Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 15 September 2022 doi:10.20944/preprints202209.0215.

v1

Article

Skin Cancer Classification and Comparison of Pretrained Mod-

els Performance using Transfer Learning
Subroto Singha 1* and Priyangka Roy 2,

1 Division of Enrollment Management, Texas A&M University-Commerce, 2600 W Neal St., Commerce, TX
75428, USA; ssingha@loemail.tamuc.edu
2 Department of Marketing & Business Analytics, Texas A&M University-Commerce, 2600 W Neal St., Com-

merce, TX 75428, USA; proy@leomail.tamuc.edu

* Correspondence: ssingha@loemail.tamuc.edu; Tel.: +1-903-886-5076

Abstract: Skin cancer is an uncommon but serious malignancy. Dermoscopic images examination
and biopsy are required for cancer detection. Deep learning (DL) is extremely effective in learning
characteristics and predicting malignancies. However, DL requires a large number of images to
train. Image augmentation and transferring learning were employed to overcome the lack of images
issue. In this study we divided images into two categories: benign and malignant. To train and test
our models, we used the public ISIC 2020 database. Melanoma is classified as malignant in the ISIC
2020 dataset. Along with categorization, the dataset was studied to demonstrate variation. The per-
formance of three top pretrained models was then benchmarked in terms of training and validation
accuracy. Three optimizers were employed to optimize the loss: RMSProp, SGD, and ADAM. Using
ResNet, VGG16, and MobileNetV2, we obtained training accuracy of 98.73%, 99.12%, and 99.76%,
respectively. Using these three pretrained models, we attained a validation accuracy of 98.39%.

Keywords: pretrained model; transfer learning; skin cancer; deep learning; ISIC 2020

1. Introduction

One of the most frequent types of cancer is skin cancer. Melanoma is responsible for
75% of skin cancer fatalities, according to the American Cancer Society. As a result, der-
matologists examine each patient's moles for melanoma, the most common kind. UV
radiation causes melonocyte cells in the human body to be damaged, leading in mela-
noma, a kind of skin cancer. Skin cancer is the most frequent malignancy, accounting for
one-third of all malignancies globally, according to the World Health Organization
(WHO). Basal cell carcinoma (BCC), squamous cell carcinoma (SCC), and malignant
melanoma are the most common kinds (MM). Non-melanocytic cancer is caused by BCC
and SCC, and the majority of skin cancers are non-melanocytic, although melanoma is
the deadliest because it spreads quickly if not diagnosed and treated early. Thus, several
organs in the human body, such as malignant genes, hair color, and an increased inci-
dence of benign melanocytic nevi and dysplastic nevi, are implicated in Melanoma diag-
nosis. Skin cancer is really caused by abnormal melanocyte cell growth, which affects
surrounding tissues by multiplying and spreading through lymph nodes [1]. Skin cancer
is presently a major public health concern, with over 123,000 new cases diagnosed each
year worldwide. Surprisingly, melanoma accounts for 75% of skin cancer mortality. Fur-
thermore, according to the American Cancer Society, 100,000 new instances of mela-
noma will be detected by the end of 2020, with 7,000 people dying from the condition.
Melanoma is responsible for around 9000 deaths in the United States each year [2]. How-
ever, if this cancer is not detected early, the cost of treatment is significant, costing over
$134,000 in its fourth stage [3]. Dermatologists first detect melanoma by examining pho-
tos of the cancer and moles, among other treatment options. They also check for "ugly
ducklings" or outlier lesions around the moles that might be melanoma. If they're lucky,
the outcomes will be precise and accurate. As a result, artificial intelligence (AI) has the
potential to assist physicians in accurately identifying melanoma. AI-based detection

© 2022 by the author(s). Distributed under a Creative Commons CC BY license.

Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 15 September 2022 doi:10.20944/preprints202209.0215.v1

technologies have the potential to increase dermatological accuracy. Existing AI ap-

