You are on page 1of 5

Journal of Advanced Research in Applied Artificial Intelligence and Neural Network

Volume 5, Issue 1&2 - 2018, Pg. No. 1-5


Peer Reviewed Journal
Research Article

Convolution Neural Network for Feature


Extraction in Skin Disease Detection
Seema Kolkur1, Dhananjay Kalbande2, Vidya Kharkar3
Asst. Prof. Department of Computer Engg, Thadomal Shahani Engg, College Bandra, Mumbai, India.
1

Dean (Industry Relations), Department of Computer Engg, Sardar Patel Institute of Tech, Andheri, Mumbai, India.
2

3
Professor, Department of Skin and VD., Seth G.S.M.C & K.E.M Hosp. Parel, Mumbai-400012, India.

Abstract
Skin Diseases are becoming very common now days. Number of people suffering from skin diseases
is increasing rapidly. Human judgment on diagnosis of skin diseases is sometimes subjective and not
reproducible. To achieve more reliable and objective accuracy computer aided diagnosis may be used.
With advancement in medical imaging, image based classification is been increasingly used for disease
detection in medical field. Feature engineering is very important for any classifier to achieve maximum
results. Convolution Neural Networks (CNN) can learn features on its own reducing total time required for
development of such systems and at the same time increasing level of accuracy. We acquired more than
850 original images for two skin diseases from department of Skin and VD, KEM Hospital, Mumbai. We
have used CNN for feature extraction from input images of two skin diseases. These features are fed to
Support Vector Machine (SVM) for classification. The results indicate CNN can be feasibly used of feature
extraction in skin disease detection.

Categories and Subject Descriptors

I.5.2 [Pattern Recognition]: Design Methodology–classifier design and evaluation.

General Terms

Machine learning, Classification, Feature Extraction.

Keywords: Deep Learning, Convolution Neural Network

Introduction With the advancement in the field of computer technology


and medical imaging, computer aided disease detection and
Skin diseases are very common in everyday life of many diagnosis is becoming a reality. A machine can find some
across the world. In India, dermatological practices are of details form images which possible a human cannot.
not that widespread in all parts. There are very few Computers can efficiently read images while humans
dermatologists currently practicing in India [1].Very few are error-prone [2]. Neural Network based classifiers are
hospitals follow these practices using the advanced widely used for disease detection. Feature extraction is
technology. Delayed diagnosis of some of the diseases may an important aspect of designing any classifier system.
results in increasing the chances of spreading the diseases. Accuracy of a classifier depends on feature set used for
So, it is important to have early detection of skin diseases. classification. Unfortunately, it is equally very tedious and

Corresponding Author: Seema Kolkur, Thadomal Shahani Engg, College Bandra, Mumbai, India.
E-mail Id: kolkur.seema@gmail.com
Orcid Id: https://orcid.org/0000-0003-2703-203X
How to cite this article: Kolkur S, Kalbande D, Kharkar V. Convolution Neural Network for Feature Extraction in Skin Disease Detection.
J Adv Res Appl Arti Intel Neural Netw 2018; 4(1&2): 1-5.

Copyright (c) 2018 Journal of Advanced Research in Applied Artificial Intelligence and Neural Network
Kolkur S et al.
J. Adv. Res. Appl. Arti. Intel. Neural Netw. 2018; 5(1&2) 2

time consuming task. CNNs can be a real great help here.


CNN can learn features on their own from given large
database.

Related Work
Skin Disease Detection is a widely researched area in the
field of Machine Learning and Artificial Intelligence. With
current technologies in the field of Medical Imaging it
is easier to acquire good quality multiple images. Many
researchers in past have used images for diagnosis of a
disease besides numbers and text. Skin Disease Detection
is basically a classification task. Binary Classification is used
to detect presence or absence of a disease in given image.
Different types of features are used by researchers forsuch
classification systems,Color Features, Texture Features,
Dermoscopic Features and Fractal Features being commonly
used. Figure 1.Methodology of Proposed Work

