- Email: [email protected]
Retinal vessel segmentation employing ANN technique by Gabor and moment invariants-based features
Applied Soft Computing 22 (2014) 94–100
Contents lists available at ScienceDirect
Applied Soft Computing journal homepage: www.elsevier.com/locate/asoc
Retinal vessel segmentation employing ANN technique by Gabor and moment invariants-based features S. Wilfred Franklin a , S. Edward Rajan b,∗ a b
CSI Institute of Technology, Thovalai, Tamil Nadu, India Mepco Schlenk Engineering College, Sivakasi, Tamil Nadu, India
a r t i c l e
i n f o
Article history: Received 16 March 2012 Received in revised form 4 July 2012 Accepted 21 April 2014 Available online 21 May 2014 Keywords: Diabetic retinopathy Retinal vessel segmentation Artificial neural networks Retinal images Retinal vasculature
a b s t r a c t Diabetic retinopathy (DR) is the major ophthalmic pathological cause for loss of eye sight due to changes in blood vessel structure. The retinal blood vessel morphology helps to identify the successive stages of a number of sight threatening diseases and thereby paves a way to classify its severity. This paper presents an automated retinal vessel segmentation technique using neural network, which can be used in computer analysis of retinal images, e.g., in automated screening for diabetic retinopathy. Furthermore, the algorithm proposed in this paper can be used for the analysis of vascular structures of the human retina. Changes in retinal vasculature are one of the main symptoms of diseases like hypertension and diabetes mellitus. Since the size of typical retinal vessel is only a few pixels wide, it is critical to obtain precise measurements of vascular width using automated retinal image analysis. This method segments each image pixel as vessel or nonvessel, which in turn, used for automatic recognition of the vasculature in retinal images. Retinal blood vessels are identified by means of a multilayer perceptron neural network, for which the inputs are derived from the Gabor and moment invariants-based features. Back propagation algorithm, which provides an efficient technique to change the weights in a feed forward network is utilized in our method. The performance of our technique is evaluated and tested on publicly available DRIVE database and we have obtained illustrative vessel segmentation results for those images. © 2014 Published by Elsevier B.V.
Introduction Diabetic retinopathy (DR) is a result of long-term diabetes and it is a severe and the most common sight threatening eye disease, which causes blindness among working-age people around the world [1]. Major vision loss due to DR is highly preventable with proper screening and timely diagnosis at the earlier stages. However, DR is not painful and the diabetic patients perceive no symptom until visual loss occurs and hence they need periodical eye-fundus examination to ensure that treatment is received in time. Evaluation of the characteristics of retinal blood vessels plays an important role in the diagnosis of diseases based on vascular pathology. The various features of retinal vessels such as length, width, tortuosity and branching pattern provide new techniques to diagnose various diseases like diabetes, arteriosclerosis, hypertension, cardiovascular disease and stroke. Retinal images provide valuable information related to human eye, by which the vascular
∗ Corresponding author. Tel.: +91 04562 235308. E-mail addresses: [email protected] (S.W. Franklin), [email protected] (S.E. Rajan). http://dx.doi.org/10.1016/j.asoc.2014.04.024 1568-4946/© 2014 Published by Elsevier B.V.
condition can be accurately observed and analyzed [2]. The only part of the central circulation that can be viewed directly and analyzed is the retinal vessel. Changes in blood vessel structure and vessel distribution, caused by diabetic retinopathy can lead to new vessel growth, which in turn instigates vision impairment. Hence retinal vessel segmentation becomes an essential tool for the detection of any variations that occurs in blood vessels. Retinal blood vessel segmentation gives the detailed information about the location of vessels which helps in the screening of diabetic retinopathy, e.g., in the detection of micro-aneurysms it helps to reduce the number of false positives [3,4]. Observations based on retinal vessel segmentation are quiet complex and hence the traditional way of diagnosis in which the ophthalmologist identifies the anomalies present in the retinal images by examining it, becomes tiresome or even impossible. As the actual size of a typical blood vessel in human retina is very small, just a few pixels wide, measurement of vascular width using retinal image segmentation process becomes critical and challenging. One of the possible solutions for these measurements is the use of computer based automated analysis of optic fundus, a technique widely accepted by the medical community. This computer based analysis helps the medical personnel to detect the changes in blood flow and vessel distribution. It also enables the
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
detection of extra vessel growth. For automatic analysis of optic fundus that detects the blood vessel irregularities, segmentation of the vessels from the background is made initially and thereby further processing like feature extraction can be achieved. In human retina, one of the most important organs is the optic nerve which acts as the convergent point of the blood vessel network. The central retinal artery and central retinal vein flow out through the optic nerve that supplies blood to the upper layers of the retina. Besides, the optic nerve acts as a channel to convey the information from the eye to the brain. In early stages, most of the retinal pathologies affects locally and does not distress the entire retina. But retinal pathology on or near the optic nerve may severely affect the vision even at the early stages because optic nerve is the most essential part for vision [5]. A few observations found in several important retinopathies are attenuation changes, focal narrowing and occlusion of retinal arteries. The diameter and shape of a retinal vessel plays a key role as indicators in ophthalmologic studies. These changes give valuable information to identify the successive stages of diseases and their response to various therapies. The optic nerve can be observed in a close view of the retinal fundus, where the optic disc is the portion of the nerve that is visible or perceivable by the eye. Fundus imaging is one of the popular clinical procedures available to record this close view observations of the retina. This fundus imaging procedure is also used for the diagnosis and evaluation of the healthy and non-healthy retinas of human eye [6]. In a healthy retina the optic nerve has a standard identifiable size, shape, color and location relative to the blood vessels. Nevertheless, in a retina containing lesions, any one or more of these properties may be deviated from its standard level and show a large variance. At various stages of the disease, the vascular network in retina is very much affected and hence various morphological changes occur in retinal vessels. We can substantially observe enough geometrical changes in diameter, branching angle, length or tortuosity in the retinal blood vessels due to diseases. The segmentation and measurement of retinal blood vessels can be used to grade the severity of certain diseases. The sign of risk level for diabetic retinopathy is the variation in width of retinal vessels within the fundus. One of the most important tools for the prediction of proliferate diabetic retinopathy is the abnormal variation in diameter along the vein. Moreover, the various retinal microvascular abnormalities predicted are seen to be the early symptoms for the risk of stroke. In all these cases, the desired focus is on the variation in diameter of the vessel and not in the exact diameter of the vessel. An alternative application of retinal vessel segmentation is biometric identification using distinctive retinal vascular network for each individual. This research paper is organized as follows: Section “Overview of retinal vessel segmentation” reviews various other vessel segmentation methods, Section “Methodology” describes and illustrates the methods of our proposal for vessel segmentation, Section “Experimental results analysis” presents its results and compares the performance with some of the techniques previously published and Section “Discussion and conclusion” puts forward the discussion and concludes this work.
