ISSN: 1748-0345 (Online)
www.tagajournal.com
AN EFFICIENT SEGMENTATION OF RETINAL BLOOD VESSEL USING QUANTUM EVOLUTIONARY ALGORITHM G.V. Shrichandran1, S. Sathiyamoorthy2, P. D. Sheba Kezia Malarchelvi3 1 Research scholar, 2&3 Professor J. J. College of Engineering and Technology, Tiruchirappalli, Tamilnadu, India Emailid- gvshrichandran@gmail.com
Abstract--- Retinal blood vessels segmentation is mandatory for retinal image analysis, diagnose and treatment of specific diseases. However, the manual analysis of the retinal image is time consuming, therefore, it is necessary for developing an automatic analysis of retinal fundus images. It is achieved by employing the image processing techniques for efficient segmentation. The existing methods face certain limitation it can be overcome by implementing Quantum Evolutionary Algorithm (QEA) segmentation process based image processing techniques. Initially, the input is preprocessed for noise reduction by using Frost filter, the pre-processed image undergoes feature extraction by Gabor filter & local binary pattern to extract the retina vessel features. The extracted features are segmented by QEA and the Active Contour based Snake (ACS) method for efficient boundary detection is utilized for segmentation of blood vessels and to determine the discontinuities in the image and the Diabetic Retinopathy disease is detected from the fundus image by applying Improved Relevance Vector Machine (IRVM) is employed to classify the Diabetic Retinopathy disease. The main objective of this paper is to perform segmentation of retinal blood vessels and detection of diseases diabetic retinopathy from the retinal fundus images. The experimental results are carried out on MESSIDOR, STARE, DRIVE and VAMPIRE retinal images to estimate the performance of the proposed method in terms of certain performance metrics which shows outperforms well when compared to other techniques. Keywords---Retina Blood Vessel, Segmentation, Diabetic Retinopathy Frost filter and Quantum Evolutionary Algorithm
I.
INTRODUCTION
Retinal fundus images play an essential part in diagnosing and treatment of diabetic retinopathy, cardiovascular and ophthalmologic diseases. Diabetic retinopathy (DR) represents around five percent of the reasons for visual deficiency all around, representing to just about five million visually impaired as expressed by World Health Organization [15]. An early detection of DR is guaranteed through the regular examination of retinal images in diabetic patients, thus reducing the occurrence of blindness. Most of the retinal diseases can be distinguished through changes occurs in retinal vasculature of fundus images. In this way, segmentation of retinal blood vessels helps in identifying the presence of disease. The retina is the tissue, covering the interior surface of the eye which contains the light-sensitive cells (photoreceptors). Photoreceptors transfer light into neural signals that are passed to the brain through the optic nerves [16]. A fundus camera system is normally utilized for capturing retinal images to save the retina condition. Retinal image contains important diagnostic data which helps to identify whether the retina is diseased or non-diseased. Blood vessels appear as networks of mesh-like structure or tree-like structure of either dark red or orange-red fibers that started inside the optic disc and dynamically reducing width. The morphological characteristics of the retina such as length, width, and branching is essential for diagnosis, screening, early detection and treatment of various diseases such as stroke, vein occlusions, glaucoma, hypertension diabetes and arteriosclerosis which aid a appropriate detection and treatment in an early stage. In addition, the investigation of retinal blood vessels can support in the assessment of retinal image segmentation in terms of the relationship between vessel and hypertensive retinopathy, retinopathy of prematurity, reflectivity, arteriolar narrowing, mosaic synthesis, biometric identification, foveola vascular region detection and computer-assisted laser surgery. While cardiovascular and ophthalmologic diseases have a severe impact on human’s life, so retinal blood vessels analysis is more significant. The patients might not notice a vision loss until it turned out to be serious, consequently, early diagnosis and appropriate treatment are necessary to prevent visual hinder and blindness. It is of great significance in Š 2018 SWANSEA PRINTING TECHNOLOGY LTD
671
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
clinical applications to expose important information about retina for treatment. So it is mandatory to develop an efficient technique for segmentation of retinal vessel for diagnosis.
Figure 1 Structure of Retina
Retinal vessel segmentation can improve screening for retinopathy by decreasing the quantity of falsepositive results from the patient which is taken under various circumstances by depicting the optic disc region and fovea. In any case, manual detection of blood vessels is complicated since the blood vessels in a retinal image are obscure and low contrast. In modern healthcare community, medical imaging has turned into the most vital tool due to visual documentation and to record the patient's information for diagnosing. By employing computational approaches, for this reason, would help in the efficient retinal analysis [27]. Image processing techniques are widely used in medical diagnosis for efficient segmentation. Pre-processing, feature extraction, segmentation, and classification are the common image processing techniques used for segmentation of retinal blood vessel. Automatic segmentation of retinal blood vessels is a prevailing tool for medical diagnostics and enhance the diagnostic performance of less specialized physicians. For this reason, the segmentation method is utilized because of accurate and reliable. The important process of segmentation is to distinguish features and the background from a given image. The basic strategies for retinal blood segmentation are supervised methods and unsupervised methods [30]. Supervised methods acquire segmentation results with labeled images whereas unsupervised methods do not need a labeled image. Automatic blood vessel segmentation in the images can help speed diagnosis and improve the diagnostic performance of less specialized physicians. A crucial step in feature extraction is segmentation of the original image. This work focuses on developing a new approach for retinal blood vessel segmentation, to overcome the flaws of existing algorithms by comparing their performances to achieve efficient performance. In this paper, unsupervised method for segmentation of retinal blood vessel from color fundus retina image for classifying the image as disease affected blood vessel or nondisease [21] affected blood vessel. These proposed method based on generating a feature vector for every image pixel by utilizing training samples (known classes) to design a classifier to classify these training samples into their subsequent classes and to classify the disease as diabetic retinopathy. The QEA with ACS algorithm is utilized for segmentation of blood vessels and the disease is detected from the fundus image by applying IRVM classification technique The remaining sections of this paper are organized as various researchers on retinal blood vessel segmentation based certain techniques and its limitations has been reviewed in section 2, section 3 presents the detailed description of proposed methodology section 4 shows the experimental results and section 5 concludes this work and provides future enhancement II.
