Informatics in Medicine Unlocked 18 (2020) 100282

Contents lists available at ScienceDirect

Informatics in Medicine Unlocked

journal homepage: http://www.elsevier.com/locate/imu

Skin cancer detection: Applying a deep learning based model driven

architecture in the cloud for classifying dermal cell images
Mohammad Ali Kadampur *, Sulaiman Al Riyaee
Department of Information Management, College of Computer & Information Sciences, Al-Imam Muhammed Ibn Saud Islamic University, Riyadh, India

A R T I C L E I N F O A B S T R A C T

Keywords: Background: Skin cancer is a common form of cancer, and early detection increases the survival rate.
Deep learning Objective: To build deep learning models to classify dermal cell images and detect skin cancer.
AI Methods: A model-driven architecture in the cloud, that uses deep learning algorithms in its core implementa­
Model driven architecture
tions, is used to construct models that assist in predicting skin cancer with improved accuracy. The study il­
Deep cognition studio
CNN
lustrates the method of building models and applying them to classify dermal cell images.
Cancer Results: The deep learning models built here are tested on standard datasets, and the metric area under the curve
Image classification of 99.77% was observed.
Conclusions: A practitioner can use the model-driven architecture and quickly build the deep learning models to
predict skin cancer.

1. Introduction research solutions by developing computer image analysis algorithms

[[26]]. The majority of these algorithmic solutions were parametric,
Skin cancer is a dangerous and widespread disease [1] Each year meaning that they required data to be normally distributed. As the na­
there are approximately 5.4 million new cases of skin cancer are ture of data cannot be controlled, these methods would be insufficient to
recorded in USA alone [2,4].The global statistics are equally alarming accurately diagnose the disease. However, non-parametric solutions do
[3][5]. Recent reports show that from 2008 to 2018, there has been a not rely on the constraint that the data is in normal distribution form.
53% increase in new melanoma cases diagnosed annually [1,4]. The In this paper, an augmented assistance to the dermatologist is pro­
mortality rate of this disease is expected to rise in the next decade. The vided using deep learning. The essence of the approach is that a com­
survival rate is less than 14% [8–10] if diagnosed in later stages. How­ puter is trained to determine the problem by analyzing the skin cancer
ever, if the skin cancer is detected at early stages then the survival rate is images. The novelty of the presentation is that the computer model can
nearly 97% [3]. This demands the early detection of skin cancer. This be developed without having any programming knowledge. The average
paper addresses the issue of early diagnosis, with improved accuracy. accuracy of diagnosis using this model is found to be approximately
It is found that [18,22,23] a skilled dermatologist usually follows a 98.89% and the best is 100%. The machine assisted diagnosis presented
series of steps, starting with naked eye observation of suspected lesions, here overcomes the problem of delay, accuracy, and scarcity of derma­
then dermoscopy (magnifying lesions microscopically) and followed by tologists in public health.
biopsy. This would consume time and the patient may advance to later Studies suggest that in the area of skin cancer detection and image
stages. Moreover accurate diagnosis is subjective, depending on the skill classification, there exists a plethora of research papers. A detailed
of the clinician. It is found that the best dermatologist has an accuracy of survey of these methods is available in Refs. [1,7,8]. Each of these pa­
less than 80% in correctly diagnosing the skin cancer [6]. Adding to pers [1,7,8] used the then available state-of-art methods and claim
these difficulties, there are not many skilled dermatologists available performance improvements. The popular methods used for image clas­
globally in public healthcare. sifications vary from application of decision tree algorithms [9,10]
In order to diagnose skin cancer speedily at the earliest stage and Bayesian classifiers [11–13], support vector machines [8,14], to a va­
solve some of the aforementioned problems, there has been extensive riety of Artificial Intelligence based approaches [15,16]. However a

* Corresponding author.
E-mail addresses: mkadampur@imamu.edu.sa (M.A. Kadampur), saalriyaee@imamu.edu.sa (S. Al Riyaee).
URL: http://www.imamu.edu.sa (M.A. Kadampur).

