Sei sulla pagina 1di 13

PERFORMANCE ANALYSIS OF HOOKWORM DETECTION

USING DEEP CONVOLUTIONAL NEURAL NETWORK


ANUBALA V P Dr. N. MUTHUKUMARAN, M.E., Ph.D
P.G Scholar/ ECE, Professor/ECE,
Anna University, Chennai, Anna University, Chennai.
Francis Xavier Engineering College, Francis Xavier Engineering College,
Tirunelveli-627003, TamilNadu, INDIA. Tirunelveli-627003, Tamilnadu, INDIA.
Email: anubala.vp@gmail.com Email: kumaranece@gmail.com

Abstract: - Hookworm is an infection that is caused by a parasitic bloodsucking


roundworm. Due to poor sanitation in the developing countries of the tropics and
subtropics, hookworm infection is a leading cause of maternal and child morbidity. The
previous work focuses on bleeding region detection and frame localization is solved with
Support Vector Machine (SVM) and KNN classifier. However the hookworms
demonstrate different shapes, widths and bend orientations. These challenges pose a
great difficulty for automatic hookworm detection. The proposed system consists of two
Convolutional Neural Networks, specifically edge taking out network along with
hookworm taxonomy network, which avoid the edge attribute caching and speed up the
categorization. To combine the tubular regions obtained from the edge taking out
network and the feature maps from hookworm taxonomy network, two edge pooling
layers are introduced. In this proposed method, the accuracy and the area under curve
analysis in detecting the hookworm shows its prospective for clinical application. The
tool used for this process is MATLAB Software.
Keywords: Hookworm, morbidity, edge feature caching, Convolutional Neural
Network, Support Vector Machine, KNN Classifier.

1. INTRODUCTION:
Roundworm which is an Intestinal worm that causes an infection known as
Hookworm Infection. The larvae and the adult worms that live in the small intestine of
human can cause intestinal disease [1]. The two main species of hookworm infecting
humans are Necator americanus and Ancylostoma duodenale. It affects lungs, skin,
and small intestine of human. According to Centres for Disease Control and Prevention,
approximately 576 to 740 million people worldwide get affected.
Itching and rashes are the first signs of infection. If this Infection lasts for a long
time, it will make the person Anemic and a condition called Ascites [2]. In this paper,
Hookworm Detection Framework (HDF) is designed and analyzed for images that model
the visual appearances and tubular patterns of hookworms by using Structured Random
Forests for Edge Extraction and Convolutional Neural Network for Hookworm
Classification Network.
This paper is organized as follows. Section II gives a brief overview of related
works. Sections III elaborates about the proposed work which has two networks namely,
edge extraction network and hookworm classification network, respectively. Section IV
discusses the results. Finally, this paper is concluded with a summary.

2. PROBLEM IDENTIFICATION:
The Wireless Capsule Endoscopy is used as an imaging capsule to view the
whole gastrointestinal tract. This technique is helpful to examine entire small intestine
of human without any pain [3]. The major drawback of this technique is that too many
images need to be manually examined by health care providers.
The bleeding regions from WCE video can be detected by grouping the pixels
through superpixel segmentation to reduce the computational complexity while
maintaining high diagnostic accuracy [4]. Features of each superpixel are extracted
using the red ratio in RGB space and fed into support vector machine for
classification. Most GI tract tumors start from the glandular cells that result into
mucus and make up the inner lining of the digestive system. For WCE images, the
physicians mainly use color and texture pattern to judge and discriminate status of the
mucosa of inner tract of GI [5]. Local Binary Patterns (LBP) is a local texture
descriptor to describe the intensity distribution. The image pixels in the
neighbourhood are first compared to the centre pixel value, and then the differences
between the central pixel and its neighbours are labelled using the indicator function
defined as follows:
1, 𝑥≥0
𝑠 (𝑥 ) = { ………. (2.1)
0, 𝑥<0
The values of the pixels in the neighbourhood are then multiplied by the
binomial weights 2p given to the corresponding pixels [6]. This method is referred to
as LBP. Finally, values of the products are summed up to obtain an LBP number of
this neighbourhood. Wavelet transform may provide local characterization to better
analyze the mucosa of the inner GI tract, leading to more capable image analysis using
information at different scales. Moreover, wavelet transform may obtain information
along different orientations in an image. In addition, our previous preliminary work
has demonstrated that wavelet-based texture provides more information for tumor
detection in WCE images [7] [8].
Support Vector Machine is a state-of-the-art technique used for classification,
and it has been shown to demonstrate better accuracy and computational advantages
over some other traditional classification approaches. An SVM constructs a binary
classifier from a set of labelled patterns called training set [9] [10].
3. PROBLEM SOLUTION:
3.1 INTRODUCTION ABOUT THE PROPOSED SYSTEM
The technique used in this proposed system is Convolutional Neural Network.
It consists of two networks namely, Edge Extraction Network and Hookworm
Classification Network to simultaneously model visual appearances and tubular
patterns of hookworms. Structured Random Forests is one of the edge extraction
network which is used in this proposed method to extract the edges of the hookworm.
The two level edge pooling layers are planned to combine the tubular regions
obtained from edge withdrawal system and the feature maps calculated from the
hookworm categorization network, to enhance the feature maps equivalent to tubular
regions.