Overview of retinal vessel segmentation Segmentation of vessels in retinal images is achieved by classifying each image pixel as vessel or nonvessel based on the local image features. In two-dimensional images, the techniques adopted for identifying vessels are based on specific properties of the vascular segments. There are two basic approaches that are generally considered for the identification of general vascular segments, which also play a key role in retinal vessel segmentation applications. The
95
algorithms used for the segmentation of blood vessels are broadly classified as pixel processing-based methods and vessel tracking or vectorial tracking methods [2]. Supplementary methods used for vessel segmentation in retinal images can be classified into two groups. The first group consists of rule-based methods and the second group consists of supervised methods, where training of manually labeled images is used. Neural network based pixel classification scheme falls under the second category [7,8], which is followed in this paper. Pixel processing-based methods normally consist of two phases. In the first phase, an enhancement procedure is implied and it selects an initial set of pixels, which is further ensured as vessels in the second phase [9]. The retinal vessel segmentation method presented in Ref. [2] consists of three processing phases. In the first phase, background normalization of monochromatic input image is performed and later thin vessel enhancement procedures are used. In the second phase, the vessel centerline candidate points are selected and subsequently these points are connected and based on vessel length, the validation of centerline segment candidates is achieved. In the third phase, vessels with different widths are enhanced and processed using binary morphological reconstruction technique and vessel filling is achieved using region growing process. Soares et al. have proposed an algorithm, where retinal blood vessels are detected using Gabor wavelets by representing each pixel by a feature vector and then by using Bayesian classifier with Gaussian mixtures, each pixel is classified as either a vessel or nonvessel pixel and thus segmentation is achieved. The concept of matched filter detection was proposed by Chaudhuri et al. [10], where twelve rotated versions of 2-D Gaussian shaped templates are used to search vessel segments along all possible directions. The resultant image produced is the binary representation of the retinal vasculature. Likewise, segmentation of retinal vessels are also obtained by matched filtering approaches using global [11] or local thresholding strategies [12]. Also for the purpose of vessel borders extraction, differential filters based on either first-or-second order derivatives are used. A two stage region growing procedure was proposed by Martinez-Perez et al. where features derived from image derivatives are used in the segmentation of retinal vessels [13,14]. As stated in Ref. [15], the edge detection, matched filtering and region growing procedures can also be collectively used for the automated detection of retinal blood vessels. Jiang et al. have proposed a technique of adaptive local thresholding based on the use of verification-based multithreshold scheme combined with classification procedures in Ref. [16], used for the detection of vessels in retinal images. A technique used for the segmentation of vessel like patterns in retinal images that combines morphological filters and cross-curvature evaluation was proposed by Zana et al. in Ref. [17]. In this approach, vessel segments alone are considered as image feature and hence using morphological filters simplifies the computation of cross-curvature. An algorithm for retinal vessel segmentation based on the classification of pixels using simple feature vector was proposed by Niemeijer et al. in Ref. [18]. A new method of segmentation of blood vessels in retinal images was proposed by Staal et al. in Ref. [19]. This supervised method is called primitive-based method and this algorithm is based on the extraction of image ridges used to compose primitives that describe the linear segments called line elements. The pixels those are assigned to the closest line element partitions the image in the form of image patches and are classified using a set of features from the corresponding line and image patch. In addition, a technique based on neural network is used to identify the retinal blood vessels, where the inputs are obtained using principal component analysis and then edge detection technique is used. Vessel tracking methods use the concept of measuring some local image properties to locate the vessel points which are used for tracing the vasculature. In these types of algorithms, both the
96
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
extraction of the image feature and the recognition of the vessel structure are performed simultaneously. An algorithm which relies on matched filtering approach was presented in Ref. [20], and is initiated by start and end points followed by automatic detection of vessel boundaries and then tracks the centerline of the vessel and finally it extracts the useful parameters of clinical interest. The unsupervised fuzzy vessel tracking algorithm presented in Ref. [21] is virtually parameter-free and it performs automatic initialization. It also overcomes the problem of vessel profile modeling, tracks fundus vessels automatically and the classification of vessel and nonvessel regions along a vessel profile is determined using fuzzy C-means clustering algorithm. The algorithm presented in Ref. [22] performs automatic adaption without manual initialization. It automatically detects the seed-point of tracing, described as local gray-level minima and it is applied directly at the image intensity level without preprocessing. Consequently, sequences of exploratory searches are initiated to detect the vessel boundaries and these models are used in the estimation of next vessel point. In Ref. [23] Chutatape et al. have proposed a technique where the tracing process starts from the circumference of the optic disc and the use of Kalman filter estimates the next point of location in the tracking procedure.
Image database We have tested and evaluated the technique used in this method by making use of the retinal color images available on DRIVE database [25]. This database has been widely used as ground truth for performance evaluation by other researchers to test their vessel segmentation methodologies. The retinal images available in this database were captured using a Canon CR5 nonmydriatic 3CCD camera with a 45◦ Field-of-View (FOV), in digital form. The DRIVE database consists of forty eye-fundus color images which are divided into training and test set, each set containing twenty images out of which the training set has three images with pathology and the test set has four images with pathology. The images available in the training set were segmented only once and the images in the test set were segmented twice, resulting in two sets, the first set is accepted as the ground truth and used for algorithm performance evaluation in literature. The complete results of the manual segmentation are also available for all the images of the two sets. The color images of the retina available in the database are of size 768 × 584 pixels and 8 bits per color channel, and were initially saved in JPEG format. Moreover, the database includes masks with the delimitation of a Field-of-View (FOV) of approximately 540 pixels in diameter for each image, which is used for the performance evaluation of our method.
Methodology This paper proposes a retinal vessel segmentation technique using Artificial Neural Networks (ANN). The main features of fundus retinal images are blood vessels, which are used to detect the presence of anomalies present in human retina automatically. These retinal blood vessels are identified by using a multilayer perceptron neural network, for which the inputs are derived from Gabor and moment invariants-based features. Using neural networks, some of the features such as hemorrhages and exudates can be detected, which helps to examine whether retinopathy is present or not. The images with retinopathy also require a grading measure, which is a complex task than identifying the presence of retinopathy. It can be examined by a trained person, because the computer based analysis detects only the changes such as neovascularisation, cotton wool spots, perifoveal exudation and vascular changes [24]. The basic step to achieve this diagnosis procedure is to locate the important regions of the fundus, such as blood vessels, automatically. The data from these segmented blood vessel regions can further be processed and analyzed for examining the features of certain eye diseases. In this paper, computer based image analysis using neural network is applied to detect and segment the blood vessels from retinal images.
Preprocessing Preprocessing of images is the first step carried out before presenting the input to the neural network. The retinal color images taken from the DRIVE database are preprocessed, where each pixel contained three values-red, green and blue. A few examples given as input images to the multilayer perceptron neural network can be seen in Fig. 1. The RGB features are extracted for the preprocessed image, its mean is calculated and the gray scale image is obtained for further processing. The true color image RGB is converted to the grayscale intensity image to evaluate its adequate ability for the segmentation of the retinal blood vessels. The conversion is done by eliminating the hue and saturation information, while retaining the luminance [26]. The given input images are further processed using mean filtering and a few examples are shown in Fig. 2. In the next step, the retinal images are further processed and smoothened using Gaussian filters and a few examples can be seen in Fig. 3. Then the Gabor features are extracted at different orientations and a few examples are shown in Fig. 4. Gabor filtering is done and
Fig. 1. Input images given to multilayer perceptron neural network.
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
97
Fig. 2. Preprocessed images obtained from original color fundus images after applying mean filtering.
Fig. 3. Preprocessed images after applying Gaussian filters.
Fig. 4. An example for Gabor features extracted at different orientations.
as per the Gabor coefficients attained, the original images are convolved. Next, the moment invariants-based features are extracted. Now the samples are taken from vessel and nonvessel regions and they are trained using neural networks. The feature vectors extracted from preprocessed retinal images are given as input to this neural network. This technique is based on pixel classification and these classification results are thresholded to classify each pixel as vessel or nonvessel. Neural network training Each pixel of a retinal image is classified as vessel or nonvessel, using neural networks. The computational structure of neural
networks and the human visual system has several similarities. In order to classify each pixel of the retinal image as vessel or not, a multi-layer perceptron neural network is used [27,28]. It is a multilayer, feed forward network using extend gradient-descent based delta-learning rule, commonly known as back propagation rule. Back propagation is a systematic method for training multi-layer artificial neural network. According to the basic definition of feed forward network, a neuron in the hidden layer will receive inputs from all neurons in the previous layer and must feed all neurons in the next layer. Back propagation provides an efficient technique to change the weights in a feed forward network with differentiable activation function units, and to learn a training set of input–output examples. If the number of layers in the feed forward network is
98
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
represented as n, and if Li represents the number of neurons in the ith layer, then the total number of weights is given by: N=
n−1
Li Li+1
where, i = 1, 2, . . ., n
(1)
i=1
The learning capacity (C) is an important parameter which determines the ability of a system to learn. For a three layer network, the learning capacity is bounded by: nh < C < nh log(h) log(nh2 )
(2)
where, h is the number of hidden nodes and n is the number of input nodes. The neural network applied is a three layer perceptron having one input node, five hidden layers and one output node. Before presenting the input to the neural network, preprocessing of the image must be done. The neural network is considered to be trained when the system error falls below an acceptable threshold level and at this stage, the vessels are identified from the background and thus segmented images are obtained.