https://doi.org/10.1016/j.imu.2019.100282
Received 5 August 2019; Received in revised form 30 November 2019; Accepted 5 December 2019
Available online 15 December 2019
2352-9148/© 2019 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
M.A. Kadampur and S. Al Riyaee Informatics in Medicine Unlocked 18 (2020) 100282

common theme in all these papers is that they are made to appear as a used 129,450 clinical skin cancer images including 3,374 dermatoscopic
job of specialists in the domain of computers and software engineering. images. The reported accuracy of classification is 72.1 � 0.9. In 2016, Yu
An essential level of programming expertise in computer languages like et al. [21] developed a CNN with over 50 layers on the ISBI 2016
Java, R, and Python is indeed required to build any of these diagnostic challenge dataset for the classification of malignant melanoma cancer.
models [17–19]. In this paper, we address methods of developing deep The best classification accuracy reported in this challenge was 85.5%. In
learning based image classification models for identification of skin 2018, Haenssle et al. [22] utilized a deep convolutional neural network
cancer, without having prior programming knowledge. The main ob­ to classify a binary diagnostic category of dermatoscopy melanocytic
jectives of this paper are: images, and reported 86.6% sensitivity and specificity for classification.
A multiclass classification using ECOC SVM and deep learning CNN was
1 To enable researchers and practitioners to develop deep learning developed by Dorj et al. in Ref. [23]. The approach was to use ECOC
models by simple plug and play art. SVM with pre-trained AlexNet Deep Learning CNN and classify multi­
2 To classify the cell images and identify Cancer with an improved class data. An average accuracy of 95.1% is reported in this work. In
degree of accuracy using deep learning. Ref. [24] Han et al. have used a deep convolutional neural network to
classify the clinical images of 12 skin diseases. The reported best clas­
2. Related works sification accuracy instance varies between 96.0% � 1%. A detailed
review of classifiers is not the scope of this paper; however a systematic
The research of skin cancer detection based on image analysis has review of deep learning classifiers is found in Ref. [25].
advanced significantly over the years. Many different techniques have
been tried [1]. The International Skin Imaging Collaboration (ISIC) 3. Models & algorithms
event of 2018 has become a de facto benchmark in skin cancer detection
by hosting a challenge contest. It is also reported that a mobile app can Using Deep Learning Studio (DLS), we can build intelligent data
be used to detect skin cancer. In all these efforts researchers have tried to discovery models. We can simply drag and drop suitable components to
improve the accuracy of diagnosis by employing different classification build an appropriate model. The DLS comes with a desktop version and a
algorithms and techniques. Image classification took to new bounds cloud version. It supports multi GPU training up to 4 GPUs in its com­
when convolutional neural network (CNN) structure was introduced by munity edition and additional GPUs in its enterprise edition. Here in this
Fukushima (1988) and later Le-Cunn (1990). They used CNNs for image work we have used the cloud version with a server GPU-XEON-E5-8 GB.
classification. CNNsbasically mimic the human visual cognition system DLS has UI components for project creation, data upload, model build­
and are considered to be the best state-of-the-art methods for image ing, training the model, testing the model, and code generation in its
classification. Though there is a plethora of literature available on image architecture. Different deep learning algorithms can be selected just by
classification, we limit our review of the literature to deep learning dragging the relevant dashboard component. DLS enables easy building
methods for skin cancer images. of deep learning models. An example model build is shown in Fig. 1.
The first breakthrough on skin cancer classification by a pre-trained The steps for constructing a deep learning model are listed below:
GoogLeNet Inception V3 CNN model came from Esteva et al. [20]. They

Fig. 1. An example deep learning model as built using DLS.