3.2 BLOCK DIAGRAM

Figure 3.1 Block Diagram of Hookworm Detection Framework

3.2.1 Description of HDF


The Wireless Capsule Endoscopy (WCE) captures the internal organs of Gastro
Intestinal (GI) tract. One of these images is fed as the input. The input image which is
in the RGB color space is first converted into HSV color space in order to get detailed
information of the picture. This HSV image is filtered using Weiner Filter to eliminate
the distortions and to enhance the image. Structured Random Forest (SRF) method is
used to get the edges of hookworm present in the image. Features are calculated by K-
means clustering. Based on the obtained features the CNN Classification is done by
performing layers of Convolution process and pooling process.
Finally, the output image is generated showing whether the hookworm is
present or not. The block diagram of Hookworm Detection Framework is shown in
Figure 3.1.

3.3 PREPROCESSING
The aim of pre-processing is to improve the image data by suppressing unwanted
distortions and to enhance some image features important for further processing. In
this project, first the input RGB image is converted into HSV color space. Further the
image is filtered using wiener filter to filter the distortions thereby enhancing the
image. The Steps that are carried out in this process is shown in Figure 3.2.

Figure 3.2 Block Diagram of Pre-processing

3.4 EDGE EXTRACTION NETWORK


Discontinuity in the gray level values is defined by an Edge. In other words, the
boundary between an object and the background can also be defined as an Edge.
There are many parameters that the shape of edges in images depends on namely, the
geometrical and optical properties of the object, the illumination conditions, and the
noise level in the images. It includes the most important information in the image, and
can provide the information of the object’s position.

3.4.1 Structured Random Forests


In Structured Random Forest (SRF) of edge extraction method, the training
random forests with structured labels pose two main challenges. One is structured
output spaces are often high dimensional and complex. Therefore scoring numerous
candidate splits directly over structured labels may be prohibitively expensive.
Another one and more critically, information gain over structured labels may not be
well defined.

3.5 SEGMENTATION
Segmentation process is used to analyze the image by segmenting the image
into multiple matrices. The edge detected image from the edge extraction network by
using the Structured Random Forest method for defining the edges of the hookworm
is fed into the segmentation block. This helps to define the tubular patterns of
hookworm more accurately from the background unwanted distortions.
3.5.1 Feature Extraction
The various features are considered to extract the tubular patterns of the
hookworm. K means clustering is used to extract the features for further work. Feature
selection means choosing an optimal subset of features based on some selection
algorithms. Both efficiency and performance of classification may be improved
through reducing the number of features or choosing the optimal subset features.

3.6 HOOKWORM CLASSIFICATION NETWORK


The performance of hookworm detection could be effectively boosted, if the
tubular regions can be appropriately integrated into the deep learning framework.
With the tubular regions induced from edge extraction network, a deep hookworm
classification network is constructed in this section, in which two edge pooling layers
are proposed to embed tubular regions into the convolutional neural network.
It consists of a deep convolutional neural network (DCNN) consists of many
neural network layers. There are two different types of layers, convolutional and
pooling, are typically alternated.

3.6.1 Convolutional Layer


The convolutional layers serve as feature extractors, learn the feature
representations of the input images. The neurons present in the convolutional layers
are arranged into feature maps. Each neuron in a feature map has a receptive field,
which is connected to a neighbourhood of neurons in the previous layer via a set of
trainable weights. In this layer, the inputs are convolved with the learned weights in
order to compute a new feature map, and the convolved results are sent through a
nonlinear activation function. Finally, all neurons within a feature map have weights
that are constrained to be equal.

