Computers in Biology and Medicine 43 (2013) 541548

Contents lists available at SciVerse ScienceDirect

Computers in Biology and Medicine

journal homepage: www.elsevier.com/locate/cbm

A novel method for retinal vessel tracking using particle lters

B. Nayebifar a, H. Abrishami Moghaddam a,b,n
a b

Department of Electrical Engineering, K.N. Toosi University of Technology, Seyed Khandan, P.O. Box 16315-1355, Tehran, Iran re brale, 80036 Amiens Cedex, France Inserm U1105 GRAMFC, Groupe de Recherches sur lAnalyse Multimodale de la Fonction Ce

a r t i c l e i n f o
Article history: Received 5 April 2012 Accepted 23 January 2013 Keywords: Retinal image processing Tracking Blood vessels Particle ltering Bifurcations

abstract
Extraction of a proper map from the vessel paths in the retinal images is a prerequisite for many applications such as identication. In this paper, we present a new approach based on particle ltering to determine and locally track the vessel paths in retina. Particle lter needs to use an acceptable probability density function (PDF) describing the blood vessels which must be provided by the retinal image. For this purpose, the product of the green and blue channels of the RGB retinal images is considered and after a median ltering stage, it is used as a PDF for tracking procedure. Then a stage of optic disc localization is performed to localize the starting points around the optic disc. With a proper set of starting points, the iterative tracking procedure initiates. First, a uniform propagation of the particles on an annular ring around each point (including starting points or ones determined as central points in the previous iteration) is performed. The particle weights are evaluated and accordingly, each particle is decided to be inside or outside the vessel. The subsequent stage is to analyze the hypothetical vectors between a central point and each of the inside vessel particles to nd ones located inside vessel. Afterwards, the particles are clustered using quality threshold clustering method. Finally, each cluster introduces a central point for pursuing the tracking procedure in the next iteration. The tracking proceeds towards a bifurcation or the end of the vessels. We introduced two criteria: automatic/manually tracked ratio (AMTR) and false/manually tracked ratio (FMTR) for evaluating the tracking results. Apart from the labeling accuracy, the average values of AMTR and FMTR were 0.7746 and 0.2091, respectively. The proposed method successfully deals with the bifurcations with robustness against noise and tracks the thin vessels. & 2013 Published by Elsevier Ltd.

1. Introduction Retina is one of the most important organs of the eye and the retinal image in a healthy person includes the following parts:

   

Blood vessels map. Optic disc. Macula. Background.

Retinal scanners provide images with low intensity in blood vessel areas and bright in other regions. Fig. 1 shows a retinal image from a healthy subject. Vessel path tracking in the retinal images is of considerable importance either in medicine or in biomedical engineering [14]
n Corresponding author at: Department of Electrical Engineering, K.N. Toosi University of Technology, Seyed Khandan, P.O. Box 16315-1355, Tehran, Iran. Tel.: 98 21 84062229. E-mail addresses: Bahador.Nayebifar@ieee.org (B. Nayebifar), moghadam@eetd.kntu.ac.ir (H. Abrishami Moghaddam).

for the following reasons: rst, retinal vessel tracking provides a vascular map with which the objects such as lesions, anatomical structures and bifurcations are easily associated [2]. Second, it can be used to localize a pathological pattern called venous beading which is crucial for progress evaluation of diabetic retinopathy. Finally, once the overall map of the retinal vascular network is available, the recognition and identication algorithms are assisted. Most of the vessel tracking methods utilize the vessel prole as a pattern to nd the vessel path in an iterative procedure [5,6]. In these methods, the algorithm starts with some predened starting points and evaluates the vessel similarity with a reference prole. In [7,8], the vessel tracking starts from starting points and follows the cues concluded from the continuities of position, diameter, direction, density and curvature. These methods evaluate the image pixel by pixel giving rise to a computationally intensive procedure. Moreover, presenting a unique model to deal with all types of vessels is impractical. In [8], by evaluating the vessel along a line perpendicular to its direction, an improved method of tracking is presented which uses a fuzzy model for the vessels. Such methods depend substantially on the vessel diameter and the gray level distribution through the vessels. Since

