Comparison of Machine Learning Techniques for Identification of Disease

Comparison of Machine Learning Techniques for Identification of
Disease
Shaveta Malik
1
, Tapas Kumar
2
, A.K Sahoo
3*
1, 2 School of Computer Science, Lingaya’s University, Faridabad, Haryana, Shavetamalik687@gmail.com
3* Scientist, Biodiversity Lab, Reverie Ecology and Fisheries Division, Barrackpore, Kolkata
Abstract:-The Fish disease causes several losses in the fish farm. A number of fungal and bacterial diseases
especially EUS (Epizootic Ulcerative Syndrome) causing Morbidity and Mortality in fish. Fish Infection caused
by Aphanomyces invadans is commonly known as EUS (Epizootic Ulcerative Syndrome) and it is due to the
primary fungal pathogen. EUS disease is still misidentified by the people. The paper proposed a combination of
feature extractor with PCA (Principle component analysis) and classifier for better accuracy. Proposed
combination of techniques gives better accuracy while identification of EUS and Non EUS infected fishes.
After experimentation, it is found that PCA improves the performance of classifier after reducing the
dimensions. Real images of EUS infected fishes have been used throughout and all the work is done in
MATLAB.
Keywords: - Epizootic Ulcerative syndrome (EUS), Principle component analysis (PCA), Features from
Accelerated Segment Test (FAST), Neural Network (NN)
I. Introduction
Fish is a dependable source of animal protein in evolving countries like INDIA. Due to large scale of mortality
occurs among the fresh water fishes, it causes a immense loss to the nation. Spreading of EUS is a semi-global
problem among the fishes of fresh water, in large natural water bodies may not be possible to control of EUS,
and Control of EUS in large natural water bodies may not be possible. Today’s major problem is to control and
treatment of EUS. The accuracy of the final diagnosis found using experiences of fish farmers or fish
veterinarian. Traditionally, Skills and experiences and the time spend by the individual defines the accuracy of
the final diagnosis. Normally infected fish will die quickly if correct and accurate treatment is not provided. In
order to solve this problem, combination of Feature extractor with PCA (Principle component analysis) applied
to extract the feature and classifier applied to classify the EUS infected and Non-EUS infected fish in order to
find the accuracy rate of EUS and Non-EUS infected fish. The infected fish will normally die very quickly if
correct and effective treatment is not provided in time. Mortality of fish will affect the loss of fish farmers,
Indian Market loss and automatically it will also affect the international market loss. The paper compares the
combination of different feature extractor with different classifier for finding the accuracy .It finds that the
proposed combination gives better accuracy. The accuracy has been found with the combination of Feature
Extractor and PCA (Principle component analysis) and feature Extractor without PCA. The dimensionality
reduction can be possible through PCA of the dataset and removes the dimensions which have the least
important information. . The data utilizes less space if number of dimensions has been reducing, it helps in
classification of larger dataset s in less time. In the classification experimentation, two classifier or classification
algorithms have been taken to find the accuracy i.e. KNN (K-Nearest Neighbour) and Neural Network. PCA has
been applied after extracting the feature from HOG (Histogram of Gradients) and FAST (Features from
Accelerated Segment Test) of each image. It has been observed through results that PCA (Principle component
analysis) improves the accuracy of classification. Many Researchers have done lot of work in many techniques
related to feature extraction and area related to the paper. Jeyanthi Suresh et al.[1] In the paper, proposed a
method or technique which automatically recognized the activity of human from the video with the feature
extractor which was the HOG & Probabilistic Neural network (PNN) classifier. The classifier was used for
classifying the actions of video experiments and results were found on Kth database and gave better
performance, 89.8% accuracy for test and 100% for the training set and measured the performance of each
featured set with different classifier..Valentin Lyubchenko et al. [2] in the paper selected the markers of colors
to distinguished the infected and Normal area, there was a drawback in the methodology of false point which
can be appeared as a disease area due to automatic allocation of color, it has the ability to change the marker
while selecting the color in the segmented image. Hitesh Chakravorty et al. [3] suggested a method in which
disease fish image recognized by using dimension reduction technique that was through PCA method and
segmentation of fish image with K-means clustering technique, segmentation was based on the color features
HSV images and Morphological operations for the area that is diseased its detection and dimensions. In which
only handpicked EUS diseased images of the fishes were considered, the proposed method or technique to
improve the diseased identification with larger accuracy as well as correctly detected diseased area. In which
extracted the features and PCA applied which is principle component analysis and converted into feature vector
Euclidian distance has been applied for classification.
International Journal of Computer Science and Information Security (IJCSIS),
Vol. 16, No. 5, May 2018
122 https://sites.google.com/site/ijcsis/
ISSN 1947-5500

