Abstract The change in morphology, diameter, branching pattern and/or tortuosity of retinal blood vessels is an important

Transcription

1 A Supervised Method for Retinal Blood Vessel Segmentation Using Line Strength, Multiscale Gabor and Morphological Features M.M. Fraz 1, P. Remagnino 1, A. Hoppe 1, Sergio Velastin 1, B. Uyyanonvara 2, S. A.Barman 1 1 Faculty of Science Engineering and Computing, Kingston University, London, United Kingdom 2 Department of Information Technology, Thammasat University, Thailand {moazam.fraz,p.remagnino,a.hoppe,sergio.velastin,s.barman}@kingston.ac.uk, bunyarit@siit.tu.ac.th Abstract The change in morphology, diameter, branching pattern and/or tortuosity of retinal blood vessels is an important indicator of various clinical disorders of the eye and the body. This paper reports a supervised methodology for segmentation of the retinal vasculature from ocular fundus images. A 7-D feature vector is constructed by computing the outputs of morphological linear operators, line strengths and oriented Gabor filters at multiple scales. The feature vector encodes the spatial intensity measures along with vessel geometry at multiple scales. A Bayesian Classifier; the Gaussian Mixture Model is used for classification of the retinal image into vessels and non-vessel pixels. The methodology is evaluated using the images of two publicly available databases, the DRIVE database and the STARE database. Method performance on both sets of test images is better than the 2 nd human observer and other existing methodologies available in the literature. I. INTRODUCTION The blood vessels in retinal images encapsulate considerable information on pathological changes induced by ophthalmologic disorders such as diabetes, hypertension and arteriosclerosis [1]. Computer-aided analysis of retinal images plays an important role in diagnosis, treatment, screening, evaluation and the clinical study of ocular disease. However, the automated segmentation of anatomical structures in the retina is a complicated task due to the presence of lesions and noise, uneven illumination, drift in intensity, lack of image contrast, varying vessel width and central vessel reflex. A substantial body of work has been performed to address automated retinal vessel segmentation which can be classified into techniques based on matched filtering, morphological processing, vessel tracking, multiscale analysis, pattern recognition and model based algorithms. Early methods of retinal vessel segmentation are based on matched filtering (MF) which was first proposed by Chaudhuri [2] and later adapted and extended by Hoover [3] and Jiang [4]. Gang and Chutatape [5] evaluated the suitability of the amplitudemodified second order Gaussian filter whereas Zhang et al. [6] proposed an extension and generalization of the MF with a first-order derivative of Gaussian. A hybrid model of the MF and ant colony algorithm for retinal vessel segmentation was proposed by Cinsdikici et al. [7]. The algorithms based on mathematical morphology for identifying vessel structures have the advantage of speed and noise resistance. Zana and Klein [8] combined morphological filters and cross-curvature evaluation to segment vessel-like patterns. Mendonca et al. [9] detected vessel centerlines in combination with multiscale morphological reconstruction. Fraz at el. [10] combined vessel centerlines with bit planes to extract the retinal vasculature. The tracking based approaches [11] segment a vessel between two points using local information and work at the level of a single vessel rather than the entire vasculature. The multiscale approaches for vessel segmentation are based on scale-space analysis [12]. The model based approaches utilize the vessel profile models [13], active contour models [14] and geometric models based on level sets [15] for vessel segmentation. The supervised segmentation methods utilize ground truth data for the classification of vessels based on given features. These methods include the use of back propagation neural networks [16, 17] and the feature vector is composed of 200 features computed in a sub window by using principal component