0010-4825/$ - see front matter & 2013 Published by Elsevier Ltd. http://dx.doi.org/10.1016/j.compbiomed.2013.01.016

542

B. Nayebifar, H. Abrishami Moghaddam / Computers in Biology and Medicine 43 (2013) 541548

Fig. 1. (a) A retinal image. (b) The corresponding image produced by the product of green and blue channels of the RGB image.

these two factors alter considerably through the vessels either in width or length, this type of tracking methods fails to present a total map for the retinal blood vessels. Some of the tracking methods have applied the Kalman lter to the system with this presumption that the vessel tracking is a linear Gaussian state space problem [9,10]. Such assumption does not hold true for all the cases as it does not consider all vessel characteristics (e.g. sudden direction variations in bifurcations). In this paper, we present a new method using particle lters which deals with the non-linear processes and with approximation through population-based sampling. This algorithm does not perform any window-based processing and utilizes the product of the green and blue channels of the RGB retinal images as a scalar probability density function (PDF) as illustrated in Fig. 1b. Considering the retinal vessels tracking as a recursive estimation problem, particle ltering can be applied to effectively address this issue. However, some basic challenges increase the problem complexities. Some of these challenges are listed as follows:

Fig. 2. The challenges in retina vessel tracking. Dashed arrows show the pathologies, boundary of the retina and optic disc. The solid lines shows the low contrast vessels that are very hard to detect.

2. Particle lters Particle lter or sequential Monte Carlo is a method based on the individual assessment of the particles and their characteristics and weighting them according to their local features. Such reasoning results in some particle to strengthen in special areas and weaken in other parts. Particle lter as a recursive Bayesian method can be further explained through system and measurement models as follows [4,12]: xn f n xn1 , vn1 yn hn xn , dn 1 2

 Optic disc has a perturbing effect on the quite uniform gray

level distribution in the retinal images.

 Pathological impacts that may appear as a series of bright

spots with darker gap in between [11].

 The dark strip along the center of a large vessel that may cause
its misinterpretation as two adjacent vessels.

 Many vessel branches start with an abrupt fall in the gray level 
values and incorrectly are assumed to be vessel edge and not a bifurcation. The vessels width is one of the main concerns especially for particle ltering approach to retinal vessel tracking. In other words, the particles must be propagated on an annular ring around each selected point with radius proportional to the vessel width. As the particles are sensitive to any change in the gray levels, the low contrast and many other imaging defects end in unexpected errors.

where n is the discrete time index, xn represents the state vector and yn is the measurement vector which can be a nonlinear function. Accordingly, vn1 and dn are the system and measurement noise, respectively. In each sequence, the PDF is represented by a set of N particles i i fx n , i 1, . . ., N g with their associated weights set fwn , i 1, . . ., N g which are evaluated to achieve the expected result. Therefore, the posterior density is estimated with an acceptable approximation as pxn 9yn 
N X i1 i where w n represents the weight of ith particle. In the nth iteration, the weights of particles must be updated recursively as i i w n wn1 i i i pyn 9x n pxn 9xn1 i i qx n 9x0:n1 , yn i i w n dxn xn

Fig. 2 gives a general overview of the aforementioned problems. This paper is organized as follows. Section 2 provides a short introduction to particle lters and its concept before invoking in detail the proposed algorithm in Section 3. The experimental results and analyses are given in Section 4 and conclusions are drawn in the last section.

