Blood vessel segmentation in pathological retinal image

Transcription

1 2014 IEEE International Conference on Data Mining Workshop Blood vessel segmentation in pathological retinal image Zhe Han, Yilong Yin*, Xianjing Meng,Gongping Yang, and Xiaowei Yan School of Computer and Technology Shandong University Jinan, China Abstract Retinal vessel segmentation is a fundamental aspect of the automatic retinal image analysis. The attributes of retinal blood vessels, such as width, tortuosity and branching pattern, play an important role in clinical diagnose. However, the edges of optic disk, fovea and edges of pathological areas have negative effects on vessel segmentation and few people focus on this problem. In this paper, we proposed a supervised method for retinal blood vessel segmentation. We design features based on local area shape combined with multi-scale local statistical features based on gray level and morphology features to solve the problems. Then, a support vector classifier is used for classification. Our algorithm is analyzed on two publicly available databases, called DRIVE and STATE. The accuracy of our method on both testing set is better than the 2nd human observer. The performance in pathological retinal images is satisfactory. Keywords-retinal vessel segmentation; pathological retinal image; support vector classifier; medical image analysis; I. INTRODUCTION Many early symptoms of diseases can be reflected on the retinal fundus, such as diabetes, glaucoma, hypertension, arteriosclerosis and so on [1,2]. Attributes of retinal blood vessels, such as width, tortuosity and branching pattern, are important indicators of diseases [3]. Retinal blood vessels are the main structure and landmark in retinal fundus image. Structures in the funds image, especially the optic disk and fovea, can be detected based on the spatial relationship with blood vessel structure [4]. Besides, the characters of retinal blood vessel could be applied to human recognition [5]. Above all, retinal blood vessel segmentation becomes an indispensable part of automatic fundus image diagnosis. There are many challenges on retinal blood segmentation. 1. Thin blood vessels especially with the one pixel width are usually lost; 2. The edges of other structures in retinal image, such as optic disk, fovea and pathological region, are often difficult to distinguish with vessels; 3. Pigmented epitheliums in the background have bad influence on vessel extracting especially in pathological images. These three problems are the main reasons for false detections. The first challenge cause error detection on vessels and other two are the main factors for errors on background. Many algorithms *Corresponding author: ylyin@sdu.edu.cn [6] specialize in the first problem and get good results. There is few people focus on the last two challenges [7, 19, 25] and few supervised methods are designed to deal with them. Our work focuses on the latter two problems. Difference of gray level between blood vessels and pathological areas is an important basis for feature definition, but many pathological areas have similar gray-levels with blood vessels. Besides, the shape of blood vessel in a local area is similar to rectangle whereas the shape pathological areas are irregular. So, we design features based on the shape of vessels in a local area and combine them with multi-scale local statistical features based on gray level and morphology features to segment retinal vessels. Support vector classifier is applied to classification of all the pixels. Our method is insensitive to the edges of optic disk, fovea pathological areas and noise points. False detection rate of background is desirable in the final results. The remainder of this paper is organized as follows. Section II gives a brief review on previous work of retinal vessel segmentation methods. Details on the feature vector and a rough description about the support vector classifier (SVC) used in our paper are presented Section III. Experimental results and analysis are provided in section IV. Conclusion and discussion are in Section V. II. RELATED WORK Previous work about the retinal vessel segmentation can be roughly divided into two classes: supervised methods and rule-based methods. Rule-based approaches can be mainly classified as matched filtering [8], vessel tracing [9], morphological processing [10, 11] and model based [12]. Retinal vessel segmentation method based on matched filtering was first proposed by Chaudhuri et al. [8]. The highest response of a group of Gaussian filters in 24 orientations was treated as the final value of a pixel and a proper threshold is selected to get a binary vessel image. Then, a further post-processing was applied to obtain the final result. Vessel tracking technique was used in retinal angiogram [9] by estimating local vessel trajectories with a given vessel initial point. Morphological filters and curvature evaluation ware combined by Zana et al. [10] to detect bright, local linear and connected vessel-like /14 $ IEEE DOI /ICDMW