ık = (tk − yk ) f (y−ink )
(5)
It is used to distribute the error at the output unit yk back to all units in the previous layer. Then each hidden unit (zj , j = 1,. . ., n) sums its delta inputs from units in the above layer as: ı−inj =
m
ıj wjk
(6)
k=1
Similarly the error information term ıj = ı−inj f(z−inj ) is computed for each hidden unit. During the final stage, the weights and biases are updated using the ı factor and the activation function is updated, using ˛ as the learning rate. Each output unit updates its weights and biases by using the weight correction term Wjk and bias correction term Wok , where, Wjk = ˛ ık zj and Wok = ˛ ık
(7)
Therefore, Wjk (new) = Wjk (old) + Wjk and
(8)
Wok (new) = Wok (old) + Wok
(9)
Similarly the weights and biases are updated for each hidden unit. Finally, the stopping condition is tested. The stopping condition may be minimization of the errors, number of epochs etc.
Algorithm Neural network using back propagation algorithm is applied, using 5/6th of the data for training and 1/6th of the data for validation [29,30]. The training algorithm used in the back propagation network is as follows: Step 1 – Read the input image. Step 2 – Separate the RGB components in R, G, B different planes. Step 3 – Morphological opening operation is performed to fill the vessel discontinuities. Step 4 – Preprocessing steps like mean filtering and Gaussian filtering are performed. Step 5 – RGB features are extracted for the preprocessed image and its mean is calculated. Step 6 – Gray scale image is obtained for further processing. Step 7 – Gabor features are extracted at different orientations. Step 8 – Moment invariants-based features are extracted. Step 9 – Samples are taken from vessel and nonvessel regions and they are trained using neural networks. Step10 – Segmentation of blood vessels is obtained as output. For each training pair, the neural network performs the following tasks. At the first stage, during the initialization of weights, a few small random variables are assigned. In the next stage, i.e., feed forward stage, each input unit (Xi ) receives an input signal and transmits this signal to all the units in the layer above i.e. hidden units, z1 , . . ., zp . Each hidden unit calculates the activation function, Zj = f (zinj ) where, z−inj = Voj +
Error information term is given as,
n
Xi Vij
(3)
i=1
Voj is the bias on hidden unit j. Each hidden unit sends this signal to all output units. Each output unit (yk , k = 1. . . m) then sums its weighted input signals and applies its activation function Yk to calculate the output signals. Yk = f (y−ink ) where, y−ink = wok +
p
zj wjk
(4)
j=1
wok is the bias on output unit k. During back propagation of errors, each output unit compares its calculated activation with its target value tk , to find out the related errors for that pattern.
Experimental results analysis The automatic vessel segmentation algorithm proposed in this paper was tested with the images of DRIVE database. This technique has been proved as one of the most valuable tools for the segmentation of blood vessels in retinal images. Illustrative vessel segmentation results for a few images from DRIVE database are shown in Fig. 5. The neural network based vessel segmentation using back propagation algorithm employed in this paper has a computationally demanding training phase. Nevertheless, it guarantees an efficient vessel classification phase and has proven excellent performance. In addition, the techniques presented here are conceptually quite easier and can be implemented very effectively. Our technique comes under the basic classification of pixel processing-based approach. Theoretically, only the proper selection of initial weights would result in successful network training and ensure better classification performance. It is found that the performance of the network improves with increased number of randomly selected training samples. Based on the experimental results, it is seen that if the number of hidden layers increases, the convergence becomes slower and the training becomes tedious. It is also observed that, if the number of hidden layer increases, a larger learning rate is essential. However, if the learning rate becomes too large, it results in oscillations. Performance evaluation We have tested and evaluated our method using the images from DRIVE database. For this purpose, the training set was formed by pixel samples from the twenty labeled training images. Tests were performed by applying Gabor filters and varying the Gabor orientations at different angles. Prediction of the moment invariants-based features was also considered. Finally, tests were performed by taking samples from vessel and nonvessel regions and these data were trained using neural networks. In our method, we have selected segmentation accuracy as the performance measure. These performance measures were evaluated using ROC curves, which are plots of true positive fractions versus false positive fractions for varying thresholds on the posterior probabilities.
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
99
Fig. 5. Segmentation results produced for retinal images shown in Fig. 1, by training the neural network on each of these images.
The accuracy was estimated by determining the true positive fraction, which is the ratio of the number of true positives to the total number of vessel pixels in the ground truth segmentations. The ratio of the number of false positives to the total number of nonvessel pixels in the ground truth segmentations determines the false positive fraction. A pair formed by a true positive fraction and a false positive fraction is plotted on the graph, which produces a curve as shown in Fig. 6. In our experiments, these fractions are calculated over all test images, considering only pixels inside the field of view. The manual segmentation result provided along with each database image seems to be the best standard for evaluating the performance measures. To exhibit the performance by employing Gabor features and moment invariants-based features using neural networks, we present the results obtained by our method and compared them with the results obtained by other existing methods, involving matched filters and adaptive local thresholding techniques. The Gabor feature extraction at different orientations used in this method increases the accuracy of the segmentation of vessels with different diameters. This method proves itself to be efficient in enhancing the vessel contrast, thereby filtering the noise. We have used a straight forward MATLAB implementation for testing the images from DRIVE database and we have obtained the best performed vessel segmentation results.
Fig. 6. ROC curve for classification on the DRIVE database using Gabor and moment invariants-based features in neural networks.
Discussion and conclusion In this paper, we have presented a new scheme which automatically tracks and segments the vessels in fundus images using computer based algorithms, without the need of any user intervention. Automated segmentation of blood vessels from fundus images provides the means for automated examination and assessment of retinal diseases by the eye care specialists. When the number of vessels in a retinal image is more, or when large number of images is captured, manual detection of the characteristics of vessels becomes impossible. Automated segmentation is the only possibility in this situation. It is technically hard to resolve for an ophthalmologist to predict the exact position of blood vessels, especially at the edges and hence this accuracy of the identification of blood vessels may tend to vary. However, the accuracy of detection and segmentation of retinal blood vessels using our method was better than other existing methods. One of the highlights of this proposed technique is its adaptability to particular image intensity properties, as most of the algorithm settings are exclusively based on threshold values computed from the image information. Our method assigns a precise classification result as vessel or nonvessel to each pixel, compared to other vessel segmentation algorithms. If proper manual segmentations are available, our approach allows the methods to be trained for different types of images such as gray-level angiograms or colored images. Furthermore, the segmentation of blood vessels from the retinal images will assist the examination of fundal disorders. Neural network based analysis has already been employed in other areas of medicine such as diagnosis of breast cancer and even in the field of ophthalmology for visual field analysis. Using neural networks, the features of diabetic retinopathy can also be detected with minimal preprocessing stages. Once after identifying the blood vessels and segmenting it, the data from these regions can be analyzed for abnormality. Of course, a few diseases present in the retina may change the appearance of these blood vessel regions which can reduce their detection. However, our algorithm was designed to minimize these risks, especially with retinal images of persons affected with diabetic retinopathy. The detection of some of the other features of diabetic retinopathy such as hemorrhages and exudates can be further evaluated by the analysis of characteristics of these blood vessels. From our vessel segmentation approach, only the outline skeleton of the segmentations for the extraction of features from the vasculature is obtained. Hence, based on the applications, different evaluation procedures can be adopted to grade the severity of the diseases. This retinal vessel segmentation technique gives knowledge about the location of vessels which paves a way for the screening of diabetic retinopathy.