RELATED WORKS
A wide variety of approaches have been proposed for retina blood vessels segmentation. In this section review of various approaches on retina blood vessels, segmentation and classifier methods are presented. Fan et al (2017) [1] proposed an automated segmentation as hierarchical image matting model is proposed to extricate blood vessels from fundus images for a retinal blood vessel. It is accomplished by using the progression and extendibility of retinal blood vessels is coordinated into the image matting model for blood vessel segmentation. It is alluded to the procedure of accurately removing a foreground object from an input Š 2018 SWANSEA PRINTING TECHNOLOGY LTD
672
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
image. It works in two stages initial step is producing a client determined trimap and second step is applying the image matting model to separate the pixels object from the obscure locales. Generally, the matting models require the client determined trimap, which isolates the input image into three areas and the region features of blood vessels are connected to produce the trimap of the input fundus image naturally. The proposed model shows the low computational complexity and outperforms well than other supervised and unsupervised techniques as far as accuracy, which accomplishes vessel segmentation. Kaur et al (2016) [2] presented an approach of blood vessel segmentation utilizing artificial intelligence by conveying fuzzy c-means and neutrosophic set. The assessment of retinal blood vessel structure can distinguish the disease, so segmentation of retinal blood vessel vasculature is essential. This technique is proposed predominantly to classify the presence of disease or not. Several retina diseases such as cotton wool spots, exudates and lesions are identified with the assistance of region developing and neural network technique. The proposed approach is evaluated on DRIVE and DIARETDB1 databases and is compared with existing strategies it demonstrates the segmentation accuracy of 98.7% and the diseased image was identified with 99% accuracy. Automated image-based disease recognition could supplement the vital part that ophthalmologists for diagnosing diabetic retinopathy as it could diminish the time expected of these specialists to predict the disease. Hassanien et al [3] recommended the approach in which artificial bee colony optimization is utilized simultaneously with the fuzzy cluster for fitness function with fractional belongs in the main level to discover coarse vessels. This procedure utilizes two level of clustering. In the primary level, artificial bee optimization is applied on the fuzzy c-means to segment the blood vessels. In the second stage, the cluster center is additionally improved utilizing the local search to recover the blood vessels which are thin and small in diameter. The greater part of the thin vessels is separated and their performance is calculated and compared to other techniques. Hasa et al (2015) [4] developed a mathematical morphology and K-means clustering for retinal blood vessel segmentation. By examining and recognizing of vasculature structures in retinal images for early identification diabetes in advanced stages. Smoothing operation on the retinal image is done by utilizing the mathematical morphology and then enhanced an image is fragmented utilizing K-means clustering finally the blood vessels are extracted. The execution comes about shows high accuracy ratio and low misclassification ratio then different strategies. Zhao et al (2015) [5] developed a segmentation model for retina blood vessel based hybrid model. A new infinite active contour model is developed depend on hybrid region information of the image. It is accomplished by utilizing a Lebesgue measure of the neighborhood of limits, takes into consideration for efficient detection of small oscillatory (branching) structures in view of the length of a feature's limits. The benefit of utilizing diverse types of region information, such as the arrangement of intensity information and local phase based enhancement map. The execution of the proposed model is assessed by applying on three public retinal image datasets (two datasets of color fundus photography and one fluorescein angiography dataset). The proposed demonstrate that outperforms well when compared with other methods. Emary et al (2014) [6] proposed an approach for automatic vessel segmentation in retinal images. Automated segmentation of retinal vessels is commonly the initial phase in the improvement of any computeraided diagnosis framework for ophthalmic disorders. This technique utilizes possibilistic fuzzy c-means (PFCM) clustering and cuckoo search algorithm, based on brood parasitic conduct of some cuckoo species in arrangement with the Levy flight performance of birds and fruit flies, is connected to drive the enhancement of the fuzzy clustering. The execution of this algorithm is carried out on DRIVE and STARE datasets for efficient segmentation. Franklin et al [7] proposed a strategy to distinguish the retinal blood vessels with the assistance of multilayer perceptron neural network. In this method, the input is derived from the three color segments i.e., red, green and blue. By utilizing this technique the diseases such as hemorrhages and exudates can be recognized. This strategy is managed one so it requires training sets which contains manually segmented images and image features and the pixel is classified as a vessel and non-vessel. The pixel is considered as a
Š 2018 SWANSEA PRINTING TECHNOLOGY LTD
673
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
feature vector which has a place with one of the classes and likewise, the classifier decides the limit between the classes. It is a straightforward and simple strategy for early recognition of diabetic retinopathy. Hou (2014) [8] presented an automated enhancement and segmentation method for blood vessels. This method diminishes the optic disc influence and underlines the vessels by utilizing morphological multidirectional top-hat transform with rotating structuring components to the background of the retinal image. For creating a vessel response image and the last blood vessel tree, an enhanced multi-scale line detector is connected. As line locators in the multiscale finder have diverse line responses, the more drawn out detectors produce more vessel responses than the shorter line detectors. To set diverse weights for various scales, every one of the responses is joined by the enhanced multi-scale finder at various scales. This technique has average accuracy, sensitivity, and specificity on the DRIVE database 0.942, 0.735, 0.969, correspondingly. Yin et al (2013) [9] developed a probabilistic tracking-based technique for automatic vessel segmentation in retinal images and vessel edge detection. PC supported detection and examination of retinal images could help specialists, enabling them to utilize a quantitative instrument for a superior analysis, particularly while examining a tremendous measure of retinal pictures in screening programs.. At first seed points of the blood vessel are chosen consequently and Bayesian method with maximum a posteriori (MAP) as a basis is utilized to recognize vessel edge points. Finally, when all the seed points are processed, the entire vascular tree is identified. The performance results are carried out on STARE and DRIVE database to assess the execution accuracy and it indicates high accuracy and sensitivity when compared with different techniques. Fraz et al (2012) [10] utilized a 9-dimensional feature vector comprising of the rearranged gray-level of the green colour channel, the aggregate of gradient orientation maps at three scales, the sum of top hat transform reactions in eight directions utilizing linear structure element, the two most extreme reactions of two orthogonal line detectors turned in twelve angles and the four maximum responses of a Gabor filter rotated at ten angles and four scales. They utilized an ensemble classifier from two hundred bagged and boosted decision trees. The performance results demonstrate that the proposed method shows the efficient result of high classification accuracy. Problem Statement Retinal fundus images assume an essential role for diagnosing and treatment of several eye diseases. The main challenging factors are a low contrast of the fundus images is the effect of the impact that several blood vessels have diverse contrast with the background and inhomogeneous illumination of the background [12]. However, the manual recognition of the retinal fundus image is time-consuming and needs the exact information so it is vital for developing an automatic analysis of retinal fundus images. Nonetheless, retinal images are hard to interpret and computational image analysis offers the possibility to build proficiency and diagnostic accuracy of the screening process. Subsequently, automated segmentation is valuable, as it decreases the time and effort required and provides better segmentation results. In this paper, automated segmentation of retinal blood vessel is proposed and to classify the retinal diseases. III.