2 structure. Fraz et al. [11] presented method for the detection of blood vessel tree by a combination of techniques for vessel centerlines detection and morphological bit plane slicing. Proper descriptors of retinal blood vessels are helpful for detection. In [13], the region growing method and a regionbased active contour model with level set implementation are exploited to extract retinal blood vessels. Besides, due to the various widths and orientations of the retinal blood vessels, multi-scale technique was applied to vessel segmentation [12]. To classify each pixel into two classes: vessel or nonvessel, supervised methods need using the labeling information while unsupervised methods do not need any labeling knowledge. Ricci and Perfetti [14] used line operators and support vector classification for retinal blood vessel segmentation. Lupascu et al. [15] proposed a supervised method by using a 41-D feature vector and AdaBoost classifier (FABC). A multilayer feed forward neural network and a 7-D feature vector, which is composed of gray-level and moment invariant-based features, are combined by Marin et al. [16] for the detection of retinal blood vessels. Staal et al. [17] proposed a supervised retinal blood segmentation method based on the extraction of image ridges by using k-nn classifier. In [18], feature vectors are composed of the pixels intensity and continuous two-dimensional Morlet wavelet transform responses taken at multiple scales. A Bayesian classifier with class-conditional probability density functions (likelihoods) described as Gaussian mixture is used. A semisupervised approach which used the radial projection is presented in [6]. Kande et al. [19] proposed an unsupervised vessel segmentation approach by using spatially weighted fuzzy C-means clustering. An ensemble classification-based retinal blood vessel segmentation approach was presented by Fraz et al. [20]. III. THE METHODOLOGY The proposed method is presented in this section. Feature vector is extracted from the green channel due to its higher contrast than red and blue channel. The detail of features vector and an introduction of support vector vlassifier (SVC) are presented in the following part. A. Feature vector Feature vector is presented in this part. The proposed features can be categorized into three classes: the novel local shape based features, multi-scale local statistically features based on gray level and multi-scale morphology features, as is detailed subsequently. 1) Local shape based features: In a local area, the shape of vessels is similar to rectangle. We can recognize a pixel based on the shape of the local area it belongs to. To get the local area for a pixel, we firstly choose a local square area for a pixel. We get a local area through searching in many directions from the current pixel, so instead of obtaining a connected area we get a line set for pixel. In other words, we fill the local area with a line set and the attributes of the line set can represent the shape of local area. Then, features are extracted from the line set. We firstly choose a local square area with size (R+1) (2 R+1) for a target pixel. The pixel is at (R+1,R+1) of the local square. For the reason that we only want to separate vessel area with back ground and almost all of the widths of blood vessels are less than 20 pixels, we set the R as 20. Secondly, we get a line starting from the target pixel and searching neighborhoods along a direction. Only the pixels with intensity less than current pixel or with the difference of no more than S compared with current pixel are included in the current line. Once a pixel does not satisfy the conditions above, searching step stops. We get a line every fifteen degrees and totally get 24 lines and it is entirely enough to fill a local square area and we set the direction number of the lines from 1 to 24 clockwise and starting from the