II.Methodology
Extract the HOG PCA (Principal KNN(K
Image (Histogram of Component Nearest
Gradients) Analysis) Neighbour)
Figure 1a). Flow Chart of the Process through K-NN
Extract The FAST (Features PCA (Principal NN(Neural
Image from Component Network)
Accelerated Analysis)
Figure 1b). Flow Chart of the Process Neural Network
In Figure 1a) shows the Flow chart of Process and steps applied to extract the features and find the
performance of classification through K-NN classifier.
In Figure 1b) shows the Flow chart of Process and steps applied to extract the features and find the
performance of classification through NN (Neural Network) classifier
The processes are broadly separated into the Four stages: - Pre-processing, Feature Extraction,
dimensionality reduction and classification.
Stage 1:- Pre-processing- Real Images have been collected and remove the noise after that segmentation has
been applied.
Stage 2:- Feature Extraction- In image processing extracting the features from the image, it is not possible to
extract the feature from a single pixel, it interact with the neighbours also, feature extractor used to extract
the feature from the image of EUS (Epizootic Ulcerative Syndrome) infected fish.
Stage 3:- Dimensionality Reduction:- After extracting the feature from HOG and FAST ,PCA (Principal
Component Analysis) will apply for the dimension reduction of the features and amount of memory used by
the data, It helps in faster classification also.
Stage 4:-Classification: - Classify the fish image into EUS Infected and Non-EUS infected through classifier
e.g. KNN(K-Nearest Neighbor) and NN(Neural Network) and find the accuracy as dataset has EUS and
Non-EUS infected fish image both.
2.1 HOG (Histogram of Gradients):-
It is based on the concept that divide the image into small area called cells and then form the blocks through
cells e.g. 4*4 pixel size cell was selected by default and blocks size is 8*8 then Calculate the edge gradients
e. g from each of the local cells 8 orientations are calculated and form the histogram of cell then normalize
it and normalize the blocks also, small changes are done in the position of window in order to not to see the
descriptor changing heavily and to get the lesser impact far from centre gradients of the descriptors. For each
pixel in order to assign magnitude weight one half of the width of descriptor known as sigma is assigned
HOG Steps (Histogram of Gradients) in Matlab
Implementation Step1:- Input the image of EUS infected fish.
Figure 2: Input Image
Step 2:- Normalize the image or gamma which is the square root of image intensity depends on what kind of the
image.
Step 3:- Orientation of gradient and its magnitude is computed.
Vol. 16, No. 5, May 2018
ISSN 1947-5500

Figure 3: Gradient computed Image
Gradient Magnitude (1)
Gradient Direction (2)
Where:
is the derivative w.r.t x (gradient in the x direction)
is the derivative w.r.t y (gradient in the y direction).
Step 4:- Create and split the window into cells and each cell represents the pixels and make the histogram of
orientation gradient.
Step 5:- Grouping the cell together into large and then normalize it.
Step6:- After extracting feature from HOG apply the Machine learning algorithm or classifier.
Figure 4: HOG descriptor Image
In Figure 4 Applied the HOG (Histogram of Gradients) to extract the features and then evaluated the
performance of classification through Machine Learning algorithm.
2.2 FAST (Features from Accelerated Segment Test):-
Fast technique recognizes the interest point in an image basis intensity of local neighbourhood. It is the fastest
and better algorithm than others, the identification of corners has been given priority over the edges[8], because
they claimed that the corners have the most innumerable features which show a strong two-dimensional
intensity change, and therefore the neighbouring points as well as the work of the algorithm, it makes pixels
comparable to a fixed radii circle and to classify a point as a corner if a circle with maximum numbers of pixels
on its radii can be drawn which are brighter or darker than its central point. The detector's main limitation here
of is that almost all the features are closer to each other.
In figures 5 shows the original image and then after applied the FAST
Figure 5 a):- Original Image Figure 5 b):- FAST (Features from Accelerated Segment Test)
2.3 PCA (Principle component analysis):-After extracting features from the HOG (Histogram of
Gradients) and FAST (Features from Accelerated Segment Test), PCA applied as features reduced by
the PCA because it is to reduce the dimensionality of the dataset and by reducing the number of
dimensions, it utilizes less space. It helps in classification on large dataset as it takes less time. After
reducing the feature space some noise and redundancies in the features are eliminated while reduce the
dimensionality.
Vol. 16, No. 5, May 2018
ISSN 1947-5500