PROPOSED SYSTEM
The retinal blood vessel is significant for medical diagnosis to identify the diseases. Diabetic retinopathy [26] is caused by diabetes which damages the retina and leads to blindness. Segmentation of blood vessels is an aid for ophthalmologists for diagnosis. This paper presents a new automatic segmentation to overcome existing method limitations such as poor segmentation due to low contrast vessels, some vessels lie too close to each other so it became merged as one big vessel it leads to difficult for detection or to wrong vessel tracking and it is hard to detect vessel robustly in the vessel lesion. A new framework is proposed to segment the blood vessels and classifier to classify the diabetic retinopathy with high accuracy. Computer-based analysis for automated segmentation of blood vessels will help ophthalmologists by screening vessel abnormalities for diagnosis. Many image processing methods are widely utilized for retinal vessels segmentation is shown in figure 2. The input image is preprocessed by frost filter to reduce noise and it undergoes feature extraction by Gabor filter & local binary pattern. The extracted features are segmented by Quantum Evolutionary Algorithm (QEA) and the Active Contour based Snake (ACS) method is proposed for efficient boundary detection. The Š 2018 SWANSEA PRINTING TECHNOLOGY LTD
674
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
segmented image is classified by IRVM to classify the diabetes retinopathy. The following steps for proposed algorithm are Step1-The retina blood vessel of 256x256 is taken from the given each dataset. The given image is read and converted into gray image. Step2- The input images are preprocessed by the frost filter to remove the noises. Step3 - The feature extraction process is done to extract the relevant features from the images by Gabor filter and local binary pattern. Step4- The extracted feature are segmented by QEA algorithm Step5- ACS is employed for boundary detection of retina blood vessel. Step6 – The segmented images are classified by IRVM to classifying the diabetic retinopathy disease Step7- Performance measure of accuracy, Sensitivity and Specificity are computed to estimate the accurate segmentation of affected blood vessel.
MESSIDOR, STARE, DRIVE and VAMPIRE retinal images
Input image
Pre-processing by Frost Filter
Gabor filter to extract the macro texture and micro texture
Local Binary pattern to extract the color and texture features ACS for Boundary detection
QEA used to segment the vessel and nonvessel region
IRVM to classify the retinal disease of Diabetic Retinopathy
Classification result
Disease
Non-disease Figure 2 Flowchart for Proposed Method
Initialization Initially, the retina datasets [14] are collected from the publicly available MESSIDOR, STARE, DRIVE, and VAMPIRE database of size 256*256. The image is converted into grayscale image so that the vessels
Š 2018 SWANSEA PRINTING TECHNOLOGY LTD
675
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
appear brighter than the background. The blood vessel is then detected by the algorithm from the center point along the vessel direction. Preprocessing Generally preprocessing is utilized for smoothing, background homogenization and removing the noise in the retinal fundus image for improving the image contrast. Mostly retinal images are the low contrast that is hard to detect the blood vessels. During acquisition and transmission the image gets corrupted by noise, so it is mandatory to reduce noises by retaining the significant image features. Generally, the intensity or brightness of an image is uniform except when it is transferred to form an image representation and this representation generate variation in the image intensity even in the absence of image information. This variation is usually random which has no specific pattern usually occur as noise and it reduces image quality. Frost filter is a linear, convolutional filter utilized to eliminate the multiplicative noise from images. When compared with the mean and median filter it has adaptive nature and furthermore, it is an exponentiallyweighted averaging filter. Frost filter works on the base of the coefficient of variation which is the proportion of standard deviation to the local mean of the corrupted image. assume that the kernel size of m*m then the center pixel value is replaced by the weighted sum of estimation of the neighborhood in the kernel. The weighting factor reduce as it leaves from involved pixel and increment with variance is represented as đ??š = ∑đ?‘šâˆ—đ?‘š đ?‘ đ?›źđ?‘’ −đ?›ź|đ?‘Ą| --- (1) 4 đ?œŽ2 đ?›ź = ( 2) ( 2) đ?‘ đ?œŽĚ… đ?œŽĚ… Where N= Normalized constant c = Local Mean đ?œŽ= Local variance đ?œŽĚ…= Image coefficient of variation value |t| = |X – X0|+|Y- Y0| s = moving kernel size Frost filter convolves the pixel values within a fixed size window with an exponential impulse response is represented as đ?‘“ = đ?‘’đ?‘Ľđ?‘?[−đ??žđ??śđ?‘Ś (đ?‘™0 )|đ?‘™|] --- (2) đ?œŽđ?‘Ś đ??śđ?‘Ś = đ?‘ŚĚ… Where K is the filter parameter, l0 represents the location of the processed pixel and |đ?‘™|is the distance measured from pixel l0. Feature Extraction This process is to extract the known feature which contains some details about the given input image. Features such as shape, histogram, texture, color, etc. are utilized to determine the information about the image. A feature is an important attribute includes image data which provides detailed information about the image to decrease the process complexity which is to be extracted. In an image not all the attributes are used for extraction. The feature extraction by Gabor filter and local binary pattern are performed to extract the retinal blood vessel. Gabor Filter Gabor filters are rapidly used for multi-directional analysis in image processing. In this research, the optimized Gabor filter is utilized for detecting the blood vessel in the retinal image. The Gabor Filters are a set of orientation and frequency sensitive bandpass filters [11] which include ideal localization in both the recurrence pattern content. The optimized Gabor filter kernels are sinusoids modulated are represented as đ?œŽđ?‘˘ = đ?‘˜ --- (3) đ?œŽ đ?œŽđ?‘Ł = đ?›śđ?‘˘ --- (4) đ?‘˘0 = đ?‘˘ đ?‘?đ?‘œđ?‘ đ?œƒ + đ?‘Ł đ?‘ đ?‘–đ?‘›đ?œƒ Š 2018 SWANSEA PRINTING TECHNOLOGY LTD