x-axis. Two categories of features are extracted though the line sets above. The first category, contains five features, is only based on the line set of a target pixel. The second category, contains three features, is based on all the line sets in a local square. For a target pixel, it is easily to get the line set and find the longest line A and the line B perpendicular to it. We set the direction of the longest line as main direction D. The length of all 24 lines is AL. The first group of features is as follows: The lengths of A and B: LA and LB; the values of (LA-LB); the value of LA 24-AL and the absolute value of AL-LB 24. On a line, the number of pixels, which have same main direction D with target pixel, is set as NLS. For the neighborhoods of a pixel, the sum of the direction numbers is set as SDN. The sum of all the LA and LB of the pixels in the neighborhoods for a target pixel are set as SLA and SLB. The number of all the pixels in neighborhood is NL and NL is equal to eight. The values of (NLSA/LA)+(NLSB/LB) and the absolute values of SDN-(DN NL), SLA-(LA NL) and SLB-(LB NL) are the second group of value. We select four kinds of points: on the central of vessel (P1), on the edges of vessel (P2), above the vessel (P3) and under the vessel (P4) in green channel image and get the corresponding line set in local area. The green channel image and line sets are shown in Fig.1. Searching based lines in all directions is illustrated. The first group features of the four points are also given in Table I. 2) Multi-scale local statistically features based on gray level: In the green channel image, blood vessels areas are always darker than background. That is to say, gray levels of pixels on vessels are smaller than gray levels of other pixels in a local area. We use this statistical gray level information of vessels to get features. We apply the formula (1) to get six features with different l and the six values of l are 1,2,5,10,15 961

3 as a dilation operation for image f with structural element di and define fe i = f di as an erosion operation for image f with structure di. We define fo i = f di as image f using di to do open operations. We define fc i = f di Figure 1. Original image is shown in (a). Searching steps are based on the lines in all directions in (b). (b1) (b2) (b3) and (b4) are the line sets one of P1, P2, P3 and P4. Table I EXAMPLES OF FEATURES P1, P2, P3 AND P4 ARE THE POINTS IN FIG.1. THE FEATURES OF LA, LB, (LA-LB),(LA 24-AL)/24 AND (AL-LB 24)/24 FOR EACH POINTS ARE GIVEN. and 20. Formula (1) is defined as: p(i, j) =nl/nl 255 (1) Where p(i,j) is the intensity at coordinate (i,j). In a local square area, with the central pixel (i,j), N means the total number of pixels in the area and n means the number of pixels with intensity no more than p(i,j). The edge length of the square is set as l. nl and Nl are the corresponding values of the local square with edge length l. It is obvious that p(i,j) is between 0 and 255. Besides, when the target pixel is near the edge of the image, we will get the smaller edge length between current l and the distance between the target pixel with the image edge to ensure that the local square do not exceed the the image. 3) Multi-scale morphology features: First of all, in order to simplify the description of features, we set the green channel of the original image as f. We set di as the disk structural element with radius of i. Each i is corresponding to a scale. We define fd i = f di as image f using di to do close operations. Three kinds of morphology features are used in our work and there are totally ten features. The first kind of feature is described below. This kind of feature contains one feature. Open operations with di are separately implemented to the green channel images. We set the mean value of the six scales as a feature: 1 f mor1 =1/6 fo i s.t i =1, 2, 3, 4, 5, 6 (2) 6. As we all know, open operation can remove bright details smaller than structural elements. In retinal image, small bright details are always noises points or pathological areas. At the same time, the smaller detail is the more sure it is noise point. We use the mean value because smaller bright details can be removed with a smaller value of i, so the mean value of all these scales can better represent the small details. The second kind of features is described below. This kind of feature contains three features. Firstly, we use (3) to get nine scales image with different size of di. The