B. Nayebifar, H. Abrishami Moghaddam / Computers in Biology and Medicine 43 (2013) 541548

i i i i i where pyn 9x n , pxn 9xn1 and qxn 9x0:n1 , yn represent the observation, state transition and the proposal density function, respectively. In our proposed method for retinal vessel tracking, the proposal density function is assumed to be a uniform distribution on an annular ring around the present point in order to obtain an efcient propagation of particles. Moreover, the state transition function (prediction) is discarded due to possible abrupt changes in vessel characteristics. Therefore, the particles are generated over the proposal density function and the particles weights are updated taking only the observations into consideration i i w n wn1 i pyn 9x n i i qx n 9x0:n1 , yn

543

and the subsequent position of the vessel is estimated through the following equation: ! N X i i ^ n mean x w 6 n dxn xn
i1

The vessel tracking differs from visual moving object tracking in the sense that the vessel is immobile and preserves its structure during the tracking process. Therefore, we must consider a proper propagation of the particles and make a correct conclusion about the posterior state of the vessels. The weights of particles assign the conjectural position of the path either in a moving object tracking system or an automatic vessel tracking system. Thus, a proper PDF must be provided to particle lter to represent a usable observation for particle ltering method.

3. The retinal vessel tracking algorithm Fig. 3 shows the ow diagram of our retinal vessel tracking algorithm. In this method, the tracking begins with some initial starting points inside a vessel and continues until the end of the vessel or a bifurcation. In bifurcations, proportional to the number of branches, the new paths are dened by new starting points. Therefore, the algorithm consists of two stages. In the rst stage, the optic disc is localized and its local effect on biasing gray level distribution is reduced. Then, using the localized optic disc, some starting points are dened for the vessel tracking algorithm. In the second stage, the tracking is performed to extract the vessel path using the particle ltering method [1315]. 3.1. Determining the starting points In this step, the optic disc must be localized. We adopted the method proposed in [3,16] for optical disc localization. First, a morphological closing is performed to suppress all the blood

vessels from the original image (Fig. 4). Then, a window larger than the size of optic disc is slid over the closed image and the window with the highest gray level uctuations is localized. It is rather probable that on healthy retinas, this window contains the optic disc. Afterward, regarding the fact that the optic disc has the sharpest edge in the closed image (Fig. 4), its edge map is extracted using the Canny edge detector. Finally, the center of gravity of the edge points is considered (with an acceptable accuracy) as the center of optic disc. Regarding to the fact that the vessels are emanated from the optic disc, the best area to locate the starting points is the beginning of the vessels around the optic disc. For this purpose, rst the original image is subtracted from the closed image as shown in Fig. 4b. The result is an enhanced image in which the effect of optic disc is suppressed and the vessels are bright and better contrasted with respect to the background (Fig. 5). This image is considered as PDF to weigh the particles. Then, an annular distribution of particles (with a radius around 10% of the retina diameter and thickness of a few pixels to contain the optic disc) is propagated to evaluate the local circumstances of the particles (approximately 700 particles) (Fig. 6). After the propagation stage, the particles are weighed according to the PDF and a local thresholding stage is performed using Otsus method (with a sectorial window containing 140 particles centered by the particle for which the threshold is determined). Particles which have a weight (gray level) higher than the local threshold are adopted and clustered to nd the particle groups for each vessel (Fig. 6). The particle clusters give the best candidates to start the tracking using (6). The consequent points are stored and each introduces a separate path to start the tracking. As this method needs a couple of points to continue the tracking, after locating the rst point, the best second point must be obtained. The following section explains how the second point is determined.

3.2. Tracking method using particle lter In the retina images, the angle between a vessel and its branches may be larger than 901 (Fig. 7b). Moreover, certain vessels show abrupt changes in their trajectories in the retinal images. Handling these special characteristics of vessels necessitates an appropriate propagation of the particles. Accordingly, the uniform propagation of particles on an annular ring (as the proposal distribution) around a (central) point will be of help for a better perception of the vessel conditions (Fig. 8a). However, the radius of the ring on which the particles are propagated must be appropriately estimated. On the one hand, this radius must be larger than the vessel width to guarantee the propagation of the particles over both the vessel and non-vessel areas. On the other hand, a too large radius may cause all

Fig. 3. The ow diagram of the retinal vessel tracking algorithm.