3.6.2 Pooling Layer


The pooling layers are used to reduce the spatial resolution of the feature maps
and thus achieve spatial invariance to input distortions and translations. Initially,
average pooling is done by taking average for all input feature maps. Recently,
maximum pooling is done by taking maximum value within a receptive field to the
next layer. Then, several convolutional and pooling layers are usually stacked on top
of each other to extract more abstract feature representations in moving through the
network.

3.6.3 Two-Level Edge Pooling


A two-level edge pooling is proposed to integrate tubular region information
into the shallow and deep parts of the classification network, respectively.
The edge pooling layer takes small rectangular blocks from the convolutional
layer and sub-samples them to produce a single output. A weighting scheme is
adopted to emphasize the feature maps corresponding to the tubular regions. The max
pooling takes the maximum values of blocks to obtain the strongest response. Feature
maps corresponding to tubular regions will be enhanced, which are potential areas of
hookworm bodies. It enhances the robustness of classification and boosts the
performance.

4. RESULT AND DISCUSSION


4.1 IMAGE WITH HOOKWORM
The Wireless Capsule Endoscopy Image shows the interior part of the small
intestine of human. This image is chosen as the input image, which is in the Red,
Green and Blue (RGB) color space. The Figure 4.1 shows the input image.

Figure 4.1 Input Image


An improved alternative to Red Blue Green color space (RGB) is a perceptual
color space such as Hue, Saturation and Value (HSV) that is closer to the human color
perception system. The HSV image is shown in Figure 4.2.

Figure 4.2 HSV Image


An image histogram is a type of histogram that acts as a graphical
representation of tonal distribution in a digital image. Figure 4.3 shows the histogram
equalization image. The pre-processed image undergoes edge extraction and
convolutional neural network. The output image is shown in Figure 4.4. After the
process completed, a message will be displayed on the screen showing the presence of
hookworm by “Hookworm Detected”. This is shown in the Figure 4.5.

Figure 4.3 Histogram Equalization Image

Figure 4.4 Output Image with Hookworm


Figure 4.5 Hookworm Detected
For the image with hookworm analysis, the area that get affected is analyzed and
the value is calculated in percentage as follows,
Affected Area = 34.9553 %

4.2 IMAGE WITHOUT HOOKWORM


For this case, to find when the input image given is a normal image and there is
no hookworm present in the image taken with the help of Wireless Capsule
Endoscopy, the following steps are done. The Figure 4.6 shows the input image that
need to be chosen.

Figure 4.6 Input Image


In order to analyze the image deeply, and to apply edge extraction and
classification network RGB color model need to be convert to HSV color model and
this is shown in Figure 4.7.
Figure 4.7 HSV Converted Image
The HSV image is then converted into gray scale and for this gray scale image
histogram equalization is estimated and this is shown Figure 4.8.

Figure 4.8 Histogram Equalization Image

The edges are extracted by using Structured Random forest technique. After
that, the various features need to be calculated and K-means clustering is used for that.
Then, Classification is done based on the CNN technique and the image is shown in
Figure 4.9.
Figure 4.9 Output Image without Hookworm
The Figure 4.10 shows a message box that displaying that the input image
given is the “Normal Image”. For the image with hookworm analysis, the area that get
affected is analyzed and the value is calculated in percentage as follows,
Affected Area = 24.906 %

Figure 4.10 Normal Image

The Performance Analysis of the existing method say Support Vector Machine
(SVM) and the proposed method say Convolutional Neural Network (CNN) is done
by considering the Accuracy factor and the Area under curve factor and is shown in
Figure 4.11 and Figure 4.12.
Figure 4.11 Accuracy
The curve in the Figure 4.12 is created by plotting the true positive rate against
the false positive rate at various threshold settings. It is an effective evaluate metric
for the binary classification, in which a larger area means a better performance.