proaches have not sufficiently considered this clinical environment. Dermatologists may
increase their diagnostic accuracy if detection algorithms use into account "contextual"
images inside the same patient to determine which shots depict a melanoma. If the AI
detects skin cancer, it will have a significant influence on dermatological clinic opera-
tions. Convolutional Neural Network (CNN) has previously performed admirably in
image classification tasks such as breast cancer image classification [4], hyperspectral
image classification [5], pap smear image classification [6], skin cancer and lesion classi-
fication [7 - 9], and others. Even when taught with hundreds or even tens of thousands
of images, CNNs are remarkably accurate at identifying medical images. Skin lesion
photographs, in our situation, are sufficient for skin cancer categorization into two bi-
nary classes: benign and malignant. In this study our contributions are as follows:
• To get the best accuracy, we used three pretrained models VGG, ResNet, and Mo-
bileNetV2;
• We utilized the ISIC 2020 database [10], a 48.9 GB dataset of 33,126 jpeg pictures in
1024x1024 pixels, 533 of which are malignant and the rest benign;
• We used SGD, ADAM, and RMSProp three optimization algorithms to minimize the
loss;
• To solve the overfitting concerns, we applied a dropout layer and data augmentation;
• We evaluated the model’s performance using accuracy;
• We analyzed the ISIC dataset in order to see location of skin cancer, type of disease
that causes cancer etc.

2. Materials and Methods

2.1 Background

Skin cancer was first discovered using human eyes in the early twentieth century.
[11] identified a few examples of such approaches, including size, bleeding, and ulcera-
tion. This procedure, however, depends mostly on physicians' eye examinations rather
than any sophisticated or advanced technology. Another detection approach, dermos-
copy of epiluminescence microscopy, has been demonstrated to be 75% to 85% accurate
[12]. A biopsy was conducted to see if a practitioner was unable to determine whether a
mole is melanoma or not [11]. Several CNN-based classifiers have recently entered the
picture, aiding dermatologists in better identifying melanoma. In fact, [3] identified skin
lesion images as benign or malignant with 95.23 percent accuracy. Their CNN was built
with four convolutional layers, the ReLu activation function, and a softmax classifier.
They used the ADAM and SGD algorithms to decrease neural network loss. They em-
ployed SGD and ADAM to add noise to increase accuracy. Using ISIC 2018 skin lesion
pictures, the proposed classifier was trained and assessed. [13] proposed a new model
that classifies skin lesions as benign or malignant using a novel regularizer technique.
Their binary classifier could distinguish between benign and malignant cancers in im-
ages. They observed that the AUCs for nevus vs melanoma lesion, seborrheic keratosis
versus basal cell carcinoma lesion, seborrheic keratosis versus melanoma lesion, and so-
lar lentigo versus melanoma lesion were 0.77, 0.93, 0.85, and 0.86, respectively. Their
model has an average accuracy of 97.49 percent. Their technique benefited doctors in
categorizing various skin lesions. [2] developed a CNN architecture for skin lesion
grouping to attain excellent dermoscopy picture group precision. They employed an
approach that merged the group layers of four different deep neural network topologies.
In terms of accuracy, their results indicated that their technique beat the other CNNs. In
this study, they developed a new CNN model based on a novel regularizer. Further-
more, [8] assessed the efficacy of their CNN on 21 board-certified dermatologists utiliz-
ing biopsy-proven clinical images and two critical binary grouping use cases. Their deep
learning CNN surpassed dermatologists in detecting skin cancer using dermoscopy and
digital imaging. In addition, [14] suggested a well-performing automated computerized
Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 15 September 2022 doi:10.20944/preprints202209.0215.v1

system, and their technique included learning. They utilized 2000 pictures from ISIC
2017 and attained an accuracy rate of 93.6 percent.

2.1 Methodology

This section discusses data preparation, data augmentation, the architectures of

pretrained models, and evaluation metrics. Three pretrained models' performance was
compared using ISIC 2020 images.

2.1.1. Data Preprocessing

Labeling the images is the first stage in data preparation. Each image in the ISIC
2020 dataset was labeled with a benign or malignant target. We divided the dataset into
two parts: train and validation. The pretrained models were trained using a train set. We
employed an 80/20 data split, with 80% of images used for training and 20% used for
validation. The final stage in data preparation was to rescale the images from 0 to 1. Be-
cause we employed RGB images with a pixel range of 0 to 255, rescaling images de-
creased training time and eliminated image pixel inconsistencies.

2.1.2. Data augmentation

Data augmentation is a strategy for correcting data imbalance. The most prevalent
approaches used in data augmentation are oversampling and undersampling. To correct
the class imbalance, oversampling uses exact copy duplicates or modified copies of the
original data from the minority class. To execute data augmentation at random, we uti-
lized rotation range=20, width shift range=0.2, height shift range=0.2, and horizontal
flag=True. Data augmentation was utilized exclusively in the training dataset to keep the
model impartial, whereas the testing dataset was kept unmodified save for rescaling the
pictures between 0 and 1.