Texture Features like contrast, correlation, energy, intensity


Data Collection
and homogeneity are extensively used in detection of We have collected our dataset of images for two skin
diseases like psoriasis and skin cancer [3-8]. [9] used fractal diseases mainly Lichen Planus and Vitiligo from the
features for skin cancer detection. Authors in [10-12] department of Skin and VD from KEM Hospital,Parel
used dermoscopic features Area, Border, Color, Diameter Mumbai. Table 1 shows image dataset statistics.
(ABCD) of a skin lesion for detection of skin cancer. Color
Table 1.Dataset Description
features mean like mean color (µc), standard deviation
(σc) and skewness of color are used for psoriasis and skin Sr. No. Disease / No of Training Testing Total
cancer detection[2,5,13]. In most of these disease detection Samples
systems SVM and neural networks with backpropagation 1 Lichen Planus 394 23 417
are used as classifiers.
2 Vitiligo 327 15 342
In recent years, CNN has been used in image recognition. Total No. of 857
Variants of CNN models have been increasingly achieving Images
better performance in different fields. Authors in [14]
and [15] have used CNN for multistage feature extraction Image Preprocessing
for classifying digits in house numbers and music genre
classification respectively. [16] showed that Random Forest Following preprocessing techniques are applied to all the
and SVM can be used with features extracted from CNN images collected.
to yield better a prediction accuracy compared to the
original CNN. Even if the CNN is not optimal (not fully • Cropping: Every image is manually cropped to
trained or overfits), can still extract good features that gives concentrate on diseased body area. Cropping of images
competitive prediction accuracy. [17-18] have implemented is very important as size of images is very large.
efficient CNN architectures to detect diabetic retinopathy

Methodology
Fig. 1 shows general methodology of the proposed system.
The system can be broadly categorized into data collection,
preprocessing, feature extraction and classification
modules. Figure 2.Image before and after cropping
Kolkur S et al.
3 J. Adv. Res. Appl. Arti. Intel. Neural Netw. 2018; 5(1&2)

• Resizing: All the cropped images are saved as 64*64 Classification


pixels for maintaining uniformity in dimensions.
For image classification, features extracted from CNN are
Feature Extraction by CNN passed to SVM for further classification. SVM classifier gives
probability of an input image being one of the two diseases.
Convolution Neural networks (CNN) are used for extracting
features from images. CNN are biologically-inspired variants Results
of MLPs. A CNN consists of an input and an output layer,
as well as multiple hidden layers. The hidden layers of a The image dataset was divided in training and testing
CNN typically consist of convolutional layers, pooling layers, dataset as shown in table 1. CNN model was trained to
fully connected layers and normalization layers. Output of fit to training data. We extracted activations from dense
every layer acts as an input to next layer. Fig. 2 shows a layer before output layer. These activations are treated
typical CNN architecture [19]. as features and were fed as an input to SVM. SVM acts
as classifier and classifies output probability of every test
Convolution layer preserves the spatial relationship between record being classified as either lichen planus or vitiligo.
pixels by learning image features using small squares of Table 3.Confusion Matrix
input data[20]. A feature map from convolution layer is
obtained by sliding a filter over an input image. CNN learns Predicted
the values of these filters on its own during the training Lichen Planus Vitiligo
process. Pooling also called as downsampling reduces the True Lichen Planus 21 2
dimensionality of each feature map but retains the most
Vitiligo 3 12
important information. The Fully Connected layer is like
a traditional Multi Layer Perceptron where every neuron
Normally performance of a classifier is measured in terms
in the previous layer is connected to every neuron on the
of precision and recall. Precision and recall for class Lichen
next layer. Convolution and pooling layers act as feature
Planus is calculated as below.
extractors while fully connected layer works as classifier.