676
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
đ?‘Ł0 = −đ?‘˘đ?‘ đ?‘–đ?‘›đ?œƒ + đ?‘Ł đ?‘?đ?‘œđ?‘ đ?œƒ 1 2
đ??şđ??š0 = đ?‘’đ?‘Ľđ?‘? {− (
2 đ?‘˘đ?œƒ
đ?œŽđ?‘˘
+
(đ?›śđ?‘Ł)2đ?œƒ đ?œŽđ?‘Ł
)} cos(2�
đ?‘˘đ?œƒ đ?œ†
+ đ?œ“)--(5)
Where đ?œŽđ?‘˘ =Standard deviation of filter in u direction to estimate the bandwidth of the filter đ?œŽđ?‘Ł = Standard deviation of filter that control the orientation selectivity of the filter θ= Angle of zero gives a filter responds to vertical feature represents the filter orientation. Îť= Wavelength of the cosine factor of the Gabor filter. đ?›ś=Spatial aspect ratio, specify the ellipticity of the Gabor function Ψ = Phase offset đ?œŽđ?‘˘ is necessary because the shapes of the filter are invariant to the image pixels. Îť and Îł values are used to maintain false positive rate and Ďˆ represents (2Ď€) rotation phase. Only optimized Gabor filters with different orientations (0 to 360intervals of sixty degrees) are utilized to convolve with the preprocessing image. The magnitude of every response is reserved and combined to generate the result image. Local Binary Pattern Gabor filters result in inaccurate performance because the image is segmented many micro textures, thus affecting the accuracy of image segmentation. The accuracy of the extracted micro texture can be improved by utilizing Local Binary Patterns (LBP). This method is utilized because it is a proficient way for texture recognition. LBP is easy to implement which integrates the structural features and numerical texture analysis. LBP is invariant in opposition to monotonic gray-scale variations and it extended to rotation invariant texture recognition. The process of LBP is performed by utilizing the eight neighbors of a pixel in the given image, by the center pixel value is assigned as a threshold. Suppose if a neighbor pixel contains higher gray value than the center pixel than it is assigned as a pixel, or else it is zero. Then the center pixel code is generated by linked with the eight ones or zeros by binary code. The LBP algorithm assigned the input image pixels as n= (0,..., 7) by thresholding with every pixel in the neighborhood of 3*3 as central pixel nc, and a binary code is generated accordingly as S (đ?‘›đ?‘– − đ?‘›đ?‘? ) is represented as 1 đ?‘–đ?‘“ đ?‘›đ?‘– ≼ đ?‘›đ?‘? đ?‘ (đ?‘Ľ) = { ---(6) 0 đ?‘œđ?‘Ąâ„Žđ?‘’đ?‘&#x;đ?‘¤đ?‘–đ?‘ đ?‘’ The binomial factor of image is assigned as 2i, and then the LBP code is given as đ?‘– đ??żđ??ľđ?‘ƒđ?‘…,đ?‘– (đ?‘Ľ, đ?‘Ś) = ∑đ?‘ −1 đ?‘–=0 đ?‘ (đ?‘›đ?‘– − đ?‘›đ?‘? )2 --- (7) where nc represents the gray-level center pixel of neighborhood pixel and ni represents the gray-levels of N equally spaced pixels on an imaging center of radius R. So, a circular neighborhood of the central pixel nc is acquired by utilizing the bilinear interpolation in which the neighborhood pixel of center ni is estimated 2đ?›ąđ?‘– 2đ?›ąđ?‘– ) , −đ?‘… cos ( đ?‘… )] đ?‘…
by[ −đ?‘… sin (
it denotes a neighborhood of points i of radius R according to the central
pixel which is represented as LBPi,R. Since correlation between pixels decreases with distance, then the texture information can be achieved from the local pixel by minimizing radius R. The differences in a neighborhood pixel are N-bit binary number, ensuing in 2N distinct values for the binary pattern. The main features of LBP are to be robust against illumination variation, fast to compute, and it doesn’t require many parameters to be situated. The extracted features from the given input image are forwarded to the segmentation to segment the affected blood vessel. Segmentation Segmentation is the process of selecting only the affected blood vessels in the given input image [20] in which the images are segmented into many segments or regions, as a result, the contents in each region assured to have same features. It also depends on several features of the image may be either color or texture. Before denoising an image, it is segmented to recover the original image. The main benefits of utilizing segmentation [22] are the process of modifying the input image into another significant representation form and it is simple
Š 2018 SWANSEA PRINTING TECHNOLOGY LTD
677
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
