied to the HAM
VI. CONCLUSION
10000 dataset, while the highest accuracy score
achieved by previous studies by reviewing four
A new improved technique for identifying melanoma
studies when applied to the same dataset was 97.78%.
from the input images is proposed in this work. A
convolutional neural network was used to create the
The experimental results of the proposed model
approach. A dense layer algorithm is used to enhance
showed that the accuracy of the model was 97.85%
the efficiency of CNN. To eliminate errors in the
when our model was applied to the DermQuest-
actual network outputs, the efficient algorithm used to
DermIS dataset, while the highest accuracy score
select the optimal weights and biases in the network
achieved by previous studies by reviewing four
is applied. With thorough study and deep data
studies when applied to the same dataset was 96.3%.
comprehension, the proposed technique was
The reason for the good results that we got, we
successful in detecting and predicting melanoma with
believe, is due to several factors: 1- The designed
good outcomes. We put the proposed method through
model‟s strength, as it contains several layers of
its paces on two datasets (HAM10000 and DermIS-
con2D 2- The presence of three dense layers with the
DermQuest). The results we obtained indicate that the
activation functions RELU and Softmax.3-The
proposed technique achieved the highest accuracy of
Adam‟s optimizer was utilized. 4-Not using the
98.03% when applied to the HAM10000 dataset. The
transfer learning method because we have created and
results obtained were compared with those of
trained the model.
previous investigations, with accuracy, specificity,
sensitivity, precision, and AUC serving as measures
This is knowing that the number of epochs is 50,
of performance. The final results showed that the
which acts to train the data. In addition, we tried as
application of the proposed model has achieved the
much as possible to use datasets containing large
best results in diagnosing skin cancer, and this
numbers of skin lesions image data. The use of big
accuracy has outperformed all the results of previous
data allows us to get good results so that we can make
studies that compared it with it. For future work, we
comparisons with other results. The tables (5 and 6)
suggest dividing the major skin lesions into several
represent the results of all the comparisons we made
subcategories to ensure the best classification of skin
with other related work results.
lesions.

REFERENCE

[1] D. R. Brenner, A. E. Poirier, S. D. Walter, W. D. King, E.


L. Franco, P. A. Demers, P. J. Villeneuve, Y. Ruan,
F.Khandwala, X. Grevers et al., “Estimating the current and
future cancer burden in canada: methodological framework
of the canadian population attributable risk of cancer
(compare) study,” BMJ open, vol. 8, no. 7, p. e022378,
2018.
[2] N. Zhang, Y.-X. Cai, Y.-Y. Wang, Y.-T. Tian, X.-L. Wang,
and B. Badami, “Skin cancer diagnosis based on optimized
convolutional neural network,” Artificial intelligence in
medicine, vol. 102, p. 101756, 2020.
[3] C. E. DeSantis, K. D. Miller, W. Dale, S. G. Mohile, H.
J.Cohen, C. R. Leach, A. Goding Sauer, A. Jemal, and R.
L.Siegel, “Cancer statistics for adults aged 85 years and
older, 2019,” CA: a cancer journal forclinicians, vol. 69, no.
6, pp. 452–467, 2019.
[4] C. Carrera, M. A. Marchetti, S. W. Dusza, G. Argenziano,
TABLE 5 The comparison of results with studies that used the R.P. Braun, A. C. Halpern, N. Jaimes, H. J. Kittler, J.
HAM 10000 dataset Malvehy,S.W . Menzies et al., “Validity and reliability of