2.1.3. Pretrained model’s architectures

Convolutional Neural Networks (CNN) are excellent in image classification and

detection. CNNs generate feature maps by stacking convolutional layers and pooling
layers. Feature maps are high-level features derived from CNNs at each level of convo-
lution. As an output layer for picture categorization, a final fully linked layer is em-
ployed. However, the layer stack in CNNs can vary greatly, and there is no previous
knowledge of what is the best architecture for a given dataset. Furthermore, a lack of
data may result in incorrect classification, and training a CNN might take a long time to
converge. Thus, transfer learning is a solution in which we employ pretrained models
that have previously been trained on some dataset and we can simply use the pretrained
weights and put our layers on top of the pretrained architecture as per our issue needs.
We used three common pretrained models in our methodology: ResNet, VGG, and Mo-
bileNet. Imagenet was used to train all of these models. Four layers were added: a flat-
ten layer, a dense layer, a dropout layer, and an output dense layer. With a learning rate
of 0.0001, models were optimized using RMSProp, SGD, and ADAM. We utilized an
early stopping strategy to avoid overtraining the models, in which we stopped training
the models when the validation accuracy reached its maximum and was no longer im-
proving. We trained each epoch for 100 steps, and our loss function was binary crossen-
tropy. Pretrained models such as VGG, ResNet, and MobileNetV2 are frequently utilized
in transfer learning for feature extraction or fine tuning. Each model is trained on
ImageNet, abbreviated ILSVRC for ImageNet Large Scale Visual Recognition Challenge.
The goal of this image classification challenge is to train a model that can properly clas-
sify an input image into 1,000 unique item categories. Models are trained using 1.2 mil-
lion training photographs, 50,000 validation photos, and 100,000 testing photos. These
Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 15 September 2022 doi:10.20944/preprints202209.0215.v1

1,000 image categories illustrate object classifications we encounter in our daily lives,
such as dog and cat breeds, household products, automobile types, and so on. As a re-
sult, pretrained models are highly excellent at extracting features with high accuracy
while requiring less training time.

2.1.4. VGGNet Architecture

The University of Oxford's Visual Geometry Group (VGG) releases VGGNet, in

which they obtained top performance in the ImageNet dataset, which has 1000 classes.
The key contribution of VGGNet was its smaller filter (3x3) in comparison to other com-
parable architectures such as AlexNet (11x11). Furthermore, VGGNet demonstrated that
increasing the depth of the architecture is more important than expanding the width in
order to enhance performance. VGG16 and VGG19 are two versions of VGGNets, with
VGG19 being more expensive to train due to its more complex design and more parame-
ters. We utilized VGG16 to train the ISIC 2020 dataset in our work.

2.1.5. ResNet

Other designs introduced deep networks by increasing network depth. Deeper net-
works introduced an issue known as vanishing gradient, which decreases neural net-
work efficiency. To tackle the vanishing gradient problem, ResNet was proposed, with
the key idea being to bypass one or more layers dubbed 'identity shortcut connections.'

2.1.6. MobileNetV2

Google introduced the MobileNetV2 concept. Because of its lightweight and mini-
mal complexity, this architecture is appropriate for mobile devices. Version 1 of Mo-
bileNet featured depthwise separable convolution, whereas version 2 introduced a supe-
rior module called inverted residual.

2.1.7. Evaluation Metrics

We used accuracy as an assessment criterion to compare the performance of the

pretrained models on the ISIC 2020 dataset. We did not evaluate additional metrics like
F1 score, Precision, Recall, and so on since our major purpose was to assess the perfor-
mance of pretrained models to determine if the method eliminates the need to design
neural network architecture from scratch and achieves high accuracy in a short training
period.
(TP + TN)
Accuracy (%) = ∗ 100
(TP + TN + FP + FN)
(1)
,

3. Results

3.1. Dataset

We utilized a freely available dataset from the International Skin Imaging

Collaboration [10]. There are 33,126 dermoscopic training pictures and 10,982 test
images in the dataset. The training dataset is skewed, which is expected given the nature
of medical data. There were 32542 benign images and 584 malignant images among all
training data. That suggests that just 1.7% of the patients had aggressive cancer. The
torso was the location of skin cancer in 51.7% of cases. Figure 1 depicts the position of
the skin from which the images were taken.
Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 15 September 2022 doi:10.20944/preprints202209.0215.v1

Figure 1. Location of skin.

Dermatologists classified benign and malignant skin tumors into nine categories. Figure 2 depicts
the disease's eight classifications; the unknown class was not included.

Figure 2. Various classes of disease.

86.5% of the individuals with moles had nevus disease, which was classed as benign. Melanoma
patients were diagnosed with malignant cancers. Table 1 explains how nine types of disorders are
classified as benign or malignant skin malignancies. Figure 3 shows sample images of the benign
and malignant classes.