544

B. Nayebifar, H. Abrishami Moghaddam / Computers in Biology and Medicine 43 (2013) 541548

Fig. 4. (a) One of the original retinal images. (b) The closed image. (c) The window containing the optic disc. (d) The optic disc edges to determine the exact position of the optic disc center.

Fig. 5. The subtraction of the original image and the closed image.

central point (Fig. 8a). In this paper, the predened value was set to 15 pixels. 2. The hypothetical vectors between the central points and the particles and the brightness prole of the pixels constituting each vector are considered. For each vector, the standard deviation of the pixel gray values is calculated and stored. 3. Two standard deviation values are experimentally determined and considered as high (Th) and low (Tl) thresholds for hysteresis thresholding. The vectors with smaller standard deviation than Tl are considered as inside vessel and the ones with larger standard deviation than Th are considered as crossing the vessel. 4. The outside particles are evaluated to dene the proper radius value. For this purpose, the vectors crossing the vessel are analyzed. Along each vector, a point corresponding to the maximum change in brightness is considered to be on the vessel wall. The distances between the central point and the wall point along each vector are stored. Afterwards, the maximum value in the stored distances is opted as the radius for particles propagation. Fig. 8 shows the particle propagation and the vector drawn between a central point and one particle. After propagation of the particles on the annular ring around the central point, the particles weights (gray values) are compared to the local thresholds and ones with higher weights are selected. The next stage is to analyze hypothetical vectors between the central points and the particles and the brightness prole of the pixels constituting each vector (Fig. 8). For each vector, the standard deviation of the

the hypothetical vectors (between a central point and the particles) cross the vessel wall and their corresponding particle is considered as outside vessel. To estimate the radius of propagation, the following process is performed:

1. A high value is selected as the predened radius for the propagation of the particles on the annular ring around the

B. Nayebifar, H. Abrishami Moghaddam / Computers in Biology and Medicine 43 (2013) 541548

545

0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 0 100 200 300 400 500 600 700 800 900 1000

Fig. 6. (a) The particles which have higher weights (gray levels) than the local threshold are illustrated. They are clustered and their groups indicate the rst starting point. (b) The dashed lines show the local thresholding using Otsus method and the solid line depicts the particle gray levels.

Fig. 7. The angle between branches: y1 is smaller than 901 and y2 is greater than 901.

Fig. 8. (a) The particle propagation and the vector drawn between the central point and one particle. (b) The gray level uctuations along the mentioned vector. (c) Some possible states that may happen along the vector.

pixel gray values is calculated and stored. The vectors with standard deviation smaller than Tl are considered as inside vessel and the ones with standard deviation larger than Th are considered as crossing the vessel. For the vectors with standard deviation between Tl and Th, the gray values along each vector are

analyzed to determine which particles are inside and outside vessel. Each vector is segmented to nd the areas with high and low gray values along it (Fig. 8). Since all central points are assumed to be inside the vessels, the hypothetical vectors are expected to have high gray values at the beginning of their

546

B. Nayebifar, H. Abrishami Moghaddam / Computers in Biology and Medicine 43 (2013) 541548

brightness prole. Therefore, if the gray values around a central point are lower than the local threshold, it will be discarded. Otherwise, the following four situations are possible (Fig. 8c):

4. Experimental results To evaluate the algorithm, the retinal images from DRIVE and STARE databases are processed. As shown in Fig. 9, this algorithm was successful in tracking the blood vessels. Some of the tracked points do not correctly represent the vessel paths. In this paper, for a better evaluation of the proposed algorithm, we use the manually segmented retinal images from DRIVE and STARE databases. Moreover, we introduce two criteria: automatic/manually tracked ratio (AMTR) which is the ratio of the number of tracked pixels to the number of the pixels in the skeletonized image of the manually segmented images and false/manually tracked ratio (FMTR) which is the ratio of the number of the false tracked pixels to the number of the pixels in the skeletonized image of the manually segmented images. The tracked pixels are the ones which have a corresponding vessel point on their related manually segmented images: AMTR tracked pixels total pixels f alse tracked pixels total pixels 7

 There is a valley in the middle of the intensity prole. This

situation corresponds to two neighboring vessels.