analysis (PCA) and corresponding edge strength. Niemeijer [18] extracted a feature vector for each pixel that consists of the Gaussian and its derivatives up to order two at multiple scales, augmented with the green plane of the RGB image and then, uses a K-nearest neighbour algorithm to estimate the probability of the pixel belonging to a vessel. Staal [19], uses ridge profiles to compute 27 features for each pixel and applies a feature selection scheme to pick those which result in better class separability by a knn classifier. In [20], 6 features are computed by employing a multiscale analysis using a Gabor wavelet transform and Gaussian Mixture Model (GMM) Bayesian classifier. Ricci [21] used line operators and Support Vector Machine (SVM) classification with only 3 features per pixel. Osareh [22] used multiscale Gabor filters for vessel candidate identification, then the features are extracted using PCA and pixels are classified using GMM and SVM classifiers. Xu [23] combined wavelets and line operators to construct a 12 dimensional feature vector and used SVM to distinguish vessel segments. Lupascu [24] introduced a feature-based Ada-Boost classifier for vessel segmentation which utilizes a 41-D feature vector at different spatial scales for each pixel. In [25], a 7-D feature vector is computed by combination of moment invariant and gray level features and a five layer feed forward neural network is used for classification. X. You [26] computed the feature vector by using the steerable complex wavelet followed by calculating the line strength [21], the SVM is used for semi-supervised classification. In this paper, a supervised algorithm for automated segmentation of retinal vessels is presented. The features are extracted using the combination of a multiscale Gabor filter /11/$ IEEE 410

2 [27], the line operators [21] and directional morphological transformation [10]. The Gaussian mixture model, a Bayesian classifier, is used to segregate the image into the vessels and non-vessels. The methodology is evaluated using the images of two publicly available databases, DRIVE [19] and STARE [3]. A number of performance measures are selected for evaluation of this algorithm which attains the highest area under the ROC among the published methods with high accuracy, sensitivity and specificity. The paper is organized as follows. Section II illustrates the proposed methodology in detail. Experimental results of the algorithm on the image data sets are discussed in Section III, and the paper is concluded in Section IV. II. THE METHODOLOGY A diagram illustrating the supervised vessel segmentation methodology is shown in Fig 1. The features are extracted from the inverted green plane of the colored retinal images, which has a higher contrast between the vessels and the background than the other channels. The vessel features of the training images are labeled using manual segmentation and are used to train a Bayesian Classifier; the Gaussian Mixture Model (GMM) classifier. The classifier will be applied to the features generated from test images to compute the segmented vascular tree. The feature vector consists of the outputs from multiscale Gabor filters [27], the line strength measures [21] of the pixels and the morphological operators [10]. A. Multiscale Gabor Features A Gabor filter is a linear filter and has been broadly used for multi-scale and multi-directional edge detection. The Gabor filter can be fine tuned to particular frequencies, scales and directions and therefore acts as low level feature extractor and background noise suppresser. The impulse response of a Gabor filter kernel is defined by the product of a Gaussian kernel and a complex sinusoid. It can be expressed as, exp 0.5 exp 2π (1) where, is the wavelength of sinusoidal factor, θ is the orientation, is the phase offset, σ is the scale of the Gaussian envelope, γ is the spatial aspect ratio, x xcosθ y sinθ and y xsinθ ycosθ. The Gabor filter response to the inverted green channel of the colored retinal image is obtained by a 2-D convolution operator and is computed in the frequency domain. The detailed procedure can be seen in [27]. The