Figure 3.Typical CNN Architecture Similarly, precision and recall for class vitiligo is calculated.
Table 2 describes architecture of our network. We have used For clinical performance of a classifier sensitivity also called
two convolution layers, two maxpooling and two dropout as True Positive Rate, specificity called True Negative Rate
layers and one dense layer. Input is 64x64 RGB image and and accuracy are also considered as performance measures.
output layer has two units representing two skin diseases. These measures are calculated using equations 1,2,3.
Both convlutional layers 1 and 2 use 32 filters of 5x5 size
with maxpooling of size 2x2. (1)
Table 2.CNN Layer Description
Sr. No. Layer Name Size (2)
0 input 3x64x64
1 conv2d1 32x60x60 (3)
2 maxpool1 32x58x58
Figure 4.Performance Metrics
3 conv2d2 32x54x54
4 maxpool2 32x52x52 precision recall f1-score support

5 dropout1 32x52x52 Lichen Planus 0.88 0.91 0.89 23

6 Dense 256 Vitiligo 0.86 0.80 0.83 15

7 dropout2 256 avg / total 0.87 0.87 0.87 38

8 Output 02
Kolkur S et al.
J. Adv. Res. Appl. Arti. Intel. Neural Netw. 2018; 5(1&2) 4

As can be seen from table 3 and 4 accuracy, Specificity and classification in 91% which shows CNN can successfully
Sensitivity of classifier is 87%, 80% and 91% respectively. be used for the stated purpose. Major advantage of a
Fig.4 shows 32 filters used in first convolutional layer. system is the saving of time and effort involved in feature
engineering. CNNs learn features on their own.

The accuracy can further be improved by adding more


number of images and/or by adding more number of layers
in CNN architecture. By adding two drop-out layers we
make sure the network does not overfit. This system can
be further extracted to classifying multiple skin diseases.

Preprocessing of images is very important task. Currently,


we manually cropped input images. This process can be
automated by using some image processing techniques
like object localization and segmentation.

References
Figure 4.32 features of size 5X5 from first layer
Discussion 1. Verma S. Nonclinical challenges of Indian dermatology–
cities vs. villages, poverty, and lack of awareness.
Deep learning and CNN is used in many computer aided International journal of dermatology 2007 Nov
diagnosis systems which are based on medical image 1;46(s2):42-5.
analysis. Below, we summarize some of the works involving 2. Al Abbadi NK, Dahir NS, Al-Dhalimi MA, Restom H.
CNN for feature extraction in medical applications. Psoriasis Detection Using Skin Color and Texture
Features 1.
In [21] authors proposed a system to classify Human 3. Sheha MA, Mabrouk MS, Sharawy A. Automatic
Epithelial-2 (HEp-2) cell images into 6 staining pattern detection of melanoma skin cancer using texture
classes. They applied transfer learning on a pre-trained analysis. International Journal of Computer Applications
deep CNN (MatConvNet) to extract the generic features. 2012 Mar;42(20):22-6.
Then most relevant features for classification were selected 4. Ramya VJ, Navarajan J, Prathipa R, Kumar LA. Detection
by feature selection. They reported accuracy of 77.1% on of melanoma skin cancer using digital camera images.
ICPR2012 cell dataset with 1457 images. ARPN Journal of Engineering and Applied Sciences 2015
Apr;10(7):3082-5.
[22] used ensemble of multiple fine-tuned CNNs (AlexNet 5. Priyanka s. Biradar, S. N. Patil, Psoriasis detection using
and GoogLeNet) for feature extraction on ImageCLEF 2016 color and texture feature’, international journal of
medical image public dataset with 6776 training images, and advanced computing and electronics technology 2015;
4166 test images and 30 classes with accuracy of 82.48% 2(4): 153-6.
6. Sonali Jadhav, D. K. Kamat, Analysis and Detection
[23] evaluates CNN performance on two different computer-
of Skin Cancer, IOSR Journal of Electronics and
aided diagnosis applications: thoracoabdominal lymph
Communication Engineering 2014 Aug; 9( 4):50-4
node detection and interstitial lung disease classification
7. Ahmed IN, Chaya P. Segmentation and Classification of
with transfer learning.
Skin Cancer Images. International Journal of Advanced
To the best of our knowledge, no work on feature extraction Research in Computer Science and Software Engineering
from skin disease images using CNN is carried out till 2014 May;4(5).
date. Most of these works, use pretrained models for 8. Jaleel JA, Salim S, Aswin RB. Diagnosis and Detection
feature extraction on a publicly available dataset. Major of skin cancer using Artificial intelligence. International
contribution of our work lies in collection of primary journal of Engineering and innovative Technology 2013
data and training a new network from scratch feature for Aug;3(2):311-5.
extraction. Our system can be used as preprocessing tool 9. Janney BJ, Rohini G, Pauline J. Detection of Malignancy
for extracting features from a large set of input images of Skin Lesions Based on Local Texture And Fractal
before any classification algorithm. Analysis. RESEARCH JOURNAL OF PHARMACEUTICAL
BIOLOGICAL AND CHEMICAL SCIENCES. 2015 Mar
Conclusion 1;6(2):408-12.
10. Goel R, Singh S. Skin Cancer Detection Using GLCM
We have used convolution network for feature extraction Matrix Analysis And SVM Classifier. International
from skin disease images. The extracted features are then Journal 2015 Jan;5(1):6-11.
given as input to SVM classifier. The overall accuracy of
Kolkur S et al.
5 J. Adv. Res. Appl. Arti. Intel. Neural Netw. 2018; 5(1&2)