2.4 HOG-PCA & FAST-PCA: - Feature vector dimensionality reduction is the work of PCA. Then on
to the extracted features we apply PCA a better accuracy is found in the case of FAST -PCA
application then in HOG-PCA. When FAST-PCA applied with the Machine learning algorithm it gives
3.8% classification accuracy higher as compared to the HOG-PCA. (Result shows in Figure 9 and 11).
2.5 KNN: - It is a supervised learning algorithm and is usually used in machine learning methods. The
best way to classify the feature vectors is basis the closest training. Being an easy and efficient that
depends on the known samples it is an important non parametric classification approach, depending on
the known samples, according to the approximate neighbours of K-nearest which classify and specifies
a class label for unknown samples. (x, F(x)) are being stored as examples of training. Being an input in
memory an n-dimensional vector (a1, a2,...,a) is termed as x and corresponding output is F(x) that is
classified basis its neighbours as per their size for classification,, the value of K-nearest has been
chosen[18] if K-nearest = 1, the only to the class of its neighbours the object is assigned, the it can
reduce the effect of noise on the major value classification of K-nearest, but can separate the
boundaries between the classes.
KNN is classified into Testing and Training Phase for classification:-
Training phase:
1) Select the images for training phase.
2) After that training images will read.
3) Pre-process and resize the each image.
4) Preprocessed image was used to extract the features (through HOG) to form a vector of features of
image that are local to the image.
5) By the local features, feature vector is constructed of the image as row in a matrix.
6) Repeat steps 2 to Step 5 for all the training images.
7) Trained the KNN technique for the phase of testing.
1) Read the images for test.
2) After applied the KNN first, identified the nearest neighbours using the function of Euclidean
distance by the Training data.
3) If the K neighbours have all the same labels, the image is labelled and exit otherwise compute pair
wise distances between the K neighbours and construct the distance matrix.
III. Proposed Methodology:-
ClassificationExtract the Pre- Feature
image (EUS Processing Extraction
PCA Through
Infected) through
NEURAL
FAST
Network (NN)
Figure 6:- Proposed Flow chart
In Figure 6, it shows the Proposed Flow chart or steps to be implemented to extract the features and find
the classification performance through Machine Learning Algorithm (Neural Network)
EUS disease detection from the image, first apply the morphological operations i.e. The image is converted into
greyscale and enhances the image, remove the noise and segmentation applied and then extract the feature from
FAST then apply the PCA dimensionality reduction of the extracted features, match the features after applying
the classifier which is neural network and find the classification accuracy.
The algorithm explained below of combination (FAST-PCA-NN) method:
1. A pixel is selected which is considered as “pe” in the image and assumed “IPE” the intensity of the
pixel.
// Meaning of a pixel under test i.e., it is an interest or feature point which is to check.
2. T is taken as the threshhold intesity set with the assumption that it will be around 19-20 percent of
the available pixels.
Vol. 16, No. 5, May 2018
ISSN 1947-5500

3. IPE is assumed to be the pixel intensity of the 16 pixels of circle surrounding th epixel "pe".
((Bresenham circle [15] of radii 3.)
4. Threshold will distinguish the "N" pixels adjacent to the 16 pixel by checking if they are above or
below it..
// (N = 12 as a part of the first form of the algorithm)
5. The intensity of pixels is comparing the 1, 5, 9 and 13 of the circle with IPE (Intensity of pixel)
first The algorithm that considered will be fast; it should be that no less than the three pixel
combination should follow rule 4, so that the interest point will exist.
6. The pixel “pe” will not be considered as an interest point or corner in it that not less than the three of
the above mentioned four pixel values I1, I5, I9, I13 are neither above nor below Ipe + T. hence, in
such situations, pixel “pe” will be rejected from considering a corner point. Only if the least 14 least
3/4 the of the pixel are considered to be falling under that criteria.
7. Then the process will repeat for all image pixels.
8. After that Apply the PCA (Principal component Analysis) for reducing the dimensions.
9. After that applied the Neural Network to train and test the image
a) Take X as a variable and X= features (Input Data) // Extracted features from FAST
Algorithm. [Input, Targets]=Datasets;
b) Create the Pattern Recognition Network
c) Then divide the data for Training ,Validation and Testing
d) Setup the data into training, validation.
e) Setup the division of data for Training ,Validation ,
f) Then the Network will be train and can be test the network after trained.
3.1 Sample of training Dataset:-
Figure 7:-Sample of Training Dataset (EUS Infected Fish)
In Figure 7 shows the Sample of EUS infected fish,The sample of EUS infected fish used in experimentation
are the real images.
3.2 Performance Comparison between the combined Techniques
Classification Accuracy Percentage
HOG-PCA-KNN 56.32%
HOG-PCA-NN 92.5%
FAST-PCA-KNN 63.32%
FAST-PCA-NN 96.3%
Vol. 16, No. 5, May 2018
ISSN 1947-5500

