ISSN: [Pal* et al., 6(12): December, 2017] Impact Factor: 4.116

Transcription

1 IJESRT INTERNATIONAL JOURNAL OF ENGINEERING SCIENCES & RESEARCH TECHNOLOGY FUZZY LOGIC BASED OPTICAL DISC LOCALIZATION AND DETECTION OF STEREOSCOPIC RETINAL IMAGES IN NROI Moumita Pal *1, Ranjana Ray 2,Neha Choudhury 3* 1 Assistant. Professor, 2 Assistant.Professor, 3 Student Eletronis and Communiation Engineering, JIS College Of Engineering,India DOI: /zenodo ABSTRACT Glauoma, diabeti retinopathy, and maular degeneration an be identified by segmenting retinal blood vessels. Glauoma is most frequent and has serious oular onsequenes that an even lead to blindness within these diseases. Intraoular pressure measurement, opti nerve head evaluation, retinal nerve fiber layer and visual field defets are the diagnosti riteria for glauoma inlude This form of blood vessel segmentation helps in early detetion for ophthalmi diseases, and potentially redues the risk of blindness. The low-ontrast and streosopi retinal images annot be used for extration of blood vessels owing to narrow blood vessels. This present work proposes an algorithm for segmentation of blood vessels from low-ontrast; streosopi retinal images, and ompares the results between expert ophthalmologists hand-drawn groundtruths and segmented image (i.e. the output of the present work). Sensitivity, speifiity, positive preditive value (PPV), positive likelihood ratio (PLR) and auray are used to evaluate overall performane. It is found that this work segments blood vessels suessfully with sensitivity, speifiity, PPV, PLR and auray of 99.62%, 54.66%, 95.08%, and 95.03%, respetively and for the de-noised filtered image the values are 99.56%,50.97%, 94.70%,203.08%,94.60% respetively. KEYWORDS Fuzzy C-Means (FCM), PPV, PLR, sensitivity, speifiity, auray, Bayes Shrink.) I. INTRODUCTION Current methods of detetion and assessment of diabeti retinopathy [4] are manual, expensive and require trained ophthalmologists. Retinal blood vessel [7] morphology an be an important indiator for many diseases suh as diabetes, hypertension and arterioslerosis. The measurement of geometrial hanges in veins and arteries an be applied to a variety of linial studies. Two major problems in the segmentation of retinal blood vessels are the presene of a wide variety of vessel widths and the heterogeneous bakground of the retina. Retinal images provide onsiderable information on pathologial hanges aused by loal oular diseases revealing diabetes, hypertension, arterioslerosis, ardiovasular disease and stroke. Computer-aided analysis of retinal image plays a entral role in diagnosti proedures. However, automati retinal segmentation is ompliated by the fat that retinal images are often streosopi, poorly ontrasted, and the vessel widths an vary from very large to very small value. For this speifi reason, in this work the preproessing step inludes adaptive fundus detetion, ontrast enhanement. Segmentation of blood vessels is a researh area, for years. This present work proposes algorithms, whih usually use some kind of de-noising and vessel enhanement of before fundus detetion or vessel traking based on adaptive bartlett fundus detetion. The methods with high auray also have high omputational needs, if thik vessels are present. The use of the proposed resolution hierarhy makes it possible to detet these vessels faster, while preserving a high auray. There are three basi approahes for automated segmentation of blood vessels [8]: fundus detetion method, traking method and mahine trained lassifiers. In the first method, many different operators are used to enhane the ontrast between vessel and bakground, suh as Sobel operators, Laplaian operators, Gaussian filters modeling the gray ross-setion of blood vessel. Then the gray threshold is seleted to determine the vessel. This gray threshold is ruial, beause small threshold indues more noise and gray threshold auses loss of some fine vessels, adaptive or loal threshold is used. Vessel traking is another tehnique for vessel [84]