100
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
While assessing the obtained segmentation results, a few limitations like partial missing of very thin vessel branches are noticed. These can be obviously noticed only in very few images as a consequence of the variability of intensity and contrast among vessels and the other regions. It can be compensated by using vessel enhancement procedure. Such misdetections can be minimized by using more flexible classification process for every image point and thereby, the overall performance of this method can be improved. One of the practical applications of this method is to apply the results of this segmentation algorithm for providing better performance analysis in computer-aided diagnosis system for retinal images. The main advantage of our method is the ability to identify and classify the image pixels as vessels or nonvessels, automatically. This allows us to analyze large number of images and to characterize many more properties of the retinal vasculature. The demonstrated effectiveness, together with its simplicity, makes this automated blood vessel segmentation method a suitable tool for the complete prescreening system for early diabetic retinopathy detection. References [1] M. Garcia, C.I. Sanchez, M.I. Lopez, D. Abasolo, R. Hornero, Neural network based detection of hard exudates in retinal images, Comput. Methods Programs Biomed. 93 (2009) 9–19. [2] A.M. Mendonc¸a, Aurélio Campilho, Segmentation of retinal blood vessels by combining the detection of centerlines and morphological reconstruction, IEEE Trans. Med. Imaging 25 (9) (2006) 1200–1213. [3] A.J. Frame, P.E. Undrill, M.J. Cree, J.A. Olson, K.C. McHardy, P.F. Sharp, J.V. Forrester, A comparison of computer based classification methods applied to the detection of microaneurysms in ophthalmic fluorescein angiograms, Comput. Biol. Med. 28 (3) (1998) 225–238. [4] M. Larsen, J. Godt, N. Larsen, H. Lund-Andersen, A.K. Sjølie, E. Agardh, H. Kalm, M. Grunkin, D.R. Owens, Automated detection of fundus photographic red lesions in diabetic retinopathy, Invest. Ophthalmol. Vis. Sci. 44 (2) (2003) 761–766. [5] C. Oyster, The Human Eye: Structure and Function, Sinauer Associates, Sunderland, MA, 1999. [6] A. Hoover, M. Goldbaum, Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels, IEEE Trans. Med. Imaging 22 (8) (2003) 951–958. [7] C. Sinthanayothin, J.F. Boyce, H.L. Cook, T.H. Williamson, Automated localization of the optic disc, fovea, and retinal blood vessels from digital color fundus images, Br. J. Ophthalmol. 83 (8) (1999) 902–910. [8] C. Sinthanayothin, J.F. Boyce, T.H. Williamson, H.L. Cook, E. Mensah, S. Lal, D. Usher, Automated detection of diabetic retinopathy on digital fundus images, Diabet. Med. 19 (2) (2002) 105–112. [9] K.A. Vermeer, F.M. Vos, H.G. Lemij, A.M. Vossepoel, A model based method for retinal blood vessel detection, Comput. Biol. Med. 34 (2004) 209–219. [10] S. Chaudhuri, S. Chateterjee, N. Katz, M. Nelson, M. Goldbaum, Detection of blood vessels in retinal images using two-dimensional matched filters, IEEE Trans. Med. Imaging 8 (3) (1989) 263–269.
[11] T. Chanwimaluang, G. Fan, An efficient algorithm for extraction of anatomical structures in retinal images, in: Proc. ICIP, 2003, pp. 1193–1196. [12] A. Hoover, V. Kouznetsova, M. Goldbaum, Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response, IEEE Trans. Med. Imaging 19 (3) (2000) 203–211. [13] M.E. Martinez-Perez, A.D. Hughes, A.V. Stanton, S.A. Thom, A.A. Bharath, K.H. Parker, Segmentation of retinal blood vessels based on the second directional derivative and region growing, in: Proc. ICIP, 1999, pp. 173–176. [14] M.E. Martinez-Perez, A.D. Hughes, A.V. Stanton, S.A. Thom, A.A. Bharath, K.H. Parker, Scale-space analysis for the characterization of retinal blood vessels, in Medical Image Computing and Computer- Assisted Intervention—MICCAI’99, C. Taylor and A. Colchester, (Eds.), Springer: New York, Lect. Notes Comput. Sci. 16794 (1999) 90–97. [15] Y. Wang, S.C. Lee, A fast method for automated detection of blood vessels in retinal images, in: IEEE Comput. Soc. Proc. Asilomar Conf., 1998, pp. 1700–1704. [16] X. Jiang, D. Mojon, Adaptive local thresholding by verification based multithreshold probing with application to vessel detection in retinal images, IEEE Trans. Pattern Anal. Mach. Intell. 254 (1) (2003) 131–137. [17] F. Zana, J.-C. Klein, Segmentation of vessel-like patterns using mathematical morphology and curvature evaluation, IEEE Trans. Med. Imaging 11 (7) (2001) 1111–1119. [18] M. Niemeijer, J. Staal, B. van Ginneken, M. Loog, M.D. Abràmoff, Comparative study of retinal vessel segmentation methods on a new publicly available database, in: M. Fitzpatrick, M. Sonka (Eds.), Proc SPIE Med. Imag., vol. 5370, 2004, pp. 648–656. [19] J. Staal, M.D. Abramoff, M. Niemeijer, M.A. Viergever, B. van Ginneken, Ridgebased vessel segmentation in color images of the retina, IEEE Trans. Med. Imaging 23 (4) (2004) 501–509. [20] L. Zhou, M.S. Rzeszotarski, L.J. Singerman, J.M. Chokreff, The detection and quantification of retinopathy using digital angiograms, IEEE Trans. Med. Imaging 13 (4) (1994) 619–626. [21] Y. Tolias, S.M. Panas, A fuzzy vessel tracking algorithm for retinal images based on fuzzy clustering, IEEE Trans. Med. Imaging 17 (2) (1998) 263–273. [22] A. Can, H. Shen, J.N. Turner, H.L. Tanenbaum, B. Roysam, Rapid automated tracing and feature extraction from retinal fundus images using direct exploratory algorithms, IEEE Trans. Inf. Technol. Biomed. 3 (2) (1999) 125–138. [23] O. Chutatape, L. Zheng, S.M. Krishnan, Retinal blood vessel detection and tracking by matched Gaussian and Kalman filters, in: Proc. 20th Annu. Int. Conf. IEEE Eng. Med. Biol., 1998, pp. 3144–3149. [24] V.S. Lee, R.M. Kingsley, E.T. Lee, et al., The diagnosis of diabetic retinopathy. Ophthalmoscopy versus fundus photography, Ophthalmology 100 (1993) 1504–1512. [25] Research Section, Digital Retinal Image for Vessel Extraction (DRIVE) Database, Univ. Med. Center Utrecht, Image Sci. Inst., Utrecht, The Netherlands, 2014 [Online]. Available: http://www.isi.uu.nl/Re-search/Databases/DRIVE [26] R.C. Gonzalez, R.E. Woods, Digital Image Processing, 229–237, Addison-Wesley, Reading, 1993, pp. 583–586. [27] D.L. Toulson, J.F. Boyce, Segmentation of MR images using neural nets, in: Proc. Br. Mach. Vision Conf., 1991, pp. 284–292. [28] D.L. Toulson, J.F. Boyce, Segmentation of MR images using neural nets (expanded version), Image Vis. Comput. 10 (1992) 324–328. [29] S.C. Lo, M.T. Freedman, J.S. Lin, et al., Automatic lung nodule detection using profile matching and back-propagation neural network techniques, J. Digit. Imaging 6 (1993) 48–54. [30] M.L. Astion, P. Wilding, The application of backpropagation neural networks to problems in pathology and laboratory medicine, Arch. Pathol. Lab. Med. 116 (1992) 995–1001.