Table 1. Disease classes divided into benign and malignant.

Disease Classes Benign Malignant

atypical melanocytic prolifer-
1 0
ation
cafe-au-lait macule 1 0
lentigo NOS 44 0
ichenoid keratosis 37 0
Melanoma 0 584
Nevus 5193 0
seborrheic keratosis 135 0
Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 15 September 2022 doi:10.20944/preprints202209.0215.v1

solar lentigo 7 0
unknown 27124 0

Figure 3. Images of Benign and Malignant Skin Cancers.

3.2 Experimental Setup

Figure 4 depicts the whole experimental setup. We employed data preparation first, followed by
data augmentation in the second stage. Three models were chosen for pretrained training. Except
for the output layer, we used all layers. A flatten layer, dense layer, dropout layer, and final out-
put layer were added. After we concluded training, we used accuracy to evaluate the performance
of our trained model.

Data Preprocessing

Data Augmentation

Pretrained Models

Add Flatten, Dense and Output Layer

Evaluate Trained Models on Testing Dataset

Figure 4. Experimental Setup.

Table 2. Summary of the result of the experiment

Pretrained Model Training Accuracy Validation Accuracy

Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 15 September 2022 doi:10.20944/preprints202209.0215.v1

VGG16 0.9912

ResNet 0.9873 0.9839

MobileNetV2 0.9976

4. Discussion
Using a massive dataset of 33,126 images, we achieved a validation accuracy of
98.39%. Using the pretrained model AlexNet, they attained an accuracy of 0.7853 and
0.9086 in [15] and [16]. Similarly, in [15], ResNet, VGGNet, and Xception were used to
obtain accuracy of 0.9208, 0.8870, and 0.9030, respectively. Table 3 displays testing accu-
racy from the literature. The greatest accuracy across all of these pretrained models was
0.9420 in [17]. One pretrained model was utilized in all of these architectures by integrat-
ing stages such as data augmentation, data standardization, and so on. [16] employed a
hybrid of AlexNet, deep convolutional neural network (DCNN), and support vector ma-
chine (SVM). However, none of the designs outperformed our attained accuracy. Further-
more, we tackled the challenges of overfitting and data preparation while minimizing the
need to develop models from scratch.

Table 3. Pretrained models accuracies in literature.

Pretrained Model Family Accuracy

AlexNet 0.7853[15], 0.9086[16]
ResNet 0.9208[15], 0.9093[16]
VGGNet 0.8870[15], 0.8602[16]
Xception 0.9030[15]
DenseNet 0.8833[16]
MobileNet 0.8826[16]
DCNN 0.9143[16]
AlexNet + SVM + DCNN 0.9420[17]

5. Conclusions
Melanoma is one of the most serious skin malignancies, shortening people's lives
dramatically. However, early identification avoids any problematic issue. AI can help de-
tect this sort of cancer at an early stage. In this work, we employed pretrained models to
compare performance while taking an evaluation metric into account (accuracy). We used
RMSProp, SGD, and ADAM to optimize the models. Pretrained models were employed
to get the maximum accuracy while spending the least amount of time creating models
from scratch. Furthermore, we addressed the issue of overfitting and offered alternative
data processing techniques with dataset insights. We achieved a validation accuracy of
98.39%, outperforming the prior pretrained model's performance despite the need for a
complex model. The findings of this study can be applied in medical science to help phy-
sicians diagnose skin cancer early and save lives.

Author Contributions: Conceptualization, S.S.; methodology, S.S. and P.R.; software, S.S.; valida-
tion, S.S. and P.R.; formal analysis, S.S.; investigation, S.S.; resources, S.S. and P.R.; data curation,
S.S.; writing—original draft preparation, S.S. and P.R.; writing—review and editing, P.R. and S.S.;
visualization, S.S.; supervision, S.S.; project administration, P.R.; funding acquisition, S.S. and P.R.
Both authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 15 September 2022 doi:10.20944/preprints202209.0215.v1

Data Availability Statement: The datasets used or analyzed during the current study are available
from the corresponding author upon reasonable request.
Conflicts of Interest: We declare that there is no conflict of interest.