Proposing an Efficient Model to Detect Melanoma Based on Dense-CNN

74
International Journal of Soft Computing and Artificial Intelligence, ISSN: 2321-404X, Volume-10, Issue-1, May-2022
http://iraj.in
dermoscopic criteria used to differentiate nevifrom [13] N. Razmjooy, B. S. Mousavi, and F. Soleymani, “A real-
melanoma: a web-based international dermoscopy society time mathematical computer method for potato inspection
study,” JAMA dermatology,vol. 152, no. 7, pp. 798–806, using machine vision,” Computers & Mathematics with
2016. Applications, vol. 63, no. 1, pp. 268– 279, 2012.
[5] A. Saez, B. Acha, and C. Serrano, “Pattern analysis in [14] R. Garg, S. Maheshwari, and A. Shukla, “Decision support
dermoscopic images,” in Computer vision techniques for the system for detection and classification of skin cancer using
diagnosis of skin Cancer. Springer, 2014, pp. 23–48. cnn,” in Innovations inComputation
[6] Y. Bengio, A. Courville, and P. Vincent, “Representation [15] E. O. Molina-Molina, S. Solorza-Caldero´n, and J. A´
learning: A review and new perspectives,” IEEE lvarez-Borrego, “Classification of dermoscopy skin lesion
transactions onpattern analysisand machine intelligence, color- 10,images using fractaldeep learning features,”
vol. 35, no. 8, pp.1798–1828, 2013. Applied Sciences,vol. no. 17, p. 5954, 2020.
[7] M. Chen, Y. Hao, K. Hwang, L. Wang, and L. [16] A. H. Shahin, A. Kamal, and M. A. Elattar, “Deep ensemble
Wang,“Disease prediction by machine learning over big learning for skin lesion classification from dermoscopic
data from Healthcare communities,” IeeeAccess, vol. 5, pp. images,” in 2018 9th CairoInternational Biomedical
8869–8879,2017. EngineeringConference (CIBEC). IEEE, 2018, pp. 150–153.
[8] M. A. Albahar, “Skin lesion classification using [17] H. Shahin, A. Kamal, and M. A. Elattar, “Deep ensemble
convolutional neural network with novel regularizer,” IEEE learning for skin lesion classification from dermoscopic
Access, vol. 7,pp. 38306–38 313,2019. images,” in 2018 9th CairoInternational Biomedical
[9] A. R. Lopez, X. Giro-i Nieto, J. Burdick, and O. EngineeringConference (CIBEC). IEEE, 2018, pp. 150–153.
Marques,“Skin lesion classification from dermoscopic [18] J.-A. Almaraz-Damian, V. Ponomaryov, S.
images using deep learning techniques,” in 2017 13th Sadovnychiy,and H. Castillejos-Fernandez, “Melanoma and
IASTED international conference on biomedical nevus skin lesion classification using handcraft and deep
engineering(BioMed). IEEE, 2017,pp.49–54. learning no. 4,feature fusion via mutual information
[10] S. Kaymak, P. Esmaili, and A. Serener, “Deep learning for measures,” Entropy,vol. 22, p. 484, 2020.
twostep classification of malignant pigmented skin [19] K. M. Hosny, M. A. Kassem, and M. M.
lesions,”in 2018 14thSymposium on Neural Networks and Foaud,“Classification of skin lesions using transfer learning
Applications(NEUREL). IEEE, 2018, pp. 1–6. and augmentation with alex-net,” PloSone, vol. 14, no. 5,
[11] M. H. Jafari, N. Karimi, E. Nasr-Esfahani, S. Samavi, S. p.e0217293, 2019.
M.R.Soroushmehr, K. Ward, and K. Najarian, “Skin lesion [20] R. Sarkar, C. C. Chatterjee, and A. Hazra, “Diagnosis of
segmentation in clinical images using deep learning,” in melanoma from dermoscopic images using a deep
23rdInternationalconference on pattern recognition depthwise separable residual convolutional network,” IET
(ICPR).IEEE 2016, 2016, pp. 337–342. Image Processing, vol. 13, no. 12, pp. 2130–2142, 2019.
[12] S. H. Kassani and P. H. Kassani, “A comparative study of
deep learning architectures on melanoma detection,” Tissue
and Cell, vol. 58, pp. 76– 83, 2019.



Proposing an Efficient Model to Detect Melanoma Based on Dense-CNN

75

View publication stats

You might also like