Figure 4.12 Area under curve

5. CONCLUSION AND FUTURE WORK


A Hookworm recognition structure is proposed for WCE images to model the
visual emergence and tubular regions of hookworms, which contains an edge finding
network and a hookworm categorization network. To incorporate the two networks,
two-level edge pooling is implanted into the shallow and deep parts of hookworm
categorization network, respectively. Experiments on a large dataset demonstrate that
the proposed framework outperforms state-of-the-art methods with hand-drafted
features as well as the off-the-shelf deep learning approaches. Since the proposed
method detects the hookworms based on the whole images in this work, the influence
of background cannot be completely avoided.
To further improve the robustness and effectiveness of the method, we will
explore the region based detection framework in our future work, which integrates the
target location and pixel-level region as the supervised information to construct an
end-to-end model. In addition, we plan to integrate temporal and spatial relationship
between consecutive images to further boost the performance, and exploit advanced
deep learning methods to detect other lesions for WCE images.

REFERENCE
1 Jun-Yan He, Xiao Wu, “Hookworm Detection in Wireless Capsule Endoscopy Images with Deep
Learning”, DOI 10.1109/TIP.2018.2801119, IEEE Transactions on Image Processing, 2018.
2 Brosch. T, L. Y. W. Tang, Y. Yoo, D. K. B. Li, A. Traboulsee, and R. Tam, “Deep 3d
convolutional encoder networks with shortcuts for multiscale feature integration applied to
multiple sclerosis lesion segmentation,” IEEE Trans. Med. Imaging, vol. 35, no. 5, pp. 1229–
1239, 2016.
3 N. Muthukumaran, 'Analyzing Throughput of MANET with Reduced Packet Loss', Wireless
Personal Communications, Vol. 97, No. 1, pp. 565-578, November 2017.
4 Fu. Y, W. Zhang, M. Mandal, and M.-H. Meng, “Computer-aided bleeding detection in wce
video,” IEEE Journal of Biomedical and Health Informatics, vol. 18, no. 2, pp. 636–642, 2014.
5 N. Muthukumaran and R. Ravi, 'The Performance Analysis of Fast Efficient Lossless Satellite
Image Compression and Decompression for Wavelet Based Algorithm', Wireless Personal
Communications, Volume. 81, No. 2, pp. 839-859, March 2015.
6 N. Muthukumaran and R. Ravi, 'Simulation Based VLSI Implementation of Fast Efficient
Lossless Image Compression System using Simplified Adjusted Binary Code & Golumb Rice
Code', World Academy of Science, Engineering and Technology, Volume. 8, No. 9, pp.1603-
1606, 2014.
7 Iddan. G, G. Meron, A. Glukhovsky, and P. Swain, “Wireless capsule endoscopy,” Nature, vol.
405, p. 417, 2000.
8 N. Muthukumaran and R. Ravi, 'Quad Tree Decomposition based Analysis of Compressed Image
Data Communication for Lossy and Lossless using WSN', World Academy of Science,
Engineering and Technology, Volume. 8, No. 9, pp. 1543-1549, 2014.
9 Krizhevsky. A, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional
neural networks,” in Advances in Neural Information Processing Systems, 2012, pp. 1097–1105.
10 M. Ruban Kingston, N. Muthukumaran, R. Ravi, 'A Novel Scheme of CMOS VCO Design with
reduce number of Transistors using 180nm CAD Tool', International Journal of Applied
Engineering Research, Volume. 10, No. 14, pp. 11934-11938, 2015.
Anubala V P was born in Tirunelveli, Tamilnadu, India in 1994. She received
the B.E (Electronics and Communication Engineering) from Anna University,
Chennai, India, in 2016 and currently she is pursuing her P.G degree in Francis
Xavier Engineering College, Tirunelveli, Tamilnadu, India. Her research
interests include, multimedia, computer vision, embedded and Image
processing.

Dr. N. Muthukumaran was born in Kaniyakumari, Tamilnadu, India, in


1984. He received the B.E (Electronics and Communication Engineering) from
Anna University, Chennai, India, in 2007 and the M.E (Applied Electronics)
from Anna University, Chennai, India, in 2010 and the Ph.D (Information and
Communication Engineering) Degree from Anna University, Chennai, India.
He is currently working as an Associate Professor in Electronics and
Communication Engineering Department & Research centre lab in Francis
Xavier Engineering College, Tirunelveli, Tamilnadu, India. His major research
interests are Image Processing/ Compression, Digital and Analog Very Large-
Scale Integration circuit design and Networks. He conducted several projects in the area of Image
processing, Image Compression, Very Large-Scale Integration and Networks. Since 2008 he has
published more than 26 journals in International and 57 National/International conferences papers.

Potrebbero piacerti anche