Contents lists available at ScienceDirect
Applied Soft Computing journal homepage: www.elsevier.com/locate/asoc
Retinal vessel segmentation employing ANN technique by Gabor and moment invariants-based features S. Wilfred Franklin a , S. Edward Rajan b,∗ a b
CSI Institute of Technology, Thovalai, Tamil Nadu, India Mepco Schlenk Engineering College, Sivakasi, Tamil Nadu, India
a r t i c l e
i n f o
Article history: Received 16 March 2012 Received in revised form 4 July 2012 Accepted 21 April 2014 Available online 21 May 2014 Keywords: Diabetic retinopathy Retinal vessel segmentation Artificial neural networks Retinal images Retinal vasculature
a b s t r a c t Diabetic retinopathy (DR) is the major ophthalmic pathological cause for loss of eye sight due to changes in blood vessel structure. The retinal blood vessel morphology helps to identify the successive stages of a number of sight threatening diseases and thereby paves a way to classify its severity. This paper presents an automated retinal vessel segmentation technique using neural network, which can be used in computer analysis of retinal images, e.g., in automated screening for diabetic retinopathy. Furthermore, the algorithm proposed in this paper can be used for the analysis of vascular structures of the human retina. Changes in retinal vasculature are one of the main symptoms of diseases like hypertension and diabetes mellitus. Since the size of typical retinal vessel is only a few pixels wide, it is critical to obtain precise measurements of vascular width using automated retinal image analysis. This method segments each image pixel as vessel or nonvessel, which in turn, used for automatic recognition of the vasculature in retinal images. Retinal blood vessels are identified by means of a multilayer perceptron neural network, for which the inputs are derived from the Gabor and moment invariants-based features. Back propagation algorithm, which provides an efficient technique to change the weights in a feed forward network is utilized in our method. The performance of our technique is evaluated and tested on publicly available DRIVE database and we have obtained illustrative vessel segmentation results for those images. © 2014 Published by Elsevier B.V.
Introduction Diabetic retinopathy (DR) is a result of long-term diabetes and it is a severe and the most common sight threatening eye disease, which causes blindness among working-age people around the world [1]. Major vision loss due to DR is highly preventable with proper screening and timely diagnosis at the earlier stages. However, DR is not painful and the diabetic patients perceive no symptom until visual loss occurs and hence they need periodical eye-fundus examination to ensure that treatment is received in time. Evaluation of the characteristics of retinal blood vessels plays an important role in the diagnosis of diseases based on vascular pathology. The various features of retinal vessels such as length, width, tortuosity and branching pattern provide new techniques to diagnose various diseases like diabetes, arteriosclerosis, hypertension, cardiovascular disease and stroke. Retinal images provide valuable information related to human eye, by which the vascular
∗ Corresponding author. Tel.: +91 04562 235308. E-mail addresses: [email protected] (S.W. Franklin), [email protected] (S.E. Rajan). http://dx.doi.org/10.1016/j.asoc.2014.04.024 1568-4946/© 2014 Published by Elsevier B.V.
condition can be accurately observed and analyzed [2]. The only part of the central circulation that can be viewed directly and analyzed is the retinal vessel. Changes in blood vessel structure and vessel distribution, caused by diabetic retinopathy can lead to new vessel growth, which in turn instigates vision impairment. Hence retinal vessel segmentation becomes an essential tool for the detection of any variations that occurs in blood vessels. Retinal blood vessel segmentation gives the detailed information about the location of vessels which helps in the screening of diabetic retinopathy, e.g., in the detection of micro-aneurysms it helps to reduce the number of false positives [3,4]. Observations based on retinal vessel segmentation are quiet complex and hence the traditional way of diagnosis in which the ophthalmologist identifies the anomalies present in the retinal images by examining it, becomes tiresome or even impossible. As the actual size of a typical blood vessel in human retina is very small, just a few pixels wide, measurement of vascular width using retinal image segmentation process becomes critical and challenging. One of the possible solutions for these measurements is the use of computer based automated analysis of optic fundus, a technique widely accepted by the medical community. This computer based analysis helps the medical personnel to detect the changes in blood flow and vessel distribution. It also enables the
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
detection of extra vessel growth. For automatic analysis of optic fundus that detects the blood vessel irregularities, segmentation of the vessels from the background is made initially and thereby further processing like feature extraction can be achieved. In human retina, one of the most important organs is the optic nerve which acts as the convergent point of the blood vessel network. The central retinal artery and central retinal vein flow out through the optic nerve that supplies blood to the upper layers of the retina. Besides, the optic nerve acts as a channel to convey the information from the eye to the brain. In early stages, most of the retinal pathologies affects locally and does not distress the entire retina. But retinal pathology on or near the optic nerve may severely affect the vision even at the early stages because optic nerve is the most essential part for vision [5]. A few observations found in several important retinopathies are attenuation changes, focal narrowing and occlusion of retinal arteries. The diameter and shape of a retinal vessel plays a key role as indicators in ophthalmologic studies. These changes give valuable information to identify the successive stages of diseases and their response to various therapies. The optic nerve can be observed in a close view of the retinal fundus, where the optic disc is the portion of the nerve that is visible or perceivable by the eye. Fundus imaging is one of the popular clinical procedures available to record this close view observations of the retina. This fundus imaging procedure is also used for the diagnosis and evaluation of the healthy and non-healthy retinas of human eye [6]. In a healthy retina the optic nerve has a standard identifiable size, shape, color and location relative to the blood vessels. Nevertheless, in a retina containing lesions, any one or more of these properties may be deviated from its standard level and show a large variance. At various stages of the disease, the vascular network in retina is very much affected and hence various morphological changes occur in retinal vessels. We can substantially observe enough geometrical changes in diameter, branching angle, length or tortuosity in the retinal blood vessels due to diseases. The segmentation and measurement of retinal blood vessels can be used to grade the severity of certain diseases. The sign of risk level for diabetic retinopathy is the variation in width of retinal vessels within the fundus. One of the most important tools for the prediction of proliferate diabetic retinopathy is the abnormal variation in diameter along the vein. Moreover, the various retinal microvascular abnormalities predicted are seen to be the early symptoms for the risk of stroke. In all these cases, the desired focus is on the variation in diameter of the vessel and not in the exact diameter of the vessel. An alternative application of retinal vessel segmentation is biometric identification using distinctive retinal vascular network for each individual. This research paper is organized as follows: Section “Overview of retinal vessel segmentation” reviews various other vessel segmentation methods, Section “Methodology” describes and illustrates the methods of our proposal for vessel segmentation, Section “Experimental results analysis” presents its results and compares the performance with some of the techniques previously published and Section “Discussion and conclusion” puts forward the discussion and concludes this work.