maximum filter response over the angle θ, spanning [0, π] in steps of π/18 is computed for each pixel in the image at different scales (σ={2,3,4,5}). The maximum response across the scales and orientation is taken as pixel feature vector which is illustrated in Fig 2(c)-(f). B. Line Strength Features. The retinal vasculature is composed of arteries and veins appearing as piecewise linear features, with variation in width and their tributaries visible within the retinal image. The concept of employing line operators for detection of linear structures in medical images is introduced in [28] which is modified and extended in [21] to incorporate the morphological attributes of retinal blood vessels. The average grey level is measured along lines of a particular length passing through the pixel under consideration at different orientations. The line with the highest average gray value is marked. The line strength of a pixel is calculated by computing the difference in the average gray values of a square sub-window centered at the target pixel with the average gray value of the marked line. The calculated line strength for each pixel is taken as pixel feature vector and is illustrated in Fig 2(g). Fig 1. Schematic overview of methodology /11/$ IEEE 411

3 (a) (c) (e) (g) (h) Fig.2. The feature vector (a)retinal Image (b)green Channel (c)- (f)gabor Filter response at scales 2,3,4,5 respectively (g) Line strength of pixels (h)sum of linear top hat transformation. The images in (b)-(h) are inverted for better visibility. C. Morphological Operator Features Mathematical morphology is a nonlinear tool in image analysis which has revealed itself as a very useful technique for quantifying retinal pathology. The basic morphological operations of opening, closing, erosion and dilation of a digital image with a structuring element are defined in [29]. The morphological opening operator, which is actually erosion followed by dilation, acts as a shape filter, erasing objects from the image which are smaller in size than the used structuring element, thus approximating the background. (b) (d) (f) Therefore enhancement of objects which are erased by the opening operation can be achieved by subtracting the opened image from the original image. Without loss of generalization, the assumption is that a vessel is a bright pattern with a Gaussian shape cross section profile on the dark background, and is piecewise connected and linear. Because of the piecewise linear nature of vessels, the morphological filters with linear structuring elements are used to enhance the vessels in the retinal image. The morphological opening using a linear structuring element oriented at a particular angle will eradicate a vessel or part of it when the structuring element cannot be contained within the vessel. This happens when the vessel and the structuring element have orthogonal directions and the structuring element is longer than the vessel width. Conversely, when the orientation of the structuring element is parallel with the vessel, the vessel will stay nearly unchanged. θ th θ e (2 a) θ th th (2 b) The morphological top-hat transformation is shown in equation (2-a) where I th is the top-hat transformed image, I is the image to be processed and S e is structuring elements for morphological opening, o and θ is the angular rotation of the structuring element. If the opening along a class of linear structuring elements is considered, a sum of top-hat along each direction will brighten the vessels regardless of their direction, provided that the length of the structuring elements is large enough to extract the vessel with the largest diameter. Therefore, the chosen structuring element is 21 pixels long 1 pixel wide and is rotated at angle spanning [0- π] in steps of π/8. Its size is approximately in the range of the diameter of the largest vessels in the retinal image. The sum of top-hat is depicted in equation (2-b), where Is th is the sum of the top-hat transformation performed with structuring element oriented at θ degrees. The set A can be defined as x 0 x π & π/8 0. In the image, every isolated round and bright zone whose diameter is less than