References
1. Kaur, R.; GholamHosseini, H.; Sinha, R.; Lindén, M. Melanoma Classification Using a Novel Deep Convolutional Neural Net-
work with Dermoscopic Images. Sensors 2022, 22, 1134, doi:10.3390/s22031134.
2. Harangi, B. Skin Lesion Classification with Ensembles of Deep Convolutional Neural Networks. Journal of Biomedical Infor-
matics 2018, 86, 25–32, doi:10.1016/j.jbi.2018.08.006.
3. Astudillo, N.M.; Bolman, R.; Sirakov, N.M. Classification with Stochastic Learning Methods and Convolutional Neural Net-
works. SN COMPUT. SCI. 2020, 1, 119, doi:10.1007/s42979-020-00126-x.
4. Spanhol, F.A.; Oliveira, L.S.; Petitjean, C.; Heutte, L. Breast Cancer Histopathological Image Classification Using Convolutional
Neural Networks. In Proceedings of the 2016 International Joint Conference on Neural Networks (IJCNN); July 2016; pp. 2560–
2567.
5. Gao, Q.; Lim, S.; Jia, X. Hyperspectral Image Classification Using Convolutional Neural Networks and Multiple Feature Learn-
ing. Remote Sensing 2018, 10, 299, doi:10.3390/rs10020299.
6. Bora, K.; Chowdhury, M.; Mahanta, L.B.; Kundu, M.K.; Das, A.K. Pap Smear Image Classification Using Convolutional Neural
Network. In Proceedings of the Proceedings of the Tenth Indian Conference on Computer Vision, Graphics and Image Pro-
cessing; Association for Computing Machinery: New York, NY, USA, December 18 2016; pp. 1–8.
7. Brinker, T.J.; Hekler, A.; Utikal, J.S.; Grabe, N.; Schadendorf, D.; Klode, J.; Berking, C.; Steeb, T.; Enk, A.H.; Kalle, C. von Skin
Cancer Classification Using Convolutional Neural Networks: Systematic Review. Journal of Medical Internet Research 2018, 20,
e11936, doi:10.2196/11936.
8. Esteva, A.; Kuprel, B.; Novoa, R.A.; Ko, J.; Swetter, S.M.; Blau, H.M.; Thrun, S. Dermatologist-Level Classification of Skin Cancer
with Deep Neural Networks. Nature 2017, 542, 115–118, doi:10.1038/nature21056.
9. Pham, T.-C.; Luong, C.-M.; Visani, M.; Hoang, V.-D. Deep CNN and Data Augmentation for Skin Lesion Classification. In Pro-
ceedings of the Intelligent Information and Database Systems; Nguyen, N.T., Hoang, D.H., Hong, T.-P., Pham, H., Trawiński,
B., Eds.; Springer International Publishing: Cham, 2018; pp. 573–582.
10. The ISIC 2020 Challenge Dataset Available online: https://challenge2020.isic-archive.com/ (accessed on 23 August 2022).
11. Banerjee, S.; Singh, S.K.; Chakraborty, A.; Das, A.; Bag, R. Melanoma Diagnosis Using Deep Learning and Fuzzy Logic. Diag-
nostics 2020, 10, 577, doi:10.3390/diagnostics10080577.
12. Winkelmann, R.R.; Farberg, A.S.; Tucker, N.; White, R.; Rigel, D.S. Enhancement of International Dermatologists’ Pigmented
Skin Lesion Biopsy Decisions Following Dermoscopy with Subsequent Integration of Multispectral Digital Skin Lesion Analy-
sis. J Clin Aesthet Dermatol 2016, 9, 53–55.
13. Albahar, M.A. Skin Lesion Classification Using Convolutional Neural Network With Novel Regularizer. IEEE Access 2019, 7,
38306–38313, doi:10.1109/ACCESS.2019.2906241.
14. Mahbod, A.; Schaefer, G.; Wang, C.; Ecker, R.; Ellinge, I. Skin Lesion Classification Using Hybrid Deep Neural Networks. In
Proceedings of the ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP);
May 2019; pp. 1229–1233.
15. Architectures on Melanoma Detection. Tissue and Cell 2019, 58, 76–83, doi:10.1016/j.tice.2019.04.009.
16. Ali, M.S.; Miah, M.S.; Haque, J.; Rahman, M.M.; Islam, M.K. An Enhanced Technique of Skin Cancer Classification Using Deep
Convolutional Neural Network with Transfer Learning Models. Machine Learning with Applications 2021, 5, 100036,
doi:10.1016/j.mlwa.2021.100036.
17. Dorj, U.-O.; Lee, K.-K.; Choi, J.-Y.; Lee, M. The Skin Cancer Classification Using Deep Convolutional Neural Network. Multimed
Tools Appl 2018, 77, 9909–9924, doi:10.1007/s11042-018-5714-1.