Table: 1:- Comparison between classifications Accuracy of different combination Techniques
The Table 1 shows among all combination Technique, the Proposed Combination Technique shows
96.3% accuracy as it gives higher accuracy as compared to others in the paper.
3.2 a) Graph between All combinations:-
Figure 8:- Classification Accuracy between Combined Technique
In Figure 8 the Graph shows the Performance comparison between Existing and Proposed Combination
of Technique.
Performance Evaluation Accuracy - TP+TN (3)
TP+TN+FP+FN
In Performance Evaluation Accuracy find the positive and negative rate to classify the EUS and Non-
EUS infected fish.
3.3 Accuracy through HOG-PCA-NN and FAST-PCA-NN:-
After applying the Feature extraction through HOG and FAST and get the classification accuracy through
Neural Network. It extract the 4356 features in order to get a neural network to successfully learn task, it must
be trained first. The training database is then divided into testing set and training set. Neural network was
trained using the train set. To get the better result train the neural network many times and get the average of
classification accuracy. In which input or feature extracted by the feature extractor is 4356 and has taken 10
hidden layers which give the output. Testing set is used to test the neural network. To find the hidden neurons,
in an architecture the dataset is partitioned into test and train data Ttrain sets Ttesting[16]. The test set is used to
test the ability of the network [1]. Network pattern recognition is be implemented.
3.4 Results and Analysis:-
3.4 a) The Result shows the classification accuracy through HOG-PCA-NN
Figure:-9 Confusion Matrix (HOG-PCA-NN)
Vol. 16, No. 5, May 2018
ISSN 1947-5500

Figure 9 shows the confusion matrix of HOG-PCA-NN and gives the performance accuracy with Non-EUS and
EUS fish.
Figure:-10 Receiver Operating Characteristic Curve
In Figure 10 shows the ROC known as Receiver’s Operating Characteristic curve of (HOG-PCA-NN)
which gives the graph between True Positive Rate Vs False Positive Rate with the EUS and Non-EUS
infected fish.
3.4 b):- FAST-PCA-NN
The results shows the classification accuracy through FAST-PCA-NN
Figure 11:-Confusion Matrix (FAST-PCA-NN
In Figure 11 shows the Confusion Matrix of (FAST-PCA-NN) which gives 96.3 % accuracy in correct detection
of EUS disease fish and 3.7% not correctly classified ,the graph shows the Target class Vs output class, It tells
the False positive(FP) and False Negative(FN), True positive(TP) and True Negative(TN).(Performance
Accuracy)
Vol. 16, No. 5, May 2018
ISSN 1947-5500

Figure 12:- Receiver Operating Characteristic Curve
In Figure 12 shows the ROC curve known as Receiver Operating Characteristic Curve (FAST-PCA-NN) shows
the Receive Operating Specification curve which gives the graph between True Positive Rate (TPR) Vs False
Positive Rate (FPR) with the EUS and Non-EUS infected fish.
The ROC curve area shows the perfectly prediction when it comes 1 as said by the properties of ROC curve in
Figure 12; an area of .5 represents a worthless detection or random prediction.
IV.Conclusion
The Experimental evaluation for performance Comparison shows the proposed combination (FAST-PCA-NN)
gives better accuracy as compared to the other combinations in the paper. Proposed combination gives 3.8%
better accuracy than other (HOG-PCA-NN) when it combines with PCA because it reduces the dimensionality
of the dataset by reducing the number of dimensions. The Experimentation has been done on MATLAB
Environment and on real images of EUS Infected.
Acknowledgement
Collection of the EUS infected fish’s images have been done from National Bureau of Fish Genetic Resources
(NBFGR, Lucknow) and ICAR-(CIFRI), Kolkata. Thanks to Dr. A.K Sahoo (CIFRI, Kolkata) and Dr .P.K
Pradhan (NBFGR, Lucknow).
References
[1] A. Jeyanthi Suresh , P.Asha ,”Human Action Recognition in Video using Histogram of Oriented Gradient
(HOG) Features and Probabilistic Neural Network (PNN)”, International Journal of Innovative Research in
Computer and Communication Engineering, Vol. 4, Issue 7, July 2016
[2]. Valentin Lyubchenko,Rami Matarneh,Oleg Kobylin,Vyacheslav Lyashenko, “Digital Image Processing
Techniques for Detection and Diagnosis of Fish Diseases”, International Journal of Advanced Research in
Computer Science and Software Engineering, Volume 6, Issue 7, July 2016
[3]. Hitesh Chakravorty, Rituraj Paul & Prodipto Das , “Image Processing Technique To Detect Fish Disease”,
International Journal of Computer Science & Security (IJCSS), Volume (9) : Issue (2) : 2015 121
[4]. Navneet Dalal, Bill Triggs,”Histograms of Oriented Gradients for Human Detection”International
Conference on Computer Vision & Pattern Recognition (CVPR ’05), Jun 2005, San Diego, United States.
[5]. E. Rosten, R. Porter, T. Drummond,” FASTER and better: a machine learning approach to corner detection
“IEEE Trans Pattern Anal Mach Intel, 32 (2010), pp. 105–119
Vol. 16, No. 5, May 2018
ISSN 1947-5500