Overview of retinal vessel segmentation Segmentation of vessels in retinal images is achieved by classifying each image pixel as vessel or nonvessel based on the local image features. In two-dimensional images, the techniques adopted for identifying vessels are based on specific properties of the vascular segments. There are two basic approaches that are generally considered for the identification of general vascular segments, which also play a key role in retinal vessel segmentation applications. The
95
algorithms used for the segmentation of blood vessels are broadly classified as pixel processing-based methods and vessel tracking or vectorial tracking methods [2]. Supplementary methods used for vessel segmentation in retinal images can be classified into two groups. The first group consists of rule-based methods and the second group consists of supervised methods, where training of manually labeled images is used. Neural network based pixel classification scheme falls under the second category [7,8], which is followed in this paper. Pixel processing-based methods normally consist of two phases. In the first phase, an enhancement procedure is implied and it selects an initial set of pixels, which is further ensured as vessels in the second phase [9]. The retinal vessel segmentation method presented in Ref. [2] consists of three processing phases. In the first phase, background normalization of monochromatic input image is performed and later thin vessel enhancement procedures are used. In the second phase, the vessel centerline candidate points are selected and subsequently these points are connected and based on vessel length, the validation of centerline segment candidates is achieved. In the third phase, vessels with different widths are enhanced and processed using binary morphological reconstruction technique and vessel filling is achieved using region growing process. Soares et al. have proposed an algorithm, where retinal blood vessels are detected using Gabor wavelets by representing each pixel by a feature vector and then by using Bayesian classifier with Gaussian mixtures, each pixel is classified as either a vessel or nonvessel pixel and thus segmentation is achieved. The concept of matched filter detection was proposed by Chaudhuri et al. [10], where twelve rotated versions of 2-D Gaussian shaped templates are used to search vessel segments along all possible directions. The resultant image produced is the binary representation of the retinal vasculature. Likewise, segmentation of retinal vessels are also obtained by matched filtering approaches using global [11] or local thresholding strategies [12]. Also for the purpose of vessel borders extraction, differential filters based on either first-or-second order derivatives are used. A two stage region growing procedure was proposed by Martinez-Perez et al. where features derived from image derivatives are used in the segmentation of retinal vessels [13,14]. As stated in Ref. [15], the edge detection, matched filtering and region growing procedures can also be collectively used for the automated detection of retinal blood vessels. Jiang et al. have proposed a technique of adaptive local thresholding based on the use of verification-based multithreshold scheme combined with classification procedures in Ref. [16], used for the detection of vessels in retinal images. A technique used for the segmentation of vessel like patterns in retinal images that combines morphological filters and cross-curvature evaluation was proposed by Zana et al. in Ref. [17]. In this approach, vessel segments alone are considered as image feature and hence using morphological filters simplifies the computation of cross-curvature. An algorithm for retinal vessel segmentation based on the classification of pixels using simple feature vector was proposed by Niemeijer et al. in Ref. [18]. A new method of segmentation of blood vessels in retinal images was proposed by Staal et al. in Ref. [19]. This supervised method is called primitive-based method and this algorithm is based on the extraction of image ridges used to compose primitives that describe the linear segments called line elements. The pixels those are assigned to the closest line element partitions the image in the form of image patches and are classified using a set of features from the corresponding line and image patch. In addition, a technique based on neural network is used to identify the retinal blood vessels, where the inputs are obtained using principal component analysis and then edge detection technique is used. Vessel tracking methods use the concept of measuring some local image properties to locate the vessel points which are used for tracing the vasculature. In these types of algorithms, both the
96
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
extraction of the image feature and the recognition of the vessel structure are performed simultaneously. An algorithm which relies on matched filtering approach was presented in Ref. [20], and is initiated by start and end points followed by automatic detection of vessel boundaries and then tracks the centerline of the vessel and finally it extracts the useful parameters of clinical interest. The unsupervised fuzzy vessel tracking algorithm presented in Ref. [21] is virtually parameter-free and it performs automatic initialization. It also overcomes the problem of vessel profile modeling, tracks fundus vessels automatically and the classification of vessel and nonvessel regions along a vessel profile is determined using fuzzy C-means clustering algorithm. The algorithm presented in Ref. [22] performs automatic adaption without manual initialization. It automatically detects the seed-point of tracing, described as local gray-level minima and it is applied directly at the image intensity level without preprocessing. Consequently, sequences of exploratory searches are initiated to detect the vessel boundaries and these models are used in the estimation of next vessel point. In Ref. [23] Chutatape et al. have proposed a technique where the tracing process starts from the circumference of the optic disc and the use of Kalman filter estimates the next point of location in the tracking procedure.
Image database We have tested and evaluated the technique used in this method by making use of the retinal color images available on DRIVE database [25]. This database has been widely used as ground truth for performance evaluation by other researchers to test their vessel segmentation methodologies. The retinal images available in this database were captured using a Canon CR5 nonmydriatic 3CCD camera with a 45◦ Field-of-View (FOV), in digital form. The DRIVE database consists of forty eye-fundus color images which are divided into training and test set, each set containing twenty images out of which the training set has three images with pathology and the test set has four images with pathology. The images available in the training set were segmented only once and the images in the test set were segmented twice, resulting in two sets, the first set is accepted as the ground truth and used for algorithm performance evaluation in literature. The complete results of the manual segmentation are also available for all the images of the two sets. The color images of the retina available in the database are of size 768 × 584 pixels and 8 bits per color channel, and were initially saved in JPEG format. Moreover, the database includes masks with the delimitation of a Field-of-View (FOV) of approximately 540 pixels in diameter for each image, which is used for the performance evaluation of our method.
Methodology This paper proposes a retinal vessel segmentation technique using Artificial Neural Networks (ANN). The main features of fundus retinal images are blood vessels, which are used to detect the presence of anomalies present in human retina automatically. These retinal blood vessels are identified by using a multilayer perceptron neural network, for which the inputs are derived from Gabor and moment invariants-based features. Using neural networks, some of the features such as hemorrhages and exudates can be detected, which helps to examine whether retinopathy is present or not. The images with retinopathy also require a grading measure, which is a complex task than identifying the presence of retinopathy. It can be examined by a trained person, because the computer based analysis detects only the changes such as neovascularisation, cotton wool spots, perifoveal exudation and vascular changes [24]. The basic step to achieve this diagnosis procedure is to locate the important regions of the fundus, such as blood vessels, automatically. The data from these segmented blood vessel regions can further be processed and analyzed for examining the features of certain eye diseases. In this paper, computer based image analysis using neural network is applied to detect and segment the blood vessels from retinal images.
Preprocessing Preprocessing of images is the first step carried out before presenting the input to the neural network. The retinal color images taken from the DRIVE database are preprocessed, where each pixel contained three values-red, green and blue. A few examples given as input images to the multilayer perceptron neural network can be seen in Fig. 1. The RGB features are extracted for the preprocessed image, its mean is calculated and the gray scale image is obtained for further processing. The true color image RGB is converted to the grayscale intensity image to evaluate its adequate ability for the segmentation of the retinal blood vessels. The conversion is done by eliminating the hue and saturation information, while retaining the luminance [26]. The given input images are further processed using mean filtering and a few examples are shown in Fig. 2. In the next step, the retinal images are further processed and smoothened using Gaussian filters and a few examples can be seen in Fig. 3. Then the Gabor features are extracted at different orientations and a few examples are shown in Fig. 4. Gabor filtering is done and
Fig. 1. Input images given to multilayer perceptron neural network.
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
97
Fig. 2. Preprocessed images obtained from original color fundus images after applying mean filtering.
Fig. 3. Preprocessed images after applying Gaussian filters.