11. Goel R, Singh S. Skin Cancer Detection using GLCM Convolutional Neural Network Using Retinal Fundus
Matrix Analysis and Back Propagation Neural Images. International Conference on Artificial Neural
Network Classifier. International Journal of Computer Networks 2017 Sep 11 (pp. 635-642). Springer, Cham.
Applications 2015 Jan 1;112(9). 19. Peng M, Wang C, Chen T, Liu G. NIRFaceNet: A
12. Achakanalli S, Sadashivappa G. STATISTICAL ANALYSIS Convolutional Neural Network for Near-Infrared Face
OF SKIN CANCER IMAGE—A CASE STUDY. International Identification. Information. 2016 Oct 27;7(4):61.
Journal of Electronics and Communication Engineering 20. Wu J. Introduction to convolutional neural networks.
2014 May;3(3):1-0. National Key Lab for Novel Software Technology.
13. Ballerini L, Fisher RB, Aldridge B, Rees J. Non-melanoma Nanjing University. China. 2017 May 1.
skin lesion classification using colour image data in 21. Phan HT, Kumar A, Kim J, Feng D. Transfer learning of
a hierarchical K-NN classifier. In Biomedical Imaging a convolutional neural network for HEp-2 cell image
(ISBI), 2012 9th IEEE International Symposium on 2012 classification. InBiomedical Imaging (ISBI), 2016 IEEE
May 2 (pp. 358-361). IEEE. 13th International Symposium on 2016 Apr 13 (pp.
14. Sermanet P, Chintala S, LeCun Y. Convolutional neural 1208-1211). IEEE.
networks applied to house numbers digit classification. 22. Kumar A, Kim J, Lyndon D, Fulham M, Feng D. An
InPattern Recognition (ICPR), 2012 21st International ensemble of fine-tuned convolutional neural networks
Conference on 2012 Nov 11 (pp. 3288-3291). IEEE. for medical image classification. IEEE journal of
15. Li TL, Chan AB, Chun A. Automatic musical pattern biomedical and health informatics 2017 Jan;21(1):31-
feature extraction using convolutional neural network. 40.
InProc. Int. Conf. Data Mining and Applications 2010 23. Shin HC, Roth HR, Gao M, Lu L, Xu Z, Nogues I, Yao
Mar. J, Mollura D, Summers RM. Deep convolutional
16. Athiwaratkun B, Kang K. Feature representation neural networks for computer-aided detection: CNN
in convolutional neural networks. arXiv preprint architectures, dataset characteristics and transfer
arXiv:1507.02313. 2015 Jul 8. learning. IEEE transactions on medical imaging. 2016
17. Pratt H, Coenen F, Broadbent DM, Harding SP, Zheng Y. May;35(5):1285-98.
Convolutional neural networks for diabetic retinopathy.
Procedia Computer Science. 2016 Jan 1;90:200-5. Date of Submission: 2018-04-17
18. García G, Gallardo J, Mauricio A, López J, Del Carpio Date of Acceptance: 2018-05-03
C. Detection of Diabetic Retinopathy Based on a

You might also like