formulation (4) is aim to get the difference between green channel image and the green channel image after close operation and defined bellow: fclose i = f fc i s.t i =1, 2, 3, 4, 5, 6, 7, 8, 9 (3). Then, we get three features by (4). Formulation (4) is defined as follows: 3 f mor2k = fclose ( j +3 k 3) s.t k =1, 2, 3 (4) j=1. We get three features with k equal to 1, 2 and 3. We all know that close operation can remove dark details smaller than structural element. Original image subtracts the image after close operation of it can enhance dark details. In retinal image, dark details mainly contain vessels. However, vessels are in different sizes, so we use various sizes of structural structures to enhance vessels with different sizes. The third kind of features is described below. This kind of feature contains one feature. We add up the difference between dilation image and erosion image to the green 962

4 channel image in six scales. The mean value of the six scales is set as a feature. The formulation for this feature is as follows: 6 f mor3 =1/6 (f +(fd i fe i )) (5) i=1 s.ti =1, 2, 3, 4, 5, 6. Through this formulation the edges in the green channel image can be enhanced in the feature. B. Support vector classifier In linear binary classification problems, a support vector classifier is to find an optimal hyper plane, which can separate the two classes of samples with the best generalization ability [21]: W Tx + b =0 (6). w, x and b respectively denotes the weight value vector, feature vector and bias. The generalization ability refers to a classifier not only has good classification performance on the training set, but also has high prediction accuracy on the test set. To put it simply, the margin is the shortest geometrical distance between the hyper plane and the closest data points. Without loss of generality, we set the margin to be equal to 1[22]. For training set,wehave.when is the hyper plane and {x i,y i } n i =1 R m {±} W Tx + b =0 for y i =+1 W Tx + b =0 for y i = 1 W Tx + b =0 r =2/ W is the margin. Support vector classifier tries to maximize to find the maximum margin hyper plane. That equals to min 1/2 W 2 w,b (8) s.t y i (W T x i + b) 1 i =1,...,n. To solve inseparable problems which mean all the points cannot be separated linearly, kernel trick is an effective approach. The main idea of kernel trick is that it is easier to get the optimal separating hyper plane to transform nonlinear classification problems in low dimensional space, which is difficult to deal with, to the higher dimensional space. Kernel function essentially is a kind of mapping function aim to accomplish the transformation. The kernel function we used in our work is Radial Basis Function (RBF) [22], which is suitable for high-dimensional feature classification and has high computation speed. (7) Table II FOUR CONDITIONS OF VESSEL DETECTION Table III ALL THE TERMS FOR PERFORMANCE EVALUATION IV. EXPERIMENTAL RESULTS A. Materials Two publicly available databases are used to evaluate the performance of our method. The first one, named STARE (Structured Analysis of the Retina), was collected by Hoover et al. [23]. It totally has 20 raw retinal images for blood vessel segmentation. These images captured by a TopCon TRV-50 fundus camera at 35 FOV. All the images were digitized to pixels with 8 bit per color channel. There were two labeled image sets segmented by two observers and we use the first result as ground truth. The second database, named DRIVE, was established by Niemeijer et al. [24].The database contains a training set, and a testing set. Each of them has 20 images. These images are captured by a Canon CR5 3CCD camera with a 45 FOV. The images are of size pixels per color channel and have a FOV of approximately 540 pixels in diameter. There are a set A and set B respectively marked 12.7% and 12.3% pixels as vessel. Performance will be evaluated on the set A. B. Evaluation Methods In order to qualify the performance of our method, we compare our results with the gold standard, which is segmented by medical experts. We use sensitivity (SEN), specificity (SPE) and accuracy (ACC) to evaluate our method. Four conditions of detection results are shown in Table II. All the evaluation terms are defined in Table III. Sensitivity and specificity are stand for the ratio of well-classified vessels and non-vessels. C. Parameter Setting Methods Each feature is normalized after extracted from each image to decrease the difference between different images. On DRIVE database, samples are randomly selected from each training image, which means training samples are used to get the classification model. There are 20 images in STARE database and no training images are provided. We use leave-one-out strategy samples are randomly selected from a training image that is