Fig. 4. An example for Gabor features extracted at different orientations.
as per the Gabor coefficients attained, the original images are convolved. Next, the moment invariants-based features are extracted. Now the samples are taken from vessel and nonvessel regions and they are trained using neural networks. The feature vectors extracted from preprocessed retinal images are given as input to this neural network. This technique is based on pixel classification and these classification results are thresholded to classify each pixel as vessel or nonvessel. Neural network training Each pixel of a retinal image is classified as vessel or nonvessel, using neural networks. The computational structure of neural
networks and the human visual system has several similarities. In order to classify each pixel of the retinal image as vessel or not, a multi-layer perceptron neural network is used [27,28]. It is a multilayer, feed forward network using extend gradient-descent based delta-learning rule, commonly known as back propagation rule. Back propagation is a systematic method for training multi-layer artificial neural network. According to the basic definition of feed forward network, a neuron in the hidden layer will receive inputs from all neurons in the previous layer and must feed all neurons in the next layer. Back propagation provides an efficient technique to change the weights in a feed forward network with differentiable activation function units, and to learn a training set of input–output examples. If the number of layers in the feed forward network is
98
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
represented as n, and if Li represents the number of neurons in the ith layer, then the total number of weights is given by: N=
n−1
Li Li+1
where, i = 1, 2, . . ., n
(1)
i=1
The learning capacity (C) is an important parameter which determines the ability of a system to learn. For a three layer network, the learning capacity is bounded by: nh < C < nh log(h) log(nh2 )
(2)
where, h is the number of hidden nodes and n is the number of input nodes. The neural network applied is a three layer perceptron having one input node, five hidden layers and one output node. Before presenting the input to the neural network, preprocessing of the image must be done. The neural network is considered to be trained when the system error falls below an acceptable threshold level and at this stage, the vessels are identified from the background and thus segmented images are obtained.
ık = (tk − yk ) f (y−ink )
(5)
It is used to distribute the error at the output unit yk back to all units in the previous layer. Then each hidden unit (zj , j = 1,. . ., n) sums its delta inputs from units in the above layer as: ı−inj =
m
ıj wjk
(6)
k=1
Similarly the error information term ıj = ı−inj f(z−inj ) is computed for each hidden unit. During the final stage, the weights and biases are updated using the ı factor and the activation function is updated, using ˛ as the learning rate. Each output unit updates its weights and biases by using the weight correction term Wjk and bias correction term Wok , where, Wjk = ˛ ık zj and Wok = ˛ ık
(7)
Therefore, Wjk (new) = Wjk (old) + Wjk and
(8)
Wok (new) = Wok (old) + Wok
(9)
Similarly the weights and biases are updated for each hidden unit. Finally, the stopping condition is tested. The stopping condition may be minimization of the errors, number of epochs etc.
Algorithm Neural network using back propagation algorithm is applied, using 5/6th of the data for training and 1/6th of the data for validation [29,30]. The training algorithm used in the back propagation network is as follows: Step 1 – Read the input image. Step 2 – Separate the RGB components in R, G, B different planes. Step 3 – Morphological opening operation is performed to fill the vessel discontinuities. Step 4 – Preprocessing steps like mean filtering and Gaussian filtering are performed. Step 5 – RGB features are extracted for the preprocessed image and its mean is calculated. Step 6 – Gray scale image is obtained for further processing. Step 7 – Gabor features are extracted at different orientations. Step 8 – Moment invariants-based features are extracted. Step 9 – Samples are taken from vessel and nonvessel regions and they are trained using neural networks. Step10 – Segmentation of blood vessels is obtained as output. For each training pair, the neural network performs the following tasks. At the first stage, during the initialization of weights, a few small random variables are assigned. In the next stage, i.e., feed forward stage, each input unit (Xi ) receives an input signal and transmits this signal to all the units in the layer above i.e. hidden units, z1 , . . ., zp . Each hidden unit calculates the activation function, Zj = f (zinj ) where, z−inj = Voj +
Error information term is given as,
n
Xi Vij
(3)
i=1
Voj is the bias on hidden unit j. Each hidden unit sends this signal to all output units. Each output unit (yk , k = 1. . . m) then sums its weighted input signals and applies its activation function Yk to calculate the output signals. Yk = f (y−ink ) where, y−ink = wok +
p
zj wjk
(4)
j=1
wok is the bias on output unit k. During back propagation of errors, each output unit compares its calculated activation with its target value tk , to find out the related errors for that pattern.
Experimental results analysis The automatic vessel segmentation algorithm proposed in this paper was tested with the images of DRIVE database. This technique has been proved as one of the most valuable tools for the segmentation of blood vessels in retinal images. Illustrative vessel segmentation results for a few images from DRIVE database are shown in Fig. 5. The neural network based vessel segmentation using back propagation algorithm employed in this paper has a computationally demanding training phase. Nevertheless, it guarantees an efficient vessel classification phase and has proven excellent performance. In addition, the techniques presented here are conceptually quite easier and can be implemented very effectively. Our technique comes under the basic classification of pixel processing-based approach. Theoretically, only the proper selection of initial weights would result in successful network training and ensure better classification performance. It is found that the performance of the network improves with increased number of randomly selected training samples. Based on the experimental results, it is seen that if the number of hidden layers increases, the convergence becomes slower and the training becomes tedious. It is also observed that, if the number of hidden layer increases, a larger learning rate is essential. However, if the learning rate becomes too large, it results in oscillations. Performance evaluation We have tested and evaluated our method using the images from DRIVE database. For this purpose, the training set was formed by pixel samples from the twenty labeled training images. Tests were performed by applying Gabor filters and varying the Gabor orientations at different angles. Prediction of the moment invariants-based features was also considered. Finally, tests were performed by taking samples from vessel and nonvessel regions and these data were trained using neural networks. In our method, we have selected segmentation accuracy as the performance measure. These performance measures were evaluated using ROC curves, which are plots of true positive fractions versus false positive fractions for varying thresholds on the posterior probabilities.
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
99
Fig. 5. Segmentation results produced for retinal images shown in Fig. 1, by training the neural network on each of these images.
The accuracy was estimated by determining the true positive fraction, which is the ratio of the number of true positives to the total number of vessel pixels in the ground truth segmentations. The ratio of the number of false positives to the total number of nonvessel pixels in the ground truth segmentations determines the false positive fraction. A pair formed by a true positive fraction and a false positive fraction is plotted on the graph, which produces a curve as shown in Fig. 6. In our experiments, these fractions are calculated over all test images, considering only pixels inside the field of view. The manual segmentation result provided along with each database image seems to be the best standard for evaluating the performance measures. To exhibit the performance by employing Gabor features and moment invariants-based features using neural networks, we present the results obtained by our method and compared them with the results obtained by other existing methods, involving matched filters and adaptive local thresholding techniques. The Gabor feature extraction at different orientations used in this method increases the accuracy of the segmentation of vessels with different diameters. This method proves itself to be efficient in enhancing the vessel contrast, thereby filtering the noise. We have used a straight forward MATLAB implementation for testing the images from DRIVE database and we have obtained the best performed vessel segmentation results.
Fig. 6. ROC curve for classification on the DRIVE database using Gabor and moment invariants-based features in neural networks.
Discussion and conclusion In this paper, we have presented a new scheme which automatically tracks and segments the vessels in fundus images using computer based algorithms, without the need of any user intervention. Automated segmentation of blood vessels from fundus images provides the means for automated examination and assessment of retinal diseases by the eye care specialists. When the number of vessels in a retinal image is more, or when large number of images is captured, manual detection of the characteristics of vessels becomes impossible. Automated segmentation is the only possibility in this situation. It is technically hard to resolve for an ophthalmologist to predict the exact position of blood vessels, especially at the edges and hence this accuracy of the identification of blood vessels may tend to vary. However, the accuracy of detection and segmentation of retinal blood vessels using our method was better than other existing methods. One of the highlights of this proposed technique is its adaptability to particular image intensity properties, as most of the algorithm settings are exclusively based on threshold values computed from the image information. Our method assigns a precise classification result as vessel or nonvessel to each pixel, compared to other vessel segmentation algorithms. If proper manual segmentations are available, our approach allows the methods to be trained for different types of images such as gray-level angiograms or colored images. Furthermore, the segmentation of blood vessels from the retinal images will assist the examination of fundal disorders. Neural network based analysis has already been employed in other areas of medicine such as diagnosis of breast cancer and even in the field of ophthalmology for visual field analysis. Using neural networks, the features of diabetic retinopathy can also be detected with minimal preprocessing stages. Once after identifying the blood vessels and segmenting it, the data from these regions can be analyzed for abnormality. Of course, a few diseases present in the retina may change the appearance of these blood vessel regions which can reduce their detection. However, our algorithm was designed to minimize these risks, especially with retinal images of persons affected with diabetic retinopathy. The detection of some of the other features of diabetic retinopathy such as hemorrhages and exudates can be further evaluated by the analysis of characteristics of these blood vessels. From our vessel segmentation approach, only the outline skeleton of the segmentations for the extraction of features from the vasculature is obtained. Hence, based on the applications, different evaluation procedures can be adopted to grade the severity of the diseases. This retinal vessel segmentation technique gives knowledge about the location of vessels which paves a way for the screening of diabetic retinopathy.
