Skin Cancer Classification using Deep Learning Models

Marwa Kahia

, Amira Echtioui

, Fathi Kallel

and Ahmed Ben Hamida

ATMS Lab, Advanced Technologies for Medicine and Signals, ENIS, Sfax University, Sfax, Tunisia

Keywords: Melanoma, Diagnosis, VGG16, Skin Cancer, InceptionV3.

Abstract: In recent years, researches proved that Melanoma is the deadliest form of skin cancer. In the early stages, it

can be treated successfully with surgery alone and survival rates are high. A large number of methods for

Melanoma classification has been proposed to deal with this problem, but although they did not find better

ways to create the final solution. Thus, our aim is to go further and explore the classic models in order to

handle the Melanoma classification problem based on modified VGG16 and modified InceptionV3. The

conducted experiments revealed the effectiveness of our proposed method based on modified VGG16 with

73.33% of accuracy, when compared to other state-of-the-art methods on the same data sets, in terms of

finding optimal and effective solutions and improving the objective function.

1 INTRODUCTION

Melanoma is the most unsafe form of skin cancer. It

begins in the melanocytes (color- producing cells

plant in the surface subcaste of the skin). In the utmost

of cases, it's caused by ultraviolet radiation from sun

or tanning beds which produce mutations (inheritable

blights) that take the skin cells to expand fleetly and

form nasty excrescences (l. Argenziano, et al., 2000).

Melanoma causes 55 500 cancer deaths annually

which is 0.7 of all cancer deaths. The prevalence and

mortality rates of carcinoma differ from one country

to another due to the variation of ethnical and ethnical

groups (Schadendorf et al., 2018). Nasty carcinoma is

presumptive to come one of the most common nasty

excrescences in the future, with yet a ten times

advanced prevalence rate (Tadeusiewicz et al., 2010).

Visual examination of the suspicious skin area is

generally adopted by dermatologist as a first step for

the diagnosis of a malignant lesion. In fact, an

accurate diagnosis is essential because of the

resemblances of some lesion types. Furthermore, the

diagnostic accuracy correlates strongly with the

professional experience of the physician

(Tadeusiewicz et al., 2010).

On the other hand, without any further technical

support, dermatologists have a 65% to 80% accuracy

https://orcid.org/0000-0001-5255-9568

https://orcid.org/0000-0003-2041-1301

https://orcid.org/0000-0001-7986-8395

https://orcid.org/0000-0001-6713-7384

rate in melanoma diagnosis. In suspicious cases,

dermatologists explore and use dermatoscopic

images as a complementary support of the visual

inspection. In fact, the combination of both visual

inspection and dermatoscopic images eventually

results in an absolute melanoma detection accuracy of

75%-84% by dermatologists (Brinker et al., 2018)

Currently, artificial intelligence (AI) has come an

aptitude to face these problems. Several deep-literacy

infrastructures like reccurent neural networks (RNN),

convolutional neural networks (CNN), deep neural

networks (DNN), long short term memory (LSTM) are

proposed in literature to descry cancer cell. These

models are also successfully performed in classifying

skin cancer.

Several CNN architectures, like ResNet,

Inception and Xception, as well as VGG16, are

proposed in literature and specially designed for

image classification. Numerous researchers have

developed methods based on deep learning to classify

and identify skin cancer (Le et al., 2020; Garg et al.,

2019; Guan et al., 2019; Nugroho et al., 2019;

Pacheco et al., 2019).

In this work, we propose a modified InceptionV3

model for the classification of skin cancer. We

propose also a modified VGG16 model which

classifies skin cancer with a better accuracy value

554

Kahia, M., Echtioui, A., Kallel, F. and Ben Hamida, A.

Skin Cancer Classiﬁcation using Deep Learning Models.

DOI: 10.5220/0010976400003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 1, pages 554-559

ISBN: 978-989-758-547-0; ISSN: 2184-433X

Figure 1: Flowchart of the proposed method for skin cancer classification using modified VGG16 model.

compared to the state of the art.

The rest of the paper is organised as follows:

Section 2 details materials and proposed method.

Section 3 represents results and discussion. Section 4

concludes this paper.