to analyze. Here Quantum Evolutionary Algorithm (QEA) is used as segmentation process because of easy computation and high contrast to segment the blood vessels. QEA is an evolutionary computation method inspired by the ideas of quantum material science as uncertainty principle and the observer effect. Generally this method comprises of the search space and speed up convergence to identify the best search. It deals with two dissimilar populations of individuals as conventional individuals with conventional genes and population of quantum individuals, which are categorized by quantum genes. While genes generally store the data which is encoded into a single value are denoted by a scalar variable, quantum genes represent a probability density function [18]. The main idea is that individuals of a population are generated by observations of the quantum individuals, and impact of performing observations is the change of the data stored in the quantum individuals’ genes. The quantum genes probability density function is a square pulse function. The information stored in the genes is represented as a center (Îź) and the width (đ?œŽ) of the pulse. At first evolutionary process a population of identical quantum individuals is generated. Then a fixed number of observations are prepared, every observation generating a alternate conventional individual. The Pseudocode for QEA is given as begin initialize đ?‘„đ?‘˘đ?‘Ł (0) make population P(0) by observing đ?‘„đ?‘˘đ?‘Ł (0) evaluate P(0) store the best solution among P(0) while not stop criterion do râ†?r+1 make P(t) by observing đ?‘„đ?‘˘đ?‘Ł (t-1) population evaluate P(t) update đ?‘„đ?‘˘đ?‘Ł (t) using quantum store the best solution among P(t) Segmented blood vessel_Selection(đ?‘„đ?‘˘đ?‘Ł ) Print(Result) end while end
To acquire gene value from a quantum gene, a random real number r, in the interval [0, 1] is created. The gene value y is represented as −1 đ?‘Ś = đ?‘„đ?‘˘đ?‘Ł (đ?‘&#x;)--- (8) −1 where đ?‘„đ?‘˘đ?‘Ł is the inverse of the cumulative distribution function Quv connected to the probability distribution function quv (from the above equ 8, and the gene parameters are stored in the quantum gene. The indexes v and u identify the vth quantum gene of the uth quantum individual. đ?‘„đ?‘˘đ?‘Ł (đ?‘Ľ) = âˆŤ đ?‘žđ?‘–đ?‘— (đ?‘Ľ)đ?‘‘đ?‘Ľ-- (9) A fitness value is resolved through the fitness function is assigned to each conventional individual generated. After the conventional population has been made, a reproduction phase starts, in which new ordinary individuals are created from the fittest individuals of the conventional population. In addition, the fittest individuals created in the reproduction phase substitute the bad evaluated individuals in the search population. The quantum individuals’ genes are then updated so as to map the most encouraging areas of the search space by moving the particular probability density functions towards the (conventional) gene values of the better-assessed individuals. New observations are acknowledged, and the procedure is repeated until the point when a specific number of generations of individuals have been created. During this process, the probability density functions for the correspondent quantum genes of the quantum individuals tend to converge to meet the same function, and this function tends to a unit impulse function. The best regular individual at the last generation is taken as the final solution which comprises of segmented blood vessels. Active Contour Snake
Š 2018 SWANSEA PRINTING TECHNOLOGY LTD
678
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
The boundary detection method is used to segment the boundary of retinal blood vessels. It has been identified the discontinuity and sudden changes in edge to reduce weak boundaries for improving the image quality. In this process, Active Contour Snake (ACS) method has been used to detect the boundary detection of retina blood vessels. In each direction, snake initialization scheme has been chosen the nearest edge point from the retina boundary. After that, the continuous edge (i.e longest) edge has been chosen and then the boundary segmented. In this, the ACS arranged contour by the energy minimum in the image. Though the whole energy is monotonically minimizing, the time delayed discrete dynamic programming formulation guarantees the convergence of the active contour. Classifier Relevance Vector Machine (RVM) is used as classifier process to detect the diabetic retinopathy. The segmented retina blood vessels are forwarded to the classifier to classify the disease. The feature vectors in the segmented image are computed in diverse points as shape and texture image [24]. Training the classification systems with these features is to increase the classification rate. The classification process is performed by in which extracted features are given as input to the classification in order to classify the images. RVM method is popular because of its simple implementation and consistently high classification accuracy when applied to many real-world classification situations. They are based on a Bayesian formulation of a linear model with an appropriate prior that results in a sparse representation. As a consequence, they can generalize well and provide inferences at low computational cost. Generally, RVM [22] has been used for decision making purpose. In this system, RVM is used for blood vessels detection and classification process. Relevance vector machine (RVM) is a special case of a sparse linear model, where the basic functions are formed by a kernel function φ centered at the different training points: đ?‘Ś(đ?‘Ľ) = ∑đ?‘ đ?‘–=1 đ?‘¤đ?‘– đ?œ™ (đ?‘Ľ − đ?‘Ľđ?‘– )-- (10) While this model is similar in form to the Support Vector Machines (SVM), the kernel function here does not need to satisfy the Mercer’s condition, which requires φ to be a continuous symmetric kernel of a positive integral operator. In RVM, the weight parameter is optimized by using Active Contour Snake and it will improve the RVM performance. The extracted features are given as input in RVM training process. The training inputs denoted as {đ??šđ?‘– , đ?‘Ąđ?‘– }đ?‘›đ?‘–=1 , đ??šđ?‘– ∈ đ?‘… đ?‘› , đ?‘Ąđ?‘– ∈ {0,1} and đ?‘› − defines the number of features. For new inputs đ??šĚ‚ , RVM makes prediction based on SVM function. RVM takes a linear combination of basic kernel functions remodeled by a logistic sigmoid function đ?‘Ś(đ??šĚ‚ , đ?‘¤) = đ?œŽ(∑đ?‘›đ?‘–=1 đ?œ”đ?‘– đ?‘˜(đ??šđ?‘– , đ??šĚ‚ )) = đ?œŽ(đ?‘¤ đ?‘‡ đ??ž)---(11) đ?‘‡ Where K is defined as the kernel function vector of đ?‘˜(đ??šĚ‚ ) = [đ?‘˜(đ??š1 , đ??šĚ‚ ) ‌ đ?‘˜(đ??šđ?‘› , đ??šĚ‚ )] , đ?‘¤indicates the weight vector of (đ?œ”1 ‌ đ?œ”đ?‘› )đ?‘‡ , and đ?œŽ(. )represented logistic sigmoid function and is given below eqn 1