5 Figure 2. The trend of accuracy with different R on STARE (with circles on each point) and DRIVE (with asterisk on each point) Figure 3. Best and worst accuracy images on DRIVE and STARE samples are used to get a classification model. In our method, there is no overlap of training set and testing set and all the training samples are totally randomly selected, which guarantees generality and applicability of the final classifier. The parameter S used in extracting local area features is set as 5. Follow the sample selection strategy above, we change the number of samples for each training image to 1000 to get the variation trend of accuracy with the different S. We get best accuracy when S equals to 5 on both databases as it is shown in Fig.2. D. Performance of Proposed method The mean accuracy, sensitivity and specificity on DRIVE database are , and ; on DRIVE database are , and , respectively. The best and worst accuracy, sensitivity and specificity on DRIVE database are (image19), (image19), (image4) and (image06), (image8) (image14), respectively. The best and worst accuracy, sensitivity and specificity are (image18), (image12), (image4) and (image16), (image4), (image5) on STARE database. Even the worst specificity of all the images is better than 2nd human observer and some existing methods. The best and worst accuracy cases on both databases are shown in Fig.3. E. Comparison with Existing Methods Our method is compared to other methods on the two public databases in Table IV and Table V. The evaluation terms of sensitivity, specificity and accuracy are presented. Our accuracy is higher than most current methods. Specificity is far better than existing methods, which proves our good 964

6 Figure 4. Performance on the edges of optic disk, fovea and pathological areas Table IV PERFORMANCE RESULTS COMPARE TO OTHER METHODS ON THE DRIVE DATABASE. Table V PERFORMANCE RESULTS COMPARE TO OTHER METHODS ON THE STARE DATABASE performance on edges of other structures. The sensitivity is accessible and the main structures of vessels are reserved. F. Performance Analysis on Background Noises In our work, we pay more attention on the edges of other structures and pigmented epitheliums in the background, so that our method has good performance on background. As is shown in Fig.3 and Fig.4, our method is insensitive to edges of optic disk, fovea and pathological areas and pigmented epitheliums. Almost no edge of optic disk and fovea are detected. The pixels misclassified as vessels have very similar shape with vessels or connected with blood vessel. V. CONCLUSION AND DISCUSSION In this paper, a supervised retinal blood vessel segmentation method is proposed. Nineteen features are extracted for each pixel and all the pixels are classified by SVC. Feature vector are designed and extracted based on local gray level statistically features, multi-scale morphology features and the new the local area shape features. Our method has been tested on two public datasets. A few samples of false detection occur in abnormal images in the areas. The edges of OD, fovea and big pathological areas and pigmented epitheliums in background have little influence on final results. With the high specificity and accuracy, the final results have good visual appearance, which is credible for 965

7 the recognizing of underlying shape or structures. As same as many retinal blood vessel detection approaches, absence and discontinuity of thin vessels causes the main proportion of false positive rate, especially in the low contrast images. Supervised methods need long time for training time as we all known and no exception of our algorithm. The mean Training time and classification time or each image are respectively less than half hours and ten minutes. Future work will be aim to get better performance for pathological image of the method.. Our work can be combined with other methods that are good at the segmentation of thin vessels to solve the limitation of false detection on thin vessels. To reduce the time of training stage, new features based on local shape could be designed and more efficient classifiers could be explored. ACKNOWLEDGMENT