2 MATERIAL AND PROPOSED

METHOD

In this section, we will present the dataset used in this

research work and present our proposed method for

skin cancer classification.

2.1 Dataset Description

The used dataset in this present work contains three

classes: melanoma, nevus and seborrheic keratosis.

More details about this datasets are given below:

 2000 training images

(https://s3-us-west-1.amazonaws.com/udacity-

dlnfd/datasets/skin-cancer/train.zip)

- melanoma images: 374

- nevus images: 1372

- seborrheic keratosis images: 254

 150 validation images

(https://s3-us-west-1.amazonaws.com/udacity-

dlnfd/datasets/skin-cancer/valid.zip)

 600 testing images

(https://s3-us-west-1.amazonaws.com/udacity-

dlnfd/datasets/skin-cancer/test.zip)

2.2 Proposed Method

Figure 1 presents Flowchart of the proposed method.

A preprocessing stage is firstly applied on input

image. The preprocessing involves resizing all

images and increasing the number of images from

both classes melanoma and seborrheic keratosis.

Then we test the modified VGG16 model and apply

our modified InceptionV3 model.

2.2.1 Data Augmentation

We used data augmentation techniques to artificially

boost the amount of our training data because our

data collection is rather small. The increase in data is

an often-applied DL method that generates the

required number of samples. It also improves

network efficiency for a small database by

optimizing it. Shifting, Rotation, flipping,

transformation, and zooming are all examples of

traditional data augmentation procedures. We used

“Keras Image Data Generator” to apply image

augmentations during training in this investigation.

As shown in section 2.1, the number of images of

class 'Nevus' is 1372. In order to balance the number

of images for all three considered classes, we applied

the data augmentation technique to augment the size

of both classes 'Melanoma' and 'seborrheic keratosis'.

In this work, we choose a vertical flip, a

horizontal flip and a 45-degree rotation for data

augmentation. As a result, we got 1372 images for

each class.

2.2.2 Skin Cancer Classification using

Modified VGG16 Model

Figure 2 shows the flowchart of the proposed method

for the classification of skin cancer using the VGG16

model. In this paper, modified VGG16 begin by five

blocks, the first two blocks include two convolutional

layers with a Relu activation function and Max

Pooling followed by three blocks. Each block enclose

three convolutional layers with a Relu activation

function and Max Pooling. An adaptative Avg

Skin Cancer Classiﬁcation using Deep Learning Models

555

Figure 2: Flowchart of the modified VGG16 for skin cancer classification.

Figure 3: Flowchart of the modified InceptionV3 for skin cancer classification.

Pooling and two blocks follow these blocks. Each

block contains linear layer, ReLu activation function,

and Dropout Layer. Finally, a linear layer is used to

predict the class of images.

We fine-tuned this model by 10 epochs. The

Adaptive Moment Estimation known as “Adam

optimizer” is used to optimize the loss function. The

adopted model is trained by a cross-entropy loss

function.

2.2.3 Skin Cancer Classification using

Modified InceptionV3 Model

Figure 3 shows the modified method for the

classification of skin cancer using the InceptionV3

model. InceptionV3 is a commonly used image

classification model that has demonstrated more than

78.1% accuracy on the ImageNet dataset. The model

itself is made up of basic symmetric and asymmetric

components including convolutions, average pooling,

maximum pooling, concatenations, drops, and fully

connected layers. Batch normalization is widely used

in the model and applied to activation inputs. The loss

is calculated via SoftMax. Our Modified InceptionV3

begins by three blocks of BasicConv2d. Each block

includes a convolutional layer and a batch

normalization step followed by 3 Modules A, module

B, 4 modules C, module D, and 2 modules E followed

by Avg Pooling, Dropout, Linear layer, ReLu,

Dropout layer and Linear layer.

3 RESULTS AND DISCUSSION

In this section, we present and discuss the obtained

classification results when both proposed models are

used. Accuracy, precision, recall and F1-score

metrics are considered for performance evaluation of

proposed classifiers. These mentioned metrics are

respectively computed according to the following

SDMIS 2022 - Special Session on Super Distributed and Multi-agent Intelligent Systems

556

equations for both modified VGG16 and modified

InceptionV3 models.

𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 





(1)

𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 





(2)

𝑟𝑒𝑐𝑎𝑙𝑙 





(3)

𝐹1  𝑠𝑐𝑜𝑟𝑒  2 ∗

∗



(4)

where TP, TN, FP and FN are respectively the True

Positive, True Negative, False Positive and False

Negative.

Both modified VGG16 and modified InceptionV3

algorithms assess the classification performance. We

achieved two experiments using the same described

dataset. We conducted the first classification

experiment considering all melanoma, nevus and

Seborrheic keratosis classes. The second

classification experiment is executed considering

only two classes: benign and malignant classes.

3.1 Classification Results: Three

Classes

In this section, we present the obtained classification

results when the three classes are considered. Table 1

presents the average accuracy results of all considered

classes for both modified VGG16 and modified

InceptionV3 models.

Table 1: Classification accuracy.

Accuracy

Modified VGG16

73.33%

Modified InceptionV3

42.00%

Table 2 details the accuracy results obtained with

three considered classes for both modified VGG16

and modified InceptionV3 models.

Table 2: Classification accuracy for three classes.

Modified

VGG16

Modified

InceptionV3

melanoma 50% 33%

nevus 54% 84%

Seborrheic keratosis 47% 24%

From tables 1, we can observe that modified

VGG16 model performs better than the modified

InceptionV3 model. In fact, the average accuracy

value obtained with modified VGG16 model is better

(73.33%) than those obtained with modified

InceptionV3 model (only 42%).

Table 2 showed that both proposed methods

present good classification performances for 'Nevus'

class with a superiority for modified InceptionV3

model. In fact, this class achieves an accuracy value

of 54% with modified VGG16 and 84% with modified

InceptionV3. However, classification performances

using both proposed methods are significantly

decreased for 'Seborrheic keratosis' class. In this

case, accuracy values are only limited to 47% and

24% for modified VGG16 and modified InceptionV3

models respectively.

3.2 Classification Results: Two Classes

In this section, we present the obtained classification

results when the two benign and malignant classes are

considered. Figure 4 shows the confusion matrix and

the ROC curves for both Modified VGG16 model.

Figure 4: Confusion matrix and ROC curve for modified

VGG16 model.

Figure 5 shows the confusion matrix and the ROC

curves for both Modified InceptionV3 model.

Skin Cancer Classiﬁcation using Deep Learning Models

557

Figure 5: Confusion matrix and ROC curve for modified

InceptionV3 model.

Table 3 reports the average results for recall,

precision and F1-score metrics computed using both

proposed VGG16 and InceptionV3 models.

Table 3: Classification performances for Malignant and

Benign classes.

Modified

VGG16

Modified

Ince

tionV3

Recall 51.35% 58.33%

Precision 95.00% 70.00%

F1-score 66.66% 63.63%

The binary classification of Malignant and Benign

classes also show that the proposed method based on

the VGG16 model achieves better performances then

the second proposed method based on InceptionV3

model. In fact, considering the proposed VGG16

model, recall, precision and F1-score values are

respectively equal to 51.35%, 95.00%, and 66.66%.

3.3 Discussion

The performances of the modified VGG16 model are

compared to three state of the art methods labelled as

KNN (Daghrir et al., 2020), SVM (Daghrir et al.,

2020) and AlexNet (Sasikala et al., 2020). Results are

summarized in Table 4.

Table 4: Comparative study for binary classification.

Method Accuracy

KNN (Da

hrir et al., 2020) 57.3%

SVM (Daghrir et al., 2020) 71.8%

AlexNet (Sasikala et al., 2020) 65.3 %

Proposed method based on

modified VGG16

73.33%

By comparing the accuracy values listed in Table

4 obtained for different considered methods, we can

observe that our modified VGG16 method performs

better than KNN, SVM, and AlexNet methods. In fact,

accuracy reached 73.33% with our proposed VGG16

method. Although the accuracy is limited to 57.3%,

the KNN method is able to hardly identify malignant

skin lesions since it is sensitive to outliers.