the length of the linear structuring element have been removed. The sum of the top-haenhance all vessels whatever their direction, including small on the filtered image will or tortuous vessels as depicted in Fig 2(h). The features computed from multiscale Gabor filters and line strength measures have a ringing effect around the areas containing pathology in the retinal image as observed in Fig. 3(b) and Fig. 3(c). Retinal image pathologies are successfully taken care by a morphological top-hat transformation as illustrated in Fig. 3(d). D. Supervised Classification. The final segmentation is obtained by using supervised classification to divide image pixels into two pixel classes, the vessel pixels (C 1 ) and nonvessel pixels (C 2 ). The Gaussian Mixture Model (GMM) classifier is employed for this purpose. GMM is a Bayesian classifier and is commonly used in a variety of computer vision applications for classification tasks [20, 22]. In this classifier, the class-conditional probability density function of the observation vector with respect to the /11/$ IEEE 412

4 (a) (c) (d) Fig.3. (a) The green channel of retinal image with pathologies (b)- (d) Vessel features computed from Gabor filter, Line strength and Morphological operator respectively. different classes is described as a linear combination of Gaussian functions, which can be modelled as, (3) Where, K represents the number of Gaussians modelling the class conditional probability density function which is also known as the likelihood. The Gaussian parameters and weights are estimated with the expectation maximization algorithm [30] for each class C i and the given number of K Gaussians. III. RESULTS AND DISCUSSION The methodology has been evaluated on two publicly available DRIVE and STARE databases. The manual segmentation of vessels by two different observers is available with the databases. The 1 st human observerr is selected as the ground truth for training of the classifier. For the DRIVE database, the training set was formed by randomly choosing (5*10 5 ) pixel samples from the 20 labeled training images. The same number of pixel samples is chosen from the first 10 images of the STARE database. In binary classification, any pixel which is identified as vessel by the algorithm and is also marked as vessel in the ground truth is a true positive. Any pixel which is marked as vessel in the segmented image but not in the ground truth image is counted as a false positive as illustrated in Table I. TABLE I. VESSEL CLASSIFICATION Vessel Present Vessel Absent Vessel detected True Positive (TP) False Positive (FP) Vessel not detected False Negative (FN) True Negative (TN) (b) The algorithm is evaluated in terms of Area Under ROC(AUC), Accuracy, Sensitivity (SN), Specificity (SP), Positive Predictive Value (PPV), Negative Predictive Value (NPV) and False Discovery Rate (FDR) and Matthews Correlation Coefficient (MCC)[31], which is often used in machine learning and is a measure of the quality of binary (two-class) classifications. These metrics are defined in Table II based on the terms in Table I. TABLE II. PERFORMANCE METRICS FOR RETINAL VESSEL SEGMENTATION Measure Description SN TP/(TP+FN) SP TN/(TN+FP) Accuracy(ACC) (TP+TN)/(TP+ +FP+TN+FN) PPV TP/(TP+FP) NPV TN/(TN+FN) FDR FP / (FP+TP) MCC (TP.TN - FP.FN) / TP FP TP FN TN FP TN FN The segmentation result obtained after classification is thresholded to produce binary images of the vascular tree. The performance results shown in Tables III and IV were obtained using the same threshold value for all the images in the same database (0.36 and 0.30 for DRIVE and STARE images, respectively). It is observed thatt the algorithm outperforms the 2 nd human observer in almost alll of the performance metrics. TABLE III. PERFORMANCE METRICS ON DRIVE DATABASE IMAGES Im ACC SN SP PPV NPV FDR MCC AV H.O AV = Average; H.O= 2 nd Human Observer /11/$ IEEE 413