2
M.A. Kadampur and S. Al Riyaee Informatics in Medicine Unlocked 18 (2020) 100282

1. Select the input block, drag and drop it into the work space. 3.2. Our tested models
2. Perform some normalization (if required, drag & drop related
blocks) MODEL1:INPUT(), BATCHNORM(3), CONV2D11(32,3) RELU,
3. Apply Convolution blocks (Choose activation ReLU preferably) CONV2D2(32,3) RELU, CONV2D3(64,3) RELU, CONV2D4(64,3) RELU,
4. Apply Pooling Blocks MAXPOOL1(2), CONV2D5(128,3) RELU, CONV2D6(128,3) RELU,
5. Apply Dropout block CONV2D7(256,3) RELU, CONV2D8(256,3) RELU, MAXPOOL2(2),
6. Repeat 3, 4, & 5 as many times as optimally needed FLATTEN(), DROP(0.3), DENSE(2) RELU, OUTPUT().
7. Use some flattening block MODEL2:INPUT(), BATCHNORM(0),CONV2D1(32,3) RELU,
8. Use some core layers (Dense blocks preferably). Ensure here that CONV2D2(32,3) RELU, MAXPOOL(8), CONV2D3(64,3) RELU,
final core layer before the output block has the proper output CONV2D4(64,3) RELU, 4*(RESNET50(TRUE,3,224,224)) DROP(0.3),
value parameter as per the data (output) categories. FLATTEN(), DENSE(128) RELU, DENSE(8) SOFTMAX, OUTPUT().
9. Apply the output block. MODEL3:INPUT(), BATCHNORM(0),CONV2D1(32,3) RELU,
10. Set the model parameters (Hyper parameter setting). CONV2D2(32,3) RELU, MAXPOOL(8), CONV2D3(64,3) RELU,
11. Finally, ensure that the Model construction checks Green “OK” CONV2D4(64,3) RELU, 4*(SQUEEZNET(TRUE,3,224,224)) DROP(0.3),
else fine-tune your selections. FLATTEN(), DENSE(128) RELU, DENSE(8) SOFTMAX, OUTPUT().
MODEL4:INPUT(), BATCHNORM(0), CONV2D1(32,3) RELU,
Fig. 2 summarizes the above steps of building deep learning models. CONV2D2(32,3) RELU, MAXPOOL(8), CONV2D3(64,3)RELU,
Setting up of model parameters is critical to the performance of the CONV2D4(64,3) RELU,4*(DENSENET201(TRUE,3,224,224)) DROP
model. There is no rule of thumb for setting these parameters. The (0.3),FLATTEN(),DENSE(128) RELU, DENSE(8) SOFTMAX, OUTPUT().
general guideline for parameter setting is that which has worked well for MODEL5: INPUT(), BATCHNORM(0), CONV2D 1(32,3)RELU,
others, just try that. ReLU is the preferred activation function for inner CONV2D 2(32,3) RELU, MAXPOOL(8), CONV2D3(64,3) RELU,
ConvNets. Filter size is usually 3 � 3 and number of filters used may vary CONV2D4(64,3) RELU, 4*(INCEPTIONV3(TRUE,3,224,224))DROP
in terms of multiples of two. A dropout percentage of 0.25–0.50 is (0.3),FLATTEN(),DENSE(128) RELU, DENSE(8) SOFTMAX, OUTPUT().
applied to speed up the training. The user must check that the number of MODEL1 above was built using 2D convolutional networks as an
outputs in the output layer is proper. If the training accuracy is less, then initial experimentation and tested on two datasets. The ROC analysis are
the epochs and convolution layers should be increased. A pooling layer presented for this model on the two datasets as shown in Figs. 4 and 5.
down-samples the volume along the spatial dimensions by a certain The remaining models use pre-trained networks, and results of experi­
factor. This reduces the size of the representation and the number of ments with these pre-trained models are presented in Table 1.
parameters which makes the model more efficient and prevents over­
fitting. Use the setting to prevent model overfitting. 4. Experiments & results

3.1. Methodology The pre-trained models above (MODEL2-MODEL5) were trained and
tested using the HAM10000 dataset, which is a benchmark public
The DLS, Model Driven Architecture Tool provides Neural Network database for machine learning. The data was divided into 80:10:10::
Modeling components as a stack of drag-drop & develop art. The Training:Validation:Testing ratios. The results of each of the above
sequence of important general steps involved as a research methodology models is presented in Table 1. The Fig. 3 shows a training instance.
in this paper are as follows:
4.1. Training
1 Data Preparation
2 Creating a project and loading the dataset In order to start training the model, one needs to start the compute
3 Building a deep learning classifier instance and push the red power button. The power button will go green
4 Tuning the model and the training instance using the remote GPU will start. The dashboard
5 Checking result shows the progress while training, accuracy changes, and completion of
6 Drawing Inferences epochs. Training will stop if there is any error in the run. If everything is
7 Code Access correct, training will stop after the stopping criteria.
8 Deploying the model as a REST API.
5. Discussion
Data preparation is a preprocessing phase in which the raw data is
processed by cleaning and formatting. The DLS accepts the data in the The models built using DLS were tested against the benchmark ma­
form of comma separated variables (.csv) files. We have used the app chine learning databases (HAM 10000) and the results are very prom­
[26] (Irfan view) for bulk naming and customizing image features. ising. A higher value of ROC suggests that the classifier model rarely fails
Creating the project and the remaining steps follow with a quick fa­ to diagnose cancer as cancer, and non-cancer as non-cancer. Among the
miliarity of the dashboard of DLS. Deploying the model as a REST API is pre-trained models, squeeznet, densenet, and inception v3 models have
left as future work. We experimented with the following set of models in a higher ROC as compared to the resnet model. These observations
our research. suggest that the models built using DLS produce better results than many
of the findings in the related work. The model building is easy with DLS
and effective.