[6]. Jeong-Seon Park, Myung-Joo Oh, and Soonhee Han, “Fish Disease Diagnosis System Based on Image
Processing of Pathogens Microscopic Images”, Frontiers in the Convergence of Bioscience and
Information Technologies 2007
[7]. Mikolajczyk,Schmid, “D.Scale and affine invariant interest point detectors”, International Journal of
Computer Vision, 2008, vol. 60, no. 1, p. 63-86
[8]. Muralidharan,R.Chandrasekar, C. “Combining local and global feature for object recognition using SVM-
KNN”. In Proceedings of International Conference on Pattern Recognition. Informatics and Medical
Engineering. 2012, ISBN 978-1-4673-1039-0/12.
[9]. Rosten, E., & Drummond, T. (2006). “Machine Learning for High-speed Corner Detection”. Proceedings of
the 9th European Conference on Computer Vision - Volume Part I (pp. 430-443). Berlin, Heidelberg: Springer-
Verlag
[10]. Burge, C. A., Mark Eakin, C., Friedman, C. S., Froelich, B., Hershberger, P. K., Hofmann, E. E., Ford, S.
E. (2014). Burge, C. A., Mark Eakin, C., Friedman, C. S., Froelich, B., Hershberger, P. K., Hofmann, E. E.,
Ford, S. E.”Climate change influences on marine infectious diseases: implications for management and
society”. Annual review of marine science, 2015,6, 249-277.
[11]. Lafferty, K. D., Harvell, C. D., Conrad, J. M., Friedman, C. S., Kent, M. L., Kuris, A. M., Saksida, S. M.
(2015), “Infectious diseases affect marine fisheries and aquaculture economics”, Annual review of marine
science, 7, 471-496.
[12]. Afferty, K. D., Harvell, C. D., Conrad, J. M., Friedman, C. S., Kent, M. L., Kuris, A. M., Saksida, S. M.
,”Infectious diseases affect marine fisheries and aquaculture economics.”,Annual review of marine
science, 2015,7, 471-496.
[13]. Narasimha-Iyer, H., et al., “Automatic Identification of Retinal Arteries and Veins From Dual-Wavelength
Images Using Structural and Functional Features. Biomedical Engineering”, IEEE Transactions on, 2007.
54(8): p. 1427-1435.
[14]. Daoliang Lia, Zetian Fua, Yanqing Duanb,”Fish-Expert: a web-based expert system for fish disease
diagnosis”, Expert Systems with Applications 23 (2002) 311–320.
[15]. A. Alahi, R. Ortiz and P. Vandergheynst, “FREAK: Fast Retina Keypoint,” IEEE Conference
on Computer Vision and Pattern Recognition, 2012.
[16]. J.-Y.Chang and J.-L. Chen, “Automated facial expression recognition system using neural networks”,
Journal of the Chinese Institute of Engineers, vol. 24, no. 3, (2001), pp. 345-356.
[17]. M. Rizon, M. F.Hashim, P. Saad, S. Yaacob, “Face Recognition using Eigenfaces and Neural Networks”,
American Journal of Applied Sciences, vol. 2, no. 6, (2006), pp. 1872-1875. .
[18] LI LiLi, ZHANG YanXia & ZHAO YongHeng k-Nearest Neighbors for automated classification of celestial
objects” Science in China Series G: Physics, Mechanics & Astronomy,jul,2008,vol. 51,no.7,pp 916-922
Vol. 16, No. 5, May 2018
ISSN 1947-5500

Comparison of Machine Learning Techniques for Identification of Disease

More Related Content

What's hot (20)

Similar to Comparison of Machine Learning Techniques for Identification of Disease (20)

Recently uploaded (20)

Comparison of Machine Learning Techniques for Identification of Disease