The work is supported by NSFC Joint Fund with Guangdong under Key Project U , Shandong Natural Science Funds for Distinguished Young Scholar under Grant No. JQ and the Research Found for the Doctoral Program of Higher Education under Grant No REFERENCES [1] Kanski, J. J., Bowling, B., Synopsis of Clinical Ophthalmology, Elsevier Health Sciences, [2] Antal, B., Hajdu, A., An ensemble-based system for microaneurysm detection and diabetic retinopathy grading, IEEE Trans. on Biomed. Eng., vol.59, no.6, , [3] Fong, D. S., Aiello, L., Gardner, T. W., King, G. L., Blankenship, G., Cavallerano, J. D., Klein, R., Retinopathy in diabetes. Diabetes care, vol.27, no. suppl 1, 84-87, [4] Foracchia, M., Grisan, E., Ruggeri, A., Detection of optic disc in retinal images by means of a geometrical model of vessel structure, EEE Trans. Med. Imag., vol.23, no.10, , [5] Akram, M. U., Tariq, A., Khan, S. A., Retinal recognition: Personal identification using blood vessels, in ICITST 2011, pp , [6] You, X., Peng, Q., Yuan, Y., Cheung, Y. M., Lei, J., Segmentation of retinal blood vessels using the radial projection and semi-supervised approach, Pattern Recogniti., vol.44, no.10, , [7] B. S. Y. Lam, Y. Gao and A. W. C. Liew, General retinal vessel segmentation using regularization-based multiconcavity modeling, IEEE Trans. Med. Imag., vol. 29, no. 7, pp. 1369C1381, Jul [8] S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and M. Goldbaum, Detection of blood vessels in retinal images using two-dimensional matched filters, IEEE Trans. Med. Imag., vol. 8, no. 3, pp. 263C269, Sep [9] Liu I., Sun Y., Recursive tracking of vascular networks in angiograms based on the detection-deletion scheme, IEEE Trans. Med. Imag., vol.12, no.2, pp , [10] F. Zana, and J.C. Klein, Segmentation of vessel-like patterns using mathematical morphology and curvature evaluation, IEEE Trans. Image Proc., vol. 10, no. 7, pp , Jul [11] A. Mendonca, and A. Campilho, Segmentation of retinal blood vessels by combining the detection of centerlines and morphological reconstruction, IEEE Trans. Med. Imag., vol. 25, no. 9, pp. 1200C1213, Sept [12] Delibasis, K. K., Kechriniotis, A. I., Tsonos, C., Assimakis., Automatic model-based tracing algorithm for vessel segmentation and diameter estimation, Comput. Meth. Prog. Bio., vol.100, no.2, pp , [13] Y. Q. Zhao, X. H. Wang, X. F. Wang, and F. Y. Shih, (2014, Jan.). Retinal Vessels Segmentation Based on Level Set and Region Growing, Pattern Recognit. [Online], Available: [14] E. Ricci and R. Perfetti, Retinal blood vessel segmentation using line operators and support vector classification, IEEE Trans. Med. Imag., vol. 26, no. 10, pp. 1357C1365, Oct [15] C. A. Lupascu, D. Tegolo, and E. Trucco, FABC: Retinal vessel segmentation using Adaboost, IEEE Trans. Inf. Technol. Biomed., vol. 14, no. 5, pp. 1267C1274, Sep [16] 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 invariantsbased features, IEEE Trans. Med. Imag., vol. 30, no. 1, pp , Jan [17] J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever, and B. van Ginneken, Ridge-based vessel segmentation in color images of the retina, IEEE Trans. Med. Imag., vol. 23, no. 4, pp , Apr [18] 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, IEEE Trans. Med. Imag.,vol. 25, no. 9, pp. 1214C1222, Sept [19] Kande, G. B., Subbaiah, P. V., Savithri, T. S., Unsupervised fuzzy based vessel segmentation in pathological digital fundus images, J. Med. Syst., vol. 34, no.5, pp , [20] M. M. Fraz, P. Remagnino, A. Hoppe, B. Uyyanonvara, A. R. Rudnicka, C. G. Owen, and S. A. Barman, An ensemble classification-based approach applied to retinal blood vessel segmentation, IEEE Trans. Biomed. Eng., vol. 59, no. 9, pp , Sept [21] Wu, X., Kumar, V., The top ten algorithms in data mining, CRC Press, [22] Cristianini, N., Shawe-Taylor, J., An introduction to support vector machines and other kernel-based learning methods, Cambridge university press,

8 [23] A, Hoover, V. Kouznetsova and M. Goldbaum, Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response, IEEE Trans. Med. Imag., vol. 19, no. 3, pp , Mar [24] Alonso Montes, C., Vilarino, D. L., Dudek, P., Penedo, M. G., Fast retinal vessel tree extraction: A pixel parallel approach, Int. J. Circ. Theor. App., vol.36, no.6, pp , [25] B. S. Y. Lam and H. Yan, A novel vessel segmentation algorithm for pathological retina images based on the divergence of vector fields,ieee Trans. Med. Imag., vol. 27, no. 2, pp. 237C246, Feb. (2008). 967