đ?œŽ(đ?‘Ž) = 1+exp(−đ?‘Ž) --(12) The logistic sigmoid function satisfies the symmetry property and is given below đ?œŽ(−đ?‘Ž) = 1 − đ?œŽ(đ?‘Ž) ---(13) So, RVM can be used as the posterior probability. For the input feature đ??šĚ‚ , the posterior probability of class đ?‘?1 can be defined as đ?‘?(đ?‘Ą = 1|đ??šĚ‚ ) = đ?‘Ś(đ??šĚ‚ , đ?‘¤) --- (14) Correspondingly, the posterior probability of class đ?‘?2 can be defined as đ?‘?(đ?‘Ą = 0|đ??šĚ‚ ) = 1 − đ?‘Ś(đ??šĚ‚ , đ?‘¤) --- (15) RVM can be treated as the posterior probability, because, to train the model. Also, it uses the key feature of Automatic Relevance Determination (ARD) prior to the weight vectorđ?‘¤, which is used to separate the hyper parameter đ?›źđ?‘– for each weight đ?œ”đ?‘– parameter. During the deduction process, a lot of the hyper parameters are driven to large values, therefore that equivalent weights are efficiently forced to zero. Accordingly, the corresponding kernel functions can be pruned out in a sparse model result. The inputs đ??šđ?‘– equivalent to the remaining nonzero weights is known as relevance vectors. Š 2018 SWANSEA PRINTING TECHNOLOGY LTD
679
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
The RVM decision model for an input vectorđ??šĚ‚ ,b ased on the đ?‘¤đ?‘€đ?‘ƒ and RVS vectors eqn (15) can be rewritten as đ?‘Ś(đ??šĚ‚ , đ?‘¤đ?‘€đ?‘ƒ ) = đ?œŽ(∑đ??šđ?‘–∈đ?‘…đ?‘‰đ?‘† đ?œ”đ?‘– đ?‘˜(đ??šđ?‘– , đ??šĚ‚ ) + đ?œ”0 ) --(16) In the RVM decision model, kernel function plays a significant role. In this improved RVM, the Elliptical Radial Basis Function (ERBF) can be used for kernel function. 2 2 2 (đ??š, đ?‘§) = đ?‘’đ?‘Ľđ?‘?(− ∑đ??ˇ đ?‘–=1(đ??šđ?‘– − đ?‘§đ?‘– ) /(đ?œŽđ?‘– . đ?‘&#x; )) --(17) Where đ?‘Ľ = (đ?‘Ľ1 , ‌ , đ?‘Ľđ??ˇ )đ?‘‡ and đ?‘§(đ?‘§1 , ‌ , đ?‘§đ??ˇ )đ?‘‡ )are defined as the D-dimension feature vectors, r represented as scale factor and đ?œŽđ?‘–2 defined variance. As a result retina blood vessel is segmented and classified by RVM to detect the diabetic retinopathy and this proposed system is helpful for the doctors for early detection and diagnosis. IV.
EXPERIMENTAL RESULTS
The evaluation of proposed Hybrid Quantum Evolutionary Algorithm (HQEA) and Active Contour Snake (ACS) method retina blood vessel segmentation and classification based relevance vector machine is simulated by MATLAB Tool under windows environment. The main aim of this research is to segment the retinal blood vessel and to classify the disease diabetes retinopathy at an early stage. The performance of the proposed system is measured by certain performance metrics to determine the segmentation and classification accuracy. Dataset Description The evaluation of proposed method is carried out on retina datasets which are collected from the publicly available datasets such as MESSIDOR, STARE, DRIVE and VAMPIRE retinal images for the purpose of evaluation and segmentation. All of the images in these datasets are centered at the macula, the center of the retina. These images are captured by using high-resolution camera. 



MESSIDOR - MESSIDOR (method to evaluate segmentation and indexing techniques in the field of retinal ophthalmology) dataset combines 1200 TIFF images. 800 images were acquired with pupil dilation and 400 without dilation. The images are taken at three different clinical sites. DRIVE (Digital Retinal Images for Vessel Extraction) consists of a total of 40 color retinal images which 7 are abnormal pathology cases showing exudates, hemorrhages and pigment epithelium changes. Each image resolution is 768 584 pixels. The set of 40 images was divided into a test and a training set, each containing 20 images. STARE (STructured Analysis of the Retina) contains 20 color retinal images, 10 of which show evidence of pathology. Each image has a resolution of 708 x 605 pixels with eight bits per color channel and is stored in PPM format. VAMPIRE comprises eight ultra-wide fields of view retinal angiographic images. Four of the images are from an AMD retina, while the other four are from a healthy retina.
Figure 3 Diabetic Retinopathy affected Retina
Performance Metrics The proposed method are simulated and measured by certain performance metrics obtained from the confusion matrix such as accuracy, sensitivity, specificity and execution time which is required to measure to Š 2018 SWANSEA PRINTING TECHNOLOGY LTD
680
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
classification performance and it is compared with other existing techniques such as Hybrid Cat swarm Optimization (HCSO) with Mean Shift Clustering (MSC) algorithm and hybrid Cuckoo Search with Artificial Fish Swarm (CS-AFS) algorithm. The measures are computed individually for each image and then averaged over the whole test image set. Table 1 shows the comparative analysis of performance metrics in term of accuracy, specificity, and sensitivity for all datasets and combined with existing methods.