On the other hand, the SVM method performs

better than the KNN and AlexNet methods due to its

adaptability and efficiency. In fact, accuracy is equal

to 71.8% with SVM method, but it is limited to only

57.3% and 65.3% with KNN and AlexNet methods

respectively. Although AlexNet achieved quiet

performance, the SVM is still considered a more

robust and powerful tool for identifying skin cancer.

4 CONCLUSIONS

In this work, we proposed two modified models for

skin cancer classification: modified VGG16 and

modified InceptionV3 models. The application of the

data augmentation showed that the reduction of the

data imbalance can be useful to improve classification

performance, but careful tuning is required, for

example, to make the data perfectly balanced training

does not necessarily result in a better model.

Performances are evaluated using different

metrics like accuracy, precision, recall and F1-score.

Two experiments are conducted. In the first

experiment, we considered melanoma, nevus and

Seborrheic keratosis classes, but in the second one,

only benign and malignant classes are considered.

Results of first experiment showed that the modified

VGG16 is a reliable multiple classifier and performs

better than modified InceptionV3 model. For second

experiment, compared to state of the art considered

methods, results showed that better accuracy values

are obtained for binary classification using modified

VGG16 model.

SDMIS 2022 - Special Session on Super Distributed and Multi-agent Intelligent Systems

558

It is clear that our proposed method given better

results compared to different others recent methods.

However, there is a need to improve its performances

in our future work. In fact, merging or concatenating

deep learning models could improve the classification

results.

REFERENCES

Brinker, T.S., Hekler, A., Utikal, J., Grabe, N.,

Schadendorf, D., Klode, J., Berking, C., Steeb, T., Enk,

A., & von Kalle, C., Skin Cancer Classification Using

Convolutional Neural Networks: Systematic Review J

Med Internet Res. 2018 Oct; 20(10): e11936

Daghrir, J., Tlig, L., Bouchouicha, M., Sayadi, M.

Melanoma skin cancer detection using deep learning

and classical machine learning techniques: A hybrid

approach. International Conference on Advanced

Technologies for Signal and Image Processing, Sep

2020.

Garg, R., Maheshwari, S., & Shukla, A. (2019). Decision

support system for detection and classification of skin

cancer using CNN. arXiv preprint arXiv:1912.03798.

Guan, Q., Wang, Y., Ping, B., Li, D., Du, J., Qin, Y., et al.

(2019). Deep convolutional neural network VGG-16

model for differential diagnosing of papillary thyroid

carcinomas in cytological images: a pilot study. Journal

of Cancer, 10(20), 4876.

Argenziano, G., Soyer H. P., De Giorgio, V., et al.

Interactive Atlas of Dermoscopy Milano, Italy: Edra

Medical Publishing and New Media; 2000.

Le, D. N., Le, H. X., Ngo, L. T., & Ngo, H. T. (2020).

Transfer learning with classweighted and focal loss

function for automatic skin cancer classification. arXiv

preprint arXiv:2009.05977.

Nugroho, A. A., Slamet, I., & Sugiyanto (2019). Skins

cancer identification system of HAMl0000 skin cancer

dataset using convolutional neural network. AIP

Conference Proceedings, 2202, Article 020039.

Pacheco, A. G., & Krohling, R. A. (2019). Recent advances

in deep learning applied to skin cancer detection. arXiv

preprint arXiv:1912.03280.

Sasikala, S., Arun Kumar, S., Shivappriya, S.N. and

Priyadharshini, T. (2020). Towards Improving Skin

Cancer Detection Using Transfer Learning. Biosc.

Biotech. Res. Comm. Special Issue Vol 13 No 11 (2020)

Pp-55-60.

Schadendorf, D., Alexander CJ van Akkooi, Carola

Berking, Klaus G Griewank, Ralf Gutzmer, Axel

Hauschild, Andreas Stang, Alexander Roesch, and

Selma Ugurel. Melanoma. The Lancet,

392(10151):971– 984, 2018.

Tadeusiewicz, R. (2010). Place and role of intelligent

systems in computer science. Computer Methods in

Materials Science;10(4), pp.193-206,2010.

Skin Cancer Classiﬁcation using Deep Learning Models

559