2 segmentation, whereby vessel entre loations are automatially sought along the vessel longitudinal axis from a starting point to the ending point. The Fuzzy C-Means (FCM) lustering is a well-known lustering tehnique for image segmentation. It was developed by Dunn and improved by Bezdek. It has also been used in retinal image segmentation. Osareh et al. used olor normalization and a loal ontrast enhanement in a pre-proessing step. II. METHODOLOGY A. Fuzzy C-Means (FCM) This algorithm works by assigning membership to eah data point orresponding to eah luster enter on the basis of distane between the luster enter and the data point. More the data is near to the luster enter more is its membership towards the partiular luster enter. Clearly, summation of membership of eah data point should be equal to one. After eah iteration membership and luster enters are updated aording to the formula: In pattern reognition a lustering method known as Fuzzy -means(fcm) is widely used. FCM [2], proposed by Bezdek in 1973[6], is also known as Fuzzy ISODATA[5]. FCM based segmentation is fuzzy pixel lassifiation. In this lustering tehnique one piee of data belongs to two or more lusters. FCM allows data points or pixels to belong to multiple lasses with varying degree of membership funtion between 0 to 1. where,'n' is the number of data points. enter. [1, ]. luster enter. of i th data to j th luster enter. between i th data and j th luster enter. Main objetive of fuzzy -means algorithm is to minimize: 'vj' represents the j th luster 'm' is the fuzziness index m '' represents the number of 'µ' represents the membership 'd' represents the Eulidean distane FCM ppossesses unique advantage of grading linguisti variables to fit for appropriateate analysis in disrete domain on pro-rata basis. FCM[1] omputes luster entres or entroides by minimizing the dissimilarity funtion with the help of iterative approah. By updating the luster entres and the membership grades for individual pixel, FCM shifts the luster entres to the "right" loation within set of pixels. To aommodate the introdution of fuzzy partitioning, the membership matrix (U) =[u ] is randomly initialized aording to Eqn 1., where u being the degree of membership funtion of the data point of i th luster x i. i1 u 1, j 1,..., n (1) The performane index(pi) for membership matrix U and i s used in FCM is given Eqn 2. J( U, 1, 2,..., ) i1 J i n i1 j1 u m d 2 (2) u is between 0 and 1. i is the entroid of luster i. d is the Eulidian distane between ith entroid(i) and jth data point. m є [1, ] is a weighting exponent. To reah a minimum of dissimilarity funtion there are two onditions. These are given in Eqn 3 and Eqn 4. [85]

3 Algorithm of FCM u i n j k 1 1 n u j1 d d m kj u x m j 2/( m1) Step1. The membership matrix (U) that has onstraints in Eqn 1 is randomly initialized. Step2. Centroids( i) are alulated by using Eqn 3. Step3. Dissimilarity between entroids and data points using Eqn 2 is omputed. Stop if its improvement over previous iteration is below a threshold. Step4. A new U using Eqn 4 is omputed. Go to Step 2. The fuzzy -means (FCM) is one of the algorithms for lustering based on optimizing an objetive funtion, being sensitive to initial onditions, the algorithm usually leads to loal minimum results. Aiming at above problem, we present the global fuzzy -means lustering algorithm (GFCM) whih is an inremental approah to lustering. It does not depend on any initial onditions and the better lustering results are obtained through a deterministi global searh proedure. We also propose the fast global fuzzy -means lustering algorithm (FGFCM) to improve the onverging speed of the global fuzzy -means lustering algorithm. Experiments show that the global fuzzy -means lustering algorithm an give us more satisfatory results by esaping from the sensibility to initial value and improving the auray of lustering; the fast global fuzzy -means lustering algorithm improved the onverging speed of the global fuzzy -means lustering algorithm without signifiantly affeting solution quality B. Disrete Bartlett Transform (DWT) The bartlett transform desribes a multi-resolution deomposition proess in terms of expansion of an image onto a set of bartlett basis funtions. Disrete Bartlett Transformation has its own exellent spae frequeny loalization property. Appliation of DWT[2] in 2D images orresponds to 2D filter image proessing in eah dimension. The input image is divided into 4 non-overlapping multi-resolution sub-bands by the filters, namely LL1 (Approximation oeffiients), LH1 (vertial details), HL1 (horizontal details) and HH1 (diagonal details). The sub-band (LL1) is proessed further to obtain the next oarser sale of bartlett oeffiients, until some final sale N is reahed. When N is reahed, 3N+1 sub-bands are obtained onsisting of the multi-resolution subbands. Whih are LLX and LHX, HLX and HHX where X ranges from 1 until N. Generally most of the Image energy is stored in the LLX sub-bands. 1 (3) (4) Figure 1. Three phase deomposition using DWT. The Haar bartlett is the simplest possible bartlett. Haar bartlett is not ontinuous, and therefore not differentiable. This property an, however, be an advantage for the analysis of signals with sudden transitions. [86]