 The intensity prole begins with high gray value around the  
central point and falls along the vector. This situation corresponds to a vector crossing the vessel wall. There is a small rise at the beginning and then a fall of the gray level along the vector. This situation is also considered as vessel crossing. There is no considerable fall of the gray values along the vector. Here, the vector is considered as totally inside vessel.

After dening the inside vessel particles, they are grouped and clustered. The clustering stage is performed using the quality threshold clustering (QT-clustering) method [17]. This method does not require any specication of the number of clusters. Therefore, it can dene any number of clusters (each of which present a new vessel path). Moreover, the quality of thresholding is an angular value which species the angular distance separating two clusters and varies in every stage according to the vessel width and particle ring radius. In the present work, a xed angular distance of 601 is used. If more than two clusters are formed, the development of a bifurcation is suspected. After dening the new points, the hypothetical vectors between the last central point and each of the new points are stored as a set of new tracked points (as indicated in Fig. 8). Each set of new points is a piecewise representation of the vessel path. Since the length of the vectors is sufciently small (from 5 to 15 pixels), they provide an acceptable approximation of the vessel trajectory. In some vessel end areas, the vessel fades locally and there is no distinct boundary for the vessel. One of the characteristics with which vessel end is detected is the angular difference between the particles in one cluster. If they have an angular difference more than a predened threshold, the vessel end will be reported.

FMTR

The average AMTR value in all images of DRIVE database for the proposed algorithm (apart from the labeling accuracy) was 0.7746 and the average FMTR value in all images was 0.2091. Some of the tracked images are depicted in Fig. 10. Moreover, the tracking algorithm was tested on STARE database (with manual starting points) and the resulted values for AMTR and FMTR were 0.7278 and 0.1262, respectively.

5. Conclusion In this algorithm, for the rst time, the particle lters are adopted to track the retinal vessels in the retinal images and extract the vessel tracking map. A method of propagation of the particles on the annular ring around each selected point is introduced. Such propagation presents our proposal distribution for particle ltering algorithm to track the vessel paths in retinal images.

Fig. 9. (a) An overview of the tracking in which each vessel is labeled according to the tracked path. The labels are assigned periodically for the ease of interpretation. Moreover, the white arrows show two types of failures in the algorithm either by not following the vessel (upward arrow) or incorrect tracking (downward arrow). The black arrow shows how the algorithm successfully handles the bifurcation. (b) The magnied picture of the bifurcation which is tracked correctly.

B. Nayebifar, H. Abrishami Moghaddam / Computers in Biology and Medicine 43 (2013) 541548

547

Fig. 10. Four images showing how the vessels are tracked and labeled. Each image has a TMSR value: (a) and (b) an original retinal image and its labeled image with AMTR 0.8520. (c) and (d) An original retinal image and its labeled image with AMTR 0.8319.

The proposed algorithm was performed on the retinal images from DRIVE and STARE databases and robustly tracked the blood vessels towards the end of the vessels. This algorithm tracks the main vessels down to the narrow ones. Despite missing some of the narrow vessels (because of the noise and imaging problems), it must not be considered as a weakness for this algorithm. Moreover, in retinal vessel tracking, each vessel introduces one path to assess the tracking procedure and narrow vessels are equally taken into consideration as wider vessels. In contrast, in retinal vessel segmentation methods, wide vessels cover larger areas. This results in high credence for segmentation methods. Hence, our proposed method which uses AMTR as an evaluation criterion must not be compared with the segmentation methods which are evaluated by the vessel areas. Accordingly, in this method, low values for AMTR seem acceptable. Future works will be focused on improving the accuracy of the PDF adopted by particle lters, the conclusion on the particles, and adaptability of the algorithm to be more sensitive to the local conditions of the particles. The data obtained from the hypothetical vectors between the central points and the particles can be stored as a history for each pixel state and used for more processing and conclusion about the pixels. Furthermore, the proposed tracking idea can be employed for the segmentation of the vessels.