Accuracy measures the exact segmentation of retinal blood vessel and classification of disease TP + TN Accuracy = TP + TN + FN + FP Specificity measures the negative segmentation of the diabetic retinopathy when the condition is actually not present. It is recognized as false-positive rate TN Specificity = TN + FP Sensitivity measures the positive segmentation of diabetic retinopathy when the condition is actually present. It is represented as false-negative rate, TP Sensitivity = TP + FN
Where True Positive (TP) denotes positive result of correctly identified blood vessel True Negative (TN) denotes negative result of correctly identified background blood vessel False Positive (FP) shows the positive result for background Vessels incorrectly classified as blood vessel False Negative (FN) shows the positive result for vessel pixels incorrectly classified as non-vessel
Table1 Comparative Analysis of Performance Measures Datasets MESSIDOR STARE DRIVE VAMPIRE
Accuracy (%) Hybrid Hybrid HCSOQEACSMSC ACS AFS 96 92.3 88.6 97.5 93 89.5 98 93.9 90.2 98.6 93 91
© 2018 SWANSEA PRINTING TECHNOLOGY LTD
Specificity (%) Hybrid Hybrid HCSOQEACSMSC ACS AFS 88.6 74.5 63.5 89 74 62.5 89.1 75.6 64.8 89.5 75 65.2
681
Sensitivity (%) Hybrid Hybrid HCSOQEACSMSC ACS AFS 91 81 72.3 91.5 82.3 73 90 83 74 92.5 80 75
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
QEA-ACS
HCSO-MSC
CS-AFS
100 98
Accuracy (%)
96 94 92 90 88 86 84 82 MESSIDOR
STARE
DRIVE
VAMPIRE
Retina Datasets Figure 4 Accuracy Comparison of Several Retina Datasets
QEA-ACS
HCSO-MSC
CS-AFS
90 80
Specificty (%)
70 60 50 40
30 20 10 0 MESSIDOR
STARE
DRIVE
VAMPIRE
Retina Datasets Figure 5 Specificity Comparison of Several Retina Datasets
© 2018 SWANSEA PRINTING TECHNOLOGY LTD
682
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
QEA-ACS
HCSO-MSC
CS-AFS
100 90 80
Sensitivity (%)
70 60
50 40 30 20 10 0 MESSIDOR
STARE
DRIVE
VAMPIRE
Retina Datasets Figure 6 Sensitivity Comparison of Several Retina Datasets
Figure 4-6 shows a comparison of performance metrics such as accuracy, specificity and sensitivity with existing methods and it is carried out for all datasets. It is clear that the proposed Hybrid Quantum Evolutionary Algorithm (HQEA) and Active Contour Snake (ACS) method for Vampire dataset show the efficient result of segmentation accuracy of correctly segmented blood vessels and positive and negative segmentation of retina blood vessels.
Table 2 Comparison of Execution time and AUC Execution Time (seconds) AUC Hybrid Hybrid Hybrid Hybrid Datasets HCSOHCSOQEACSQEACSMSC MSC ACS AFS ACS AFS MESSIDOR 0.21 0.26 0.35 0.912 0.8935 0.783 STARE 0.19 0.25 0.34 0.923 0.8952 0.796 DRIVE
0.20
0.24
0.42
0.9165
0.8991
0.802
VAMPIRE
0.15
0.23
0.45
0.953
0.9012
0.812
© 2018 SWANSEA PRINTING TECHNOLOGY LTD
683
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
QEA-ACS
HCSO-MSC
CS-AFS
0.45
Execution Time (seconds)
0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05
0 MESSIDOR
STARE
DRIVE
VAMPIRE
Retina Datasets Figure 7 Execution Time Comparison of Several Retina Datasets
Figure 7 shows the comparison of proposed and existing method in terms of Area under curve. The proposed method shows efficient AUC when compared to other methods. QEA-ACS
HCSO-MSC
CS-AFS
1 0.9 0.8
Area Under Curve
0.7 0.6 0.5 0.4
0.3 0.2 0.1 0 MESSIDOR
STARE
DRIVE
VAMPIRE
Retina Datasets Figure 8 AUC Comparison of Several Retina Datasets
Figure 8 shows the comparison of proposed and existing method in terms of execution time. The proposed method shows less execution time when compared to other methods. The high classification accuracy leads to less execution time for blood vessel segmentation. V. CONCLUSION Automated analysis of retinal vessels is essential for the diagnosis of a wide range of eye diseases and plays an important role in automatic retinal disease screening systems. In this paper hybrid Quantum Evolutionary Algorithm and Active Contour Snake is proposed for automatic segmentation of retinal blood vessel. This is accomplished by employing image processing techniques to achieve high efficient segmentation. Preprocessing, feature extraction and boundary detection methods are employed to segment the disease affected blood vessel and IRVM classifier is used to classify the diabetic retinopathy disease affected blood
Š 2018 SWANSEA PRINTING TECHNOLOGY LTD
684
TAGA JOURNAL VOL. 14
ISSN: 1748-0345 (Online)
www.tagajournal.com
vessel. The execution results show that the proposed HQEA-ACS shows the efficient segmentation and high classification accuracy to segment the retina blood vessel when compared to other existing methods. The obtained results can be helpful for the opthamalosists to identify and detect the diabetic retinopathy and can aid the person from vision loss. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20]
[21] [22] [23] [24] [25] [26] [27] [28] [29]
[30]