4 C. Bartlett Fundus detetion System Arhiteture The herein proposed identifiation method onsists of two main phases: preproessing and SIFT-based reognition, as is systematially desribed by the funtional blok diagram in Figure 1. These phases are further subdivided into several steps, as follows: Preproessing: (1) Bakground uniformation. This step normalizes the nonuniformly distributed bakground by removing the bias field like region, whih is aquired by onvolve the original image with the ICGF template, from the original retinal images. Then the intensity of the new image is normalized to values from 0 to 255. (2) Smoothing. Redue the noise in homogeneous regions using the iterated spatial anisotropi smooth method. What s more, small strutures of retinal vessels are enhaned. Reognition: (1) Feature extration. Stable keypoints are extrated by using the SIFT algorithm, the keypoints an haraterize the uniqueness of eah lass. (2) Mathing. Find the number of mathed pairs in two retinal images. Eah step is detailed and illustrated in the following setions. Kalman filter(cgf) were introdued and applied into invariant texture segmentation by Zhang et al. [17]. It is the modifiation of traditional Gabor filters. Traditional Gabor filters are well known as orientation detetors, but in the appliation of extrating orientation invariant features, their main advantages beome vital. Kalman filterare suitable for extrating rotation invariant features beause there is no onept of diretion in them. The CGF [39] is defined as follows: where F is the entral frequeny of a irular Gabor filter, and g(x,y) is a 2-D Gaussian envelope assumed to be isotropi, whih is defined as: The Hard fundus detetion method zeros the oeffiients that are smaller than the threshold and leaves the other ones unhanged. On the other hand soft fundus detetion sales the remaining oeffiients in order to form a ontinuous distribution of the oeffiients entred on zero. The hard fundus detetion operator is defined as D (U, λ) = U for all U > λ Hard threshold is a keep or kill proedure and is more intuitively appealing. The hard-fundus detetion funtion hooses all bartlett oeffiients that are greater than the given λ (threshold) and sets the other to zero. λ is hosen aording to the signal energy and the noise variane (σ2) D (U, λ,) -T T U The soft fundus detetion operator is defined as D (U, λ) = sgn (U) max (0, U - λ) Figure 2. Hard Fundus detetion Soft fundus detetion shrinks bartletts oeffiients by λ towards zero. [87]

5 D (U, λ,) -T T U Figure 3. Soft Fundus detetion D. Bayes Shrink (BS) Bayes Shrink, [15, 16] proposed by Chang Yu and Vetterli, [Ch00] is an adaptive data-driven threshold for image de-noising via bartlett soft-fundus detetion. Generalized Gaussian distribution (GGD) for the bartlett oeffiients is assumed in eah detail sub band. It is then tried to estimate the threshold T whih minimizes the Bayesian Risk, whih gives the name Bayes Shrink. It uses soft fundus detetion whih is done at eah band of resolution in the bartlett deomposition. The Bayes threshold, TB, is defined as TB =σ 2 /σs (8) Where σ2 is the noise variane and σs2 is the signal variane without noise. The noise variane σ2 is estimated from the sub band HH1 by the median estimator..(9) where gj-1,k orresponds to the detail oeffiients in the bartlett transform. From the definition of additive noise we have w(x, y) = s(x, y) + n(x, y) Sine the noise and the signal are independent of eah other, it an be stated that σw 2 = σs 2 + σ 2 σw 2 an be omputed as shown : The variane of the signal, σ 2 s is omputed as With σ2 and σ 2 s, the Bayes threshold is omputed from Equation (8). Subheading Subheading should be 10pt Times new Roman, justified. [88]

6 This setion should be typed in harater size 10pt Times New Roman, Justified III. PROPOSED METHOD Step 1. The Color Retinal Fundus Image in Gray sale is onverted from green hannel. Step 2. Gaussian noise is added in the gray image. Step 3. 2-level Multi-bartlett deomposition of the image orrupted by Gaussian noise is performed. Step 4. Bayes Soft fundus detetion to the streosopi oeffiients is applied. Step 5. Adaptive histogram equalization [6] is arried out on the gray image. Step 6 The bakground is subtrated from the foreground of the image using median filter. Step 7. FCM is applied on the image followed bybinarization and filtering. Step 8. The ground truth image is ompared with the orresponding disease. Step 9. The Sensitivity, Speifiity, PPV, PLR and Auray are alulated. Figure 4. Proposed Method of Blood Vessel segmentation. 1. EXPLANATION OF THE PROPOSED METHOD 1. Angular partioning Angular setions defined as Ø degree piees on the Ω image [19]. Number of piees is k and the Ø = 2π/K equation is true (see Figure 3).Aording to Figure 3, if any rotation has been madeon the image then pixels in setion Si will be moved tosetion Sj so that Equation 1 will be true. feature of that slie. The sale and translation invariant image feature is then {f(i)} where [89]