Conict of interest statement None declared. Acknowledgment The author would like to thank Mr. Amin Dehghani for his valuable help during the realization of this research.

References
[1] Q. Li, J. You, L. Zhang, D. Zhang, P. Bhattacharya, A new approach to automated retinal vessel segmentation using multiscale analysis, 18th International Conference on ICPR, Pattern Recognition (2006). [2] M. Lalondey, L. Gagnony, M.C. Boucherz, Non-recursive paired tracking for vessel extraction from retinal images, in: Proceedings of the Conference Vision Interface, 2000, pp. 6168. [3] Y. Jiang, A. Bainbridge-Smith, A.B. Morris, Blood vessel tracking in retinal images, in: Proceedings of Image and Vision Computing New Zealand, 2007, pp. 126131. [4] M.S. Arulampalam, S. Maskell, N. Gordon, T. Clapp, A tutorial on particle lters for online nonlinear/non-Gaussian Bayesian tracking, IEEE Trans. Signal Process. (2002). [5] W. Huang, Automatic Detection and Quantication of Blood Vessels in the Vicinity of the Optic Disc in Digital Retinal Images, Masters Thesis, University of Waikato, 2006.

548

B. Nayebifar, H. Abrishami Moghaddam / Computers in Biology and Medicine 43 (2013) 541548

[6] S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, M. Goldbaum, Detection of blood vessels in retinal images using two-dimensional matched lters, IEEE Trans. Med. Imag. (1989). [7] Y. Sun, Automated identication of vessel contours in coronary arteriograms by an adaptive tracking algorithm, IEEE Trans. Med. Imag. (1989). [8] S. Tolias, S. Panas, A fuzzy vessel tracking algorithm for retinal images based on fuzzy clustering, IEEE Trans. Med. Imag. (1998). [9] T. Behrens, K. Rohr, H. Stiehl, Segmentation of tubular structures in 3D images using a combination of the Hough transform and a Kalman lter, in: Proceedings of the DAGM: Symposium on Pattern Recognition, vol. 2191, 2001, pp. 406413. [10] D. Lesage, E.D. Angelini, I. Bloch, G. Funka-Lea, A review of 3D vessel lumen segmentation techniques: models, features and extraction schemes, Med. Image Anal. (2009). [11] M. Sofka, C.V. Stewart, Retinal vessel extraction using multiscale matched lters, Condence Edge Meas. (2005). [12] H. Shim, D. Kwon, I.D. Yun, S.U. Lee, Robust segmentation of cerebral arterial segments by a sequential Monte Carlo method: particle ltering, Comput. Methods Prog. Biomed. (2006).

[13] C. Florin, N. Paragios, J. Williams, Particle lters, a quasi-Monte Carlo solution for segmentation of coronaries, in: Proceedings of the MICCAI, 2005, pp. 246253. [14] D. Lesage, E.D. Angelini, I. Bloch, G. Funka-Lea, Medial-based Bayesian tracking for vascular segmentation: application to coronary arteries in 3D CT angiography, in: Fifth IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2008, pp. 268271. [15] K. Allen, C. Yau, J. Noble, A recursive, stochastic vessel segmentation framework that robustly handles bifurcations, in: Proceedings of the Medical Image Understanding Analytics, 2008. [16] M. Schaap, R. Manniesing, I. Smal, T. van Walsum, A. van der Lugt, W. Niessen, Bayesian tracking of tubular structures and its application to carotid arteries in CTA, Part II, LNCS 4792, Proceedings of the International Conference on Medical Image Computing and Computer Assisted Intervention, 2007, pp. 562570. [17] C. Sammut, G.I. Webb, Encyclopedia of Machine Learning, Springer, US, 2010.