100
S.W. Franklin, S.E. Rajan / Applied Soft Computing 22 (2014) 94–100
While assessing the obtained segmentation results, a few limitations like partial missing of very thin vessel branches are noticed. These can be obviously noticed only in very few images as a consequence of the variability of intensity and contrast among vessels and the other regions. It can be compensated by using vessel enhancement procedure. Such misdetections can be minimized by using more flexible classification process for every image point and thereby, the overall performance of this method can be improved. One of the practical applications of this method is to apply the results of this segmentation algorithm for providing better performance analysis in computer-aided diagnosis system for retinal images. The main advantage of our method is the ability to identify and classify the image pixels as vessels or nonvessels, automatically. This allows us to analyze large number of images and to characterize many more properties of the retinal vasculature. The demonstrated effectiveness, together with its simplicity, makes this automated blood vessel segmentation method a suitable tool for the complete prescreening system for early diabetic retinopathy detection. References [1] M. Garcia, C.I. Sanchez, M.I. Lopez, D. Abasolo, R. Hornero, Neural network based detection of hard exudates in retinal images, Comput. Methods Programs Biomed. 93 (2009) 9–19. [2] A.M. Mendonc¸a, Aurélio Campilho, Segmentation of retinal blood vessels by combining the detection of centerlines and morphological reconstruction, IEEE Trans. Med. Imaging 25 (9) (2006) 1200–1213. [3] A.J. Frame, P.E. Undrill, M.J. Cree, J.A. Olson, K.C. McHardy, P.F. Sharp, J.V. Forrester, A comparison of computer based classification methods applied to the detection of microaneurysms in ophthalmic fluorescein angiograms, Comput. Biol. Med. 28 (3) (1998) 225–238. [4] M. Larsen, J. Godt, N. Larsen, H. Lund-Andersen, A.K. Sjølie, E. Agardh, H. Kalm, M. Grunkin, D.R. Owens, Automated detection of fundus photographic red lesions in diabetic retinopathy, Invest. Ophthalmol. Vis. Sci. 44 (2) (2003) 761–766. [5] C. Oyster, The Human Eye: Structure and Function, Sinauer Associates, Sunderland, MA, 1999. [6] A. Hoover, M. Goldbaum, Locating the optic nerve in a retinal image using the fuzzy convergence of the blood vessels, IEEE Trans. Med. Imaging 22 (8) (2003) 951–958. [7] C. Sinthanayothin, J.F. Boyce, H.L. Cook, T.H. Williamson, Automated localization of the optic disc, fovea, and retinal blood vessels from digital color fundus images, Br. J. Ophthalmol. 83 (8) (1999) 902–910. [8] C. Sinthanayothin, J.F. Boyce, T.H. Williamson, H.L. Cook, E. Mensah, S. Lal, D. Usher, Automated detection of diabetic retinopathy on digital fundus images, Diabet. Med. 19 (2) (2002) 105–112. [9] K.A. Vermeer, F.M. Vos, H.G. Lemij, A.M. Vossepoel, A model based method for retinal blood vessel detection, Comput. Biol. Med. 34 (2004) 209–219. [10] S. Chaudhuri, S. Chateterjee, N. Katz, M. Nelson, M. Goldbaum, Detection of blood vessels in retinal images using two-dimensional matched filters, IEEE Trans. Med. Imaging 8 (3) (1989) 263–269.
[11] T. Chanwimaluang, G. Fan, An efficient algorithm for extraction of anatomical structures in retinal images, in: Proc. ICIP, 2003, pp. 1193–1196. [12] A. Hoover, V. Kouznetsova, M. Goldbaum, Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response, IEEE Trans. Med. Imaging 19 (3) (2000) 203–211. [13] M.E. Martinez-Perez, A.D. Hughes, A.V. Stanton, S.A. Thom, A.A. Bharath, K.H. Parker, Segmentation of retinal blood vessels based on the second directional derivative and region growing, in: Proc. ICIP, 1999, pp. 173–176. [14] M.E. Martinez-Perez, A.D. Hughes, A.V. Stanton, S.A. Thom, A.A. Bharath, K.H. Parker, Scale-space analysis for the characterization of retinal blood vessels, in Medical Image Computing and Computer- Assisted Intervention—MICCAI’99, C. Taylor and A. Colchester, (Eds.), Springer: New York, Lect. Notes Comput. Sci. 16794 (1999) 90–97. [15] Y. Wang, S.C. Lee, A fast method for automated detection of blood vessels in retinal images, in: IEEE Comput. Soc. Proc. Asilomar Conf., 1998, pp. 1700–1704. [16] X. Jiang, D. Mojon, Adaptive local thresholding by verification based multithreshold probing with application to vessel detection in retinal images, IEEE Trans. Pattern Anal. Mach. Intell. 254 (1) (2003) 131–137. [17] F. Zana, J.-C. Klein, Segmentation of vessel-like patterns using mathematical morphology and curvature evaluation, IEEE Trans. Med. Imaging 11 (7) (2001) 1111–1119. [18] M. Niemeijer, J. Staal, B. van Ginneken, M. Loog, M.D. Abràmoff, Comparative study of retinal vessel segmentation methods on a new publicly available database, in: M. Fitzpatrick, M. Sonka (Eds.), Proc SPIE Med. Imag., vol. 5370, 2004, pp. 648–656. [19] J. Staal, M.D. Abramoff, M. Niemeijer, M.A. Viergever, B. van Ginneken, Ridgebased vessel segmentation in color images of the retina, IEEE Trans. Med. Imaging 23 (4) (2004) 501–509. [20] L. Zhou, M.S. Rzeszotarski, L.J. Singerman, J.M. Chokreff, The detection and quantification of retinopathy using digital angiograms, IEEE Trans. Med. Imaging 13 (4) (1994) 619–626. [21] Y. Tolias, S.M. Panas, A fuzzy vessel tracking algorithm for retinal images based on fuzzy clustering, IEEE Trans. Med. Imaging 17 (2) (1998) 263–273. [22] A. Can, H. Shen, J.N. Turner, H.L. Tanenbaum, B. Roysam, Rapid automated tracing and feature extraction from retinal fundus images using direct exploratory algorithms, IEEE Trans. Inf. Technol. Biomed. 3 (2) (1999) 125–138. [23] O. Chutatape, L. Zheng, S.M. Krishnan, Retinal blood vessel detection and tracking by matched Gaussian and Kalman filters, in: Proc. 20th Annu. Int. Conf. IEEE Eng. Med. Biol., 1998, pp. 3144–3149. [24] V.S. Lee, R.M. Kingsley, E.T. Lee, et al., The diagnosis of diabetic retinopathy. Ophthalmoscopy versus fundus photography, Ophthalmology 100 (1993) 1504