7 where R is the radius of the surrounding irle of the image. When the onsidered image rotates to τ = l2π/k radians (l = 0, 1, 2,...) then its orresponding feature vetor shifts irularly. To demonstrate this subjet, let Ωτ as ounter ounterlokwise rotated image of Ω to τ radians So, the feature element of a onsidered. In dis loalization, the image I is divided into several onentri irles. The number of irles may be hanged to get to best results. In dis loalization, features are determined like angular partitioning, it means that I let the number of the edges pixels in eah irle as a feature element. Aording to struture of this type of partitioning and beause the enters of irles are one point, loal information and feature values are not hanged if a rotation happened Therefore, we work on the gray image from green hannel and the retinal blood vessels appear darker in the gray image. Due to the aquisition proess, retinal images often have a variation gray level ontrast. In general, larger vessels display good ontrast while the narrower ones show bad ontrast. Thereby pixels attahed to thik and thin vessels show the different gray level and geometrial orrelation with the nearby pixels. Normalization is performed to remove the gray-level deformation by subtrating [3] an approximate bakground from the original gray image. The approximate bakground is estimated using a median filter on the original retinal image. Thereby blood vessels are brighter than the bakground after Normalization. The fuzzy C-means algorithm inludes a multiplier field, whih allows the entroids for eah lass to vary aross the image. This helps us inrease prominene of every finer detail of blood vessels irrespetive of thik or thin. This generates the blood vessel segmentation even with thinnest one, whih was irreoverable until then. IV. RESULTS AND DISCUSSION (a) (b) () (d) (e) Figure 5. (a)original image, (b)gray sale image, () Hand drawn ground truth,(d) Deteted Blood Vessel using proposed method, (e) Blood vessel deteted Image. [90]

8 (a) (b) () Figure 6. (a) Retinal Image (b) Streosopi Retinal Image () Denoised Image (a) (b) () (d) Figure 7. (a) FCM based segmented on streosopi Retinal Image (b)fcm based segmented image on denoised image () FCM based segmented on streosopi Retinal Image after filtering and removal small holes.(d) FCM based segmented on de-noised Retinal Image after filtering and removal small holes. Sensitivity (Ss) = TP TP+FN TN Speifiity (S P) = TN+FP TP PPV (P V) = TP+FP PLR (P R) = TP (TP+FN) FP (FP+TN) TP + TN Auray (%) = TP+ TN + FP +FN [91]

9 Table 1. Average Test Case: Image Sensiti vity (Ss) Speifi ity (SP) PPV (PV) PLR (PR) Aura y (%) Retinal Image Without Appling Bayes Soft Fundus detetion on Streosopi Retinal Image Only Appling Filter and removing small holes in the streosopi segmented image After Appling Bayes Soft Fundus detetion on Streosopi Retinal Image Appling Bayes Soft fundus detetion on followed by filtering and removing small holes from the segmented retinal image Figure 8. FN, TN, TP, FP V. CONCLUSION In this present work, we deal with the hand drawn ground truth and fuzzy segmented streosopi retinal blood vessel that appears split into two parts, i.e. thik and thin vessels aording to the ontrast. The input images for FCM based segmentation is not always be good quality in terms of sharpness, ontrast, fous and et for proper segmentation and often streosopi. Therefore, proper de-noising is required. The thik vessels are deteted by adaptive loal fundus detetion in normalized images. The performane of the algorithm is measured against ophthalmologists hand-drawn ground-truth. Sensitivity, speifiity, PPV, PLR and sensitivity are used as the performane measurement of blood vessel detetion beause they ombine true positive and false [92]