Fig. 2. Constructing a deep learning model: steps.

3
M.A. Kadampur and S. Al Riyaee Informatics in Medicine Unlocked 18 (2020) 100282

Fig. 3. An instance of training the model.

Fig. 4. ROC analysis of MODEL1 on dimizC dataset.

4
M.A. Kadampur and S. Al Riyaee Informatics in Medicine Unlocked 18 (2020) 100282

Fig. 5. ROC analysis of MODEL1 on Kaggle dataset.

manuscript “Deep Learning: Detecting Skin Cancer by Classifying

Table 1
Dermal Cell Images Using a Model Driven Architecture in the Cloud” the
Metric values of pre-trained deep learning classifiers.
following is fulfilled:
Model (Precision) (F1-Score) (ROC AUC) 1) This material is the authors’ own original work, which has not
MODEL2(resnet) 94.24 94.22 98.61 been previously published elsewhere.
MODEL3(squeeznet) 97.40 94.57 99.77 2) The paper is not currently being considered for publication
MODEL4(densenet) 97.51 96.27 99.09 elsewhere.
MODEL5(inceptionv3) 98.19 95.74 99.23
3) The paper reflects the authors’ own research and analysis in a
truthful and complete manner.
6. Conclusion & future work 4) The paper properly credits the meaningful contributions of co-
authors and co-researchers.
This paper introduced a tool in which a non-programming back­ 5) The results are appropriately placed in the context of prior and
ground person can develop complex deep learning models. It opened the existing research.
options of flexibility in designing deep learning classifiers by hinting at 6) All sources used are properly disclosed (correct citation). Literally
the general procedures and the looping patterns in the development of copying of text must be indicated as such by using quotation marks and
deep learning models (Fig. 2). giving proper reference.
The paper used Deep Learning Studio a Model Driven Architecture 7) All authors have been personally and actively involved in sub­
for Deep learning. The paper introduced the features of the DLS tool and stantial work leading to the paper, and will take public responsibility for
demonstrated the process of constructing a Deep Learning Model using its content.
the tool. The paper narrated the process of data preparation using The violation of the Ethical Statement rules may result in severe
dermal cell images and its test application into the DLS model to detect consequences.
cancer cells. The DLS models achieved an AUC of 99.77% in detecting To verify originality, your article may be checked by the originality
cancer cells from the images of cancer cells. detection software iThenticate. See also http://www.elsevier.com/edit
The paper pointed at the provision of obtaining the programming ors/plagdetect.
code for the model for further exploration by the programming I agree with the above statements and declare that this submission
specialist. The provision to download the trained model and develop follows the policies of Solid State Ionics as outlined in the Guide for
enterprise level applications is the best seed level research that this Authors and in the Ethical Statement.
paper observes for the future work. Finally, the paper achieved the set
objectives of the introductory section. Funding received

Ethical Statement for informatics in medicine unlocked None.