Fan, Z., Lu, J. and Li, W., 2017. Automated Blood Vessel Segmentation of Fundus Images Based on Region Features and Hierarchical Growth Algorithm. arXiv preprint arXiv:1701.00892. Kaur, I. and Singh, L.M., 2016. A Method of Disease Detection and Segmentation of Retinal Blood Vessels using Fuzzy C-Means and Neutrosophic Approach. Imperial Journal of Interdisciplinary Research, 2(6). Hassanien, A.E., Emary, E. and Zawbaa, H.M., 2015. Retinal blood vessel localization approach based on bee colony swarm optimization, fuzzy cmeans and pattern search. Journal of Visual Communication and Image Representation, 31, pp.186-196. Hassan, G., El-Bendary, N., Hassanien, A.E., Fahmy, A. and Snasel, V., 2015. Retinal blood vessel segmentation approach based on mathematical morphology. Procedia Computer Science, 65, pp.612-622. Zhao, Y., Rada, L., Chen, K., Harding, S.P. and Zheng, Y., 2015. Automated vessel segmentation using infinite perimeter active contour model with hybrid region information with application to retinal images. IEEE transactions on medical imaging, 34(9), pp.1797-1807. Emary, E., Zawbaa, H.M., Hassanien, A.E., Schaefer, G. and Azar, A.T., 2014, July. Retinal vessel segmentation based on possibilistic fuzzy c-means clustering optimised with cuckoo search. In Neural Networks (IJCNN), 2014 International Joint Conference on (pp. 1792-1796). IEEE. Franklin, S.W. and Rajan, S.E., 2014. Computerized screening of diabetic retinopathy employing blood vessel segmentation in retinal images. biocybernetics and biomedical engineering, 34(2), pp.117-124. Hou, Y., 2014. Automatic segmentation of retinal blood vessels based on improved multiscale line detection. Journal of Computing Science and Engineering, 8(2), pp.119-128. Yin, Y., Adel, M. and Bourennane, S., 2013. Automatic segmentation and measurement of vasculature in retinal fundus images using probabilistic formulation. Computational and mathematical methods in medicine, 2013. Fraz, M.M., Remagnino, P., Hoppe, A., Uyyanonvara, B., Rudnicka, A.R., Owen, C.G. and Barman, S.A., 2012. An ensemble classification-based approach applied to retinal blood vessel segmentation. IEEE Transactions on Biomedical Engineering, 59(9), pp.2538-2548. Kuri, S.K. and Kulkarni, J.V., 2015. Automated Segmentation of Retinal Blood Vessels using Optimized gabor filter with local entropy thresholding. International Journal of Computer Applications, 114(11). Sadek, I., Sidibé, D. and Meriaudeau, F., 2015, March. Automatic discrimination of color retinal images using the bag of words approach. In Medical Imaging 2015: Computer-Aided Diagnosis (Vol. 9414, p. 94141J). International Society for Optics and Photonics. Ghaderi, R., Hassanpour, H. and Shahiri, M., 2007, November. Retinal vessel segmentation using the 2-D Morlet wavelet and neural network. In Intelligent and Advanced Systems, 2007. ICIAS 2007. International Conference on (pp. 1251-1255). IEEE. Kumar, R.P., Albregtsen, F., Reimers, M., Edwin, B., Langø, T. and Elle, O.J., 2015. Blood vessel segmentation and centerline tracking using local structure analysis. In 6th European conference of the international federation for medical and biological engineering (pp. 122-125). Springer, Cham. Amrutkar, N., Bandgar, Y., Chitalkar, S. and Tade, S.L., 2013. Retinal blood vessel segmentation algorithm for diabetic retinopathy and abnormality detection using image substraction. International Journal of Advanced Research in Electrical, Electronics and Instrumentation Engineering, 2(4). Dhiravidachelvi, E. and Rajamani, V., 2014. A Survey on Anatomical Structures: in Fundus Retinal Images. Red, 1, p.3. Eisa, M., ElGamal, A., Ghoneim, R. and Bahey, A., 2010. Local binary patterns as texture descriptors for user attitude recognition. IJCSNS International Journal of Computer Science and Network Security, 10(6), pp.222-229. Mani, A. and Patvardhan, C., 2010. An adaptive quantum evolutionary algorithm for engineering optimization problems. computing, 1(22). Zhang, J., Li, H., Nie, Q. and Cheng, L., 2014. A retinal vessel boundary tracking method based on Bayesian theory and multi-scale line detection. Computerized Medical Imaging and Graphics, 38(6), pp.517-525. Cornforth, D.J., Jelinek, H.F., Cree, M.J., Leandro, J.J.G., Soares, J.V.B. and Cesar, R.M., 2009. Evolution of retinal blood vessel segmentation methodology using wavelet transforms for assessment of diabetic retinopathy. In Intelligent and Evolutionary Systems (pp. 171-182). Springer, Berlin, Heidelberg. Saxena, S. and Sharma, V., 2016. Retinal Blood Vessel Segmentation Using Graphical User Interface (GUI) for Diagnosis of Disease. International Journal of Innovative Research in Science, Engineering and Technology, 5, pp.7290-7295. Li, M., Ma, Z., Liu, C., Zhang, G. and Han, Z., 2017. Robust Retinal Blood Vessel Segmentation Based on Reinforcement Local Descriptions. BioMed research international, 2017. Sanjani, S.S., Boin, J.B. and Bergen, K., 2013. Blood Vessel Segmentation in Retinal Fundus Images. Guedri, H., Ben Abdallah, M., Echouchene, F. and Belmabrouk, H., 2017. Novel Computerized Method for Measurement of Retinal Vessel Diameters. Biomedicines, 5(2), p.12. Fraz, M.M., Remagnino, P., Hoppe, A., Uyyanonvara, B., Rudnicka, A.R., Owen, C.G. and Barman, S.A., 2012. Blood vessel segmentation methodologies in retinal images–a survey. Computer methods and programs in biomedicine, 108(1), pp.407-433.\ Amin, J., Sharif, M. and Yasmin, M., 2016. A Review on Recent Developments for Detection of Diabetic Retinopathy. Scientifica, 2016. Joshi, S. and Karule, P.T., 2012. Retinal blood vessel segmentation. International Journal of Engineering and Innovative Technology (IJEIT), 1(3), pp.175-178. Ricci, E. and Perfetti, R., 2007. Retinal blood vessel segmentation using line operators and support vector classification. IEEE transactions on medical imaging, 26(10), pp.1357-1365. Dua, S., Kandiraju, N. and Thompson, H.W., 2005, April. Design and implementation of a unique blood-vessel detection algorithm towards early diagnosis of diabetic retinopathy. In Information Technology: Coding and Computing, 2005. ITCC 2005. International Conference on (Vol. 1, pp. 2631). IEEE. Bhuiyan, A., Nath, B., Chua, J. and Kotagiri, R., 2007, September. Blood vessel segmentation from color retinal images using unsupervised texture classification. In Image Processing, 2007. ICIP 2007. IEEE International Conference on (Vol. 5, pp. V-521). IEEE.
© 2018 SWANSEA PRINTING TECHNOLOGY LTD
685
TAGA JOURNAL VOL. 14