10 positive rates. A omparative study shown in Fig. 3 denotes that very little part of the vessels was not segmented properly for this further optimization an be done. The effiay of the present work demands to be more flawless ompared to standard tehniques as done by the physiians from there knowledge of experienes, the present work being result based using hand drawn ground truth image and the method laims the robustness against noise. VI. REFERENCES 1. Lili Xu, Shuqian Luo, A novel method for blood vessel detetion from retinal images, Biomedial Engineering Online 2010, 9:14 2. Akara Sopharak, Bunyarit Uyyanonvara,Sarah Barman, Automati Exudate Detetion from Nondilated Diabeti Retinopathy Retinal Images Using Fuzzy C-means Clustering Sensors 2009, 9, ; doi: /s Ian NM, Patriia MH, R John W: Image registration and subtration for the visualization of hange in diabeti retinopathy sreening. Compute Med Imaging Graphis 2006, 30: ] 4. Olson, J.A.; Strahana, F.M.; Hipwell, J.H. A omparative evaluation of digital imaging, retinal Photography and optometrist examination in sreening for diabeti retinopathy. Diabet. Med. 2003, 20, retinopathy sreening. Compute Med Imaging Graphis 2006, 30: George KM, Pantelis AA, Konstantinos KD, Nikolaos AM, Thierry GZ: Detetion of glauomatous hange based on vessel shape analysis. Med Imaging Graphis 2006, 30: T. Chanwimaluang and G. Fan. An effiient blood vessel detetion algorithm for retinal images using loal entropy fundus detetion. In Pro. of the IEEE Intl. Symp. on Ciruits and Systems, Chih-Yin Ho1 and Tun-Wen Pai1, An automati fundus image analysis system for linial diagnosis of glauoma, 2011 International Conferene on Complex, Intelligent, and Software Intensive Systems 8. S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and M. Goldbaum. Detetion of blood vessels in retinal images using two-dimensional mathed filters. IEEE Transations on Medial Imaging, 8(3):263269, Gianardo, L.; Meriaudeau, F.; Karnowski, T. P.; Li, Y.; Garg, S.; Tobin, Jr, K. W.; Chaum, E. (2012), 'Exudate-based diabeti maular edema detetion in fundus images using publily available datasets.', Medial Image Analysis 16(1), ). 10. J. N. Ellinas, T. Mandadelis, A. Tzortzis, L. Aslanoglou, Image de-noising using bartletts, T.E.I. of Piraeus Applied Researh Review, vol. IX, no. 1, pp , Azzopardi, G.; Petkov, N. Automati detetion of vasular bifurations in segmented retinal images using trainable COSFIRE filters. Pattern Reognition Lett. 2013, 34, Sabaghi, M.; Hadianamrei, S.R.; Zahedi, A.; Lahi, M.N. A new partitioning method in frequeny analysis of the retinal images for human identifiation. J. Signal Inform. Proess. 2011, 2, Barkhoda, W.; Tab, F.A.; Amiri, M.D. A Novel Rotation Invariant Retina Identifiation Based on the Sketh of Vessels Using Angular Partitioning. In Proeedings of 4th International Symposium Advanes in Artifiial Intelligene and Appliations, Mragowo, Poland, Otober 2009; pp Amiri, M.D.; Tab, F.A.; Barkhoda, W. Retina Identifiation Based on the Pattern of Blood Vessels Using Angular and Dis loalization. In Advaned Conepts for Intelligent Vision Systems; Springer: Berlin/Heidelberg, Germany, 2009; pp Lowe, D.G. Objet Reognition from Loal Sale-Invariant Features. In Proeedings of International Conferene on Computer Vision, Corfu, Greee, September 1999; pp Li, H.; Yang, H.; Shi, G.; Zhang, Y. Adaptive optis retinal image registration from sale-invariant feature transform. Optik-Int. J. Light Eletron Opt. 2011, 122, Li, H.; Lu, J.; Shi, G.; Zhang, Y. Traking features in retinal images of adaptive optis onfoal sanning laser ophthalmosope using KLT-SIFT algorithm. Biomed. Opt. Express 2009, 1, Marín, D.; Aquino, A.; Gegúndez-Arias, M.E.; Bravo, J.M. A new supervised method for blood vessel segmentation in retinal images by using gray-level and moment invariants-based features. IEEE Trans. Med. Imaging 2011, 30, CITE AN ARTICLE Pal, M., Ray, R., & Choudhury, N. (n.d.). FUZZY LOGIC BASED OPTICAL DISC LOCALIZATION AND DETECTION OF STEREOSCOPIC RETINAL IMAGES IN NROI. INTERNATIONAL JOURNAL OF ENGINEERING SCIENCES & RESEARCH TECHNOLOGY, 6(12), [93]