Hereby, I Dr. Mohammad Ali Kadampur ensure that for the

5
M.A. Kadampur and S. Al Riyaee Informatics in Medicine Unlocked 18 (2020) 100282

Declaration of competing interest [8] Erkol B, Moss RH, Stanley RJ, Stoecker WV, Hvatum E. Au- tomatic lesion
boundary detection in dermoscopy images using gradient vector flow snakes. Skin
Res Technol 2005:17–26.
All authors have participated in (a) conception and design, or anal­ [9] Zorman MMM, Kokol SP, Malcic I. The limitations of decision trees and automatic
ysis and interpretation of the data; (b) drafting the article or revising it learning in real world medical decision making. J Med Syst 1997:403–15.
critically for important intellectual content; and (c) approval of the final [10] Gehrke J. Classification and regression trees. IGI global; 2009.
[11] Larsen K. Generalized n€ aıve bayes classifier. In: ACM SIGKDD explorations news
version. letter, 7 of 1. USA: ACM; 2005. p. 76–81. T.
This manuscript has not been submitted to, nor is under review at, [12] Neal RM. Springer hidelburg Germany. 1996.
another journal or other publishing venue. [13] Ruiz D, Berenguer V, Soriano A, Sanchez B. A decision support system for the
diagnosis of melanoma: a comparative approach. Expert Syst Appl Nov-Dec 2011;
The authors have no affiliation with any organization with a direct or 38(12):15217–23.
indirect financial interest in the subject matter discussed in the [14] Gilmore S, Hofmann-Wellenhof R, Soyer HP. A support vectormachine for decision
manuscript. support inmelanoma recognition. Experimental Dermatogra- phy; 2010. p. 830–5.
[15] Silver D, Huang A, Maddison CJ, Guez A, Sifre L, van den G. Mastering the game of
The following authors have affiliations with organizations with go with deep neural networks and tree search. Nature 2016:484–9. https://doi.
direct or indirect financial interest in the subject matter discussed in the org/10.1038/nature16961.
manuscript: [16] Codella1 N, Cai1 J, Abedini M, Garnavi R, Halpern A, Smith1 JR. Deep learning,
sparse coding, and SVM for melanoma recog- nition in dermoscopy images. Springe
LNCS; 2015. https://doi.org/10.1007/978-3-319-24888-2015.
Acknowledgements [17] Souza AMFF, Bostian FMS, Sugomori KY. Deep learning practical neural networks
with Java. third ed. Pack Publishing; 2017.
[18] Ramasubramaniam K, Singh A. Machine learning using R. second ed. Acta Press;
We thank DeepCognition for creating the Deep Learning program
2018.
used in this study. [19] Chollet F. Deep learning with Python. second ed. Manning Publications Company;
2017.
Appendix A. Supplementary data [20] Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S. Dermatologist-
level classification of skin cancer with deep neural networks. Nature 2017:115.
https://doi.org/10.1038/nature21056.
Supplementary data to this article can be found online at https://doi. [21] Yu L, Chen H, Dou Q, Qin J, Heng P-A. Automated melanoma recog- nition in
org/10.1016/j.imu.2019.100282. dermoscopy images via very deep residual networks. IEEE Trans Med Imaging
2017:994–1004. https://doi.org/10.1109/TMI.2016.2642839.
[22] Haenssle HA, Fink C, Schneiderbauera R, Toberer F, Buhl T, Blum A, Kalloo A,
References Hassen ABH, Thomas L, Enk A, Uhlmann L. Man against machine: diagnostic
performance of a deep learning convolutional neural network for dermoscopic
[1] Masood A, Al-Jumaily AA. Review article computer aided diagnostic support melanoma recognition in comparison to 58 derma- tologists. Ann Oncol 2018:
system for skin cancer: a review of techniques and algorithms. Int J Biomed 1836–42. https://doi.org/10.1093/annonc/mdy166.
Imaging 2013:1–22. https://doi.org/10.1155/2013/323268. [23] Dorj U-O, Lee K-K, Choi J-Y, Lee1 M. The skin cancer classification using deep
[2] Siegel R, Naishadham D, Jamal A. Cancer statistics, 2012. Cancer J. Clin. 2012: convolutional neural network. Multimedia Tools and Applica- tions; 2018.
10–29. https://doi.org/10.3322/caac.20138. p. 9909–24. https://doi.org/10.1007/s11042-018-5714-1.
[3] Siegel R, Miller KD, Jamal A. Cancer statistics, 2018. Cancer J. Clin. 2018:7–30. [24] Han SS, Kim MS, Lim W, Park GH, Park I, Chang SE. Classifica- tion of the clinical
https://doi.org/10.3322/caac.21442. images for benign and malignant cutaneous tumors us- ing a deep learning
[4] Esteva1 A, Kuprel1 B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S. algorithm. J Investig Dermatol 2018:1529–38. https://doi.org/10.1016/j.
Dermatologist-level classification of skin cancer with deep neural networks. Nature jid.2018.01.028.
2017:115–26. https://doi.org/10.1038/nature21056. [25] Brinker1 TJ, Hekler1 A, Utika JS, Grabe N, Schadendorf D, Klode J, Berking C,
[5] W. H. Organization. WHO. [link]. https://www.who.int/uv/faq/skincancer/en/in Steeb T, von Kalle1 AHEC. Skin cancer classification using convolutional neural
dex1.html. networks: systematic review. J Med Internet Res 2018:e11936. https://doi.org/
[6] Marks R. Epidemiology of melanoma, clinical and experimental dermatology. Clin 10.2196/11936.
Dermatol 2000:459–63. https://doi.org/10.1046/j.1365-2230.2000.00693.x. [26] IrfanView. [link]. https://www.irfanview.net/.
[7] Kittler H, Pehamberger H, Wolf K, Binder MJTIO. Diagnostic accuracy of
dermoscopy. Clin Dermatol 2002:159–65. https://doi.org/10.1046/j.1365-
2230.2000.00693.x.