5 TABLE IV. PERFORMANCE METRICS ON STARE DATABASE IMAGES Im ACC SN SP PPV NPV FDR MCC AV H.O AV = Average; H.O= 2 nd Human Observer. The ROC curves are plotted in Fig 4 by calculating the true and the false positive fraction on all test images through threshold variation on the image produced after classification. The measured AUC is and for the DRIVE and STARE databases, respectively. The proposed method is compared with other published methods in Table V and VI for DRIVE and STARE respectively. The algorithm outperforms other methods in terms of ACC, SN, SP and AUC. Fig.4. ROC graphs for DRIVE and STARE. TABLE V. RESULTS COMPARED TO OTHER METHODS (DRIVE DATABASE) Sr. Average Area Under Methods SN SP No. Accuracy ROC 1 2 nd Human Observer Zana [8] N.A N.A 3 Niemeijer [18] Jiang[4] N.A Al-Diri [14] N.A N.A 6 Martínez P.[32] N.A N.A 7 Chaudhuri [2] N.A Mendonca [9] N.A 9 Soares [20] Staal [19] N.A Ricci[21] N.A N.A Lam[33] N.A N.A Lupascu [24] N.A Marin[25] N.A N.A Proposed Method N.A = Not Available TABLE VI. RESULTS COMPARED TO OTHER METHODS (STARE DATABASE) Sr. No. Methods Average Accuracy SN SP Area Under ROC 1 2nd Human Observer Hoover [3] Jiang[4] N.A N.A Staal [19] Soares [20] Mendonca [9] N.A 7 Marin[25] N.A N.A Lam[33] N.A N.A Ricci[21] N.A N.A Proposed 10 Method N.A = Not Available IV. CONCLUSION In this paper, a supervised classification method for automated extraction of blood vessels in retinal images has been proposed. The pixel classification is based on computing only seven features for each pixel by employing multiscale Gabor filters, line strength measures and morphological operators, thus needing shorter computation time. The total time required to classify a single image is less than one minute running on a Dell Inspiron 1564 with an Intel Corei5 CPU at 2.27 GHz and 2 GB of RAM. The methodology has been tested on two publicly available databases and has been validated against observer studies. Certain new metrics have also been evaluated. The demonstrated performance metrics show that the proposed method outperforms the 2 nd human observer as well as the other published methods, thus making it a suitable tool for incorporating into a Diabetic Retinopathy screening system to initially identify normal retinal image /11/$ IEEE 414

6 features. The vessel segmented images are available online at page.html. REFERENCES [1] H. Leung, J. J. Wang, E. Rochtchina, A. G. Tan, T. Y. Wong, R. Klein, L. D. Hubbard, and P. Mitchell, "Relationships between Age, Blood Pressure, and Retinal Vessel Diameters in an Older Population," Investigative Ophthalmology & Vis, [2] S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and M. Goldbaum, "Detection of blood vessels in retinal images using two-dimensional matched filters," Medical Imaging, IEEE Transactions on, vol. 8, pp , [3] A. D. Hoover, V. Kouznetsova, and M. Goldbaum, "Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response," Medical Imaging, IEEE Transactions on, vol. 19, pp , [4] J. Xiaoyi and D. Mojon, "Adaptive local thresholding by verificationbased multithreshold probing with application to vessel detection in retinal images," Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 25, pp , [5] L. Gang, O. Chutatape, and S. M. Krishnan, "Detection and measurement of retinal vessels in fundus images using amplitude modified second-order Gaussian filter," Biomedical Engineering, IEEE Transactions on, vol. 49, pp , [6] B. Zhang, L. Zhang, L. Zhang, and F. Karray, "Retinal vessel extraction by matched filter with first-order derivative of Gaussian," Computers in biology and medicine, vol. 40, pp , [7] M. G. Cinsdikici and D. Aydin, "Detection of blood vessels in ophthalmoscope images using MF/ant (matched filter/ant colony) algorithm," Computer methods and programs in biomedicine, vol. 96, pp , [8] F. Zana and J. C. Klein, "Segmentation of vessel-like patterns using mathematical morphology and curvature evaluation," Image Processing, IEEE Transactions on, vol. 10, pp , [9] A. M. Mendonca and A. Campilho, "Segmentation of retinal blood vessels by combining the detection of centerlines and morphological reconstruction," Medical Imaging, IEEE Transactions on, vol. 25, pp , [10] M. M. Fraz, M. Y. Javed, and A. Basit, "Retinal Vessels Extraction Using Bitplanes," presented at Eight IASTED International Conference on Visualization, Imaging, and Image Processing, Palma De Mallorca, Spain, [11] Y. A. Tolias and S. M. Panas, "A fuzzy vessel tracking algorithm for retinal images based on fuzzy clustering," Medical Imaging, IEEE Transactions on, vol. 17, pp , [12] A. F. Frangi, W. J. Niessen, K. L. Vincken, M. A. Viergever, W. William, C. Alan, and D. Scott, "Multiscale Vessel Enhancement Filtering," in Medical Image Computing and Computer-Assisted Interventation MICCAI 98, vol. 1496, Lecture Notes in Computer Science: Springer Berlin / Heidelberg, 1998, pp [13] W. Li, A. Bhalerao, and R. Wilson, "Analysis of Retinal Vasculature Using a Multiresolution Hermite Model," Medical Imaging, IEEE Transactions on, vol. 26, pp , [14] B. Al-Diri, A. Hunter, and D. Steel, "An Active Contour Model for Segmenting and Measuring Retinal Vessels," Medical Imaging, IEEE Transactions on, vol. 28, pp , [15] K. W. Sum and P. Y. S. Cheung, "Vessel Extraction Under Non- Uniform Illumination: A Level Set Approach," Biomedical Engineering, IEEE Transactions on, vol. 55, pp , [16] R. Nekovei and S. Ying, "Back-propagation network and its configuration for blood vessel detection in angiograms," Neural Networks, IEEE Transactions on, vol. 6, pp , [17] C. Sinthanayothin, J. F. Boyce, H. L. Cook, and T. H. Williamson, "Automated localisation of the optic disc, fovea, and retinal blood vessels from digital colour fundus images," British Journal of Ophthalmology, vol. 83, pp , [18] M. N. Abramoff, J.J.Staal, B. v. Ginneken, M.Loog, and M.D, "Comparative study of retinal vessel segmentation methods on a new publicly available database," presented at SPIE Medical Imaging, [19] J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, and B. van Ginneken, "Ridge-based vessel segmentation in color images of the retina," Medical Imaging, IEEE Transactions on, vol. 23, pp , [20] J. V. B. Soares, J. J. G. Leandro, R. M. Cesar, H. F. Jelinek, and M. J. Cree, "Retinal vessel segmentation using the 2-D Gabor wavelet and supervised classification," Medical Imaging, IEEE Transactions on, vol. 25, pp , [21] E. Ricci and R. Perfetti, "Retinal Blood Vessel Segmentation Using Line Operators and Support Vector Classification," Medical Imaging, IEEE Transactions on, vol. 26, pp , [22] A. Osareh and B. Shadgar, "Automatic Blood Vessel Segmentation In Color Images Of Retina," Iranian Journal Of Science And Technology Transaction B-Engineering, vol. 33, pp , [23] L. Xu and S. Luo, "A novel method for blood vessel detection from retinal images," BioMedical Engineering OnLine, vol. 9, pp. 14, [24] C. A. Lupascu, D. Tegolo, and E. Trucco, "FABC: Retinal Vessel Segmentation Using AdaBoost," Information Technology in Biomedicine, IEEE Transactions on, vol. 14, pp , [25] D. Marin, A. Aquino, M. E. Gegundez-Arias, and J. M. Bravo, "A New Supervised Method for Blood Vessel Segmentation in Retinal Images by Using Gray-Level and Moment Invariants-Based Features," Medical Imaging, IEEE Transactions on, vol. 30, pp , [26] X. You, Q. Peng, Y. Yuan, Y.-m. Cheung, and J. Lei, "Segmentation of retinal blood vessels using the radial projection and semi-supervised approach," Pattern Recognition, vol. In Press, Corrected Proof, [27] J. R. Movellan, "Tutorial on Gabor Filters," Tutorial paper [28] R. Zwiggelaar, S. M. Astley, C. R. M. Boggis, and C. J. Taylor, "Linear structures in mammographic images: detection and classification," Medical Imaging, IEEE Transactions on, vol. 23, pp , [29] J. Serra, Image Analysis and Mathematical Morphology. Orlando, FL, USA: Academic Press, Inc, [30] S. Theodoridis and K. Koutroumbas, Pattern Recognition, 3rd Edition ed: Academic Press, [31] R. Kohavi and F. Provost, "Glossary of Terms," Machine Learning, vol. 30, pp , [32] M. E. Martinez-Perez, A. D. Hughes, A. V. Stanton, S. A. Thom, A. A. Bharath, and K. H. Parker, "Retinal Blood Vessel Segmentation by Means of Scale-Space Analysis and Region Growing," presented at Proceedings of the Second International Conference on Medical Image Computing and Computer-Assisted Intervention, London, UK, [33] B. S. Y. Lam, G. Yongsheng, and A. W. C. Liew, "General Retinal Vessel Segmentation Using Regularization-Based Multiconcavity Modeling," Medical Imaging, IEEE Transactions on, vol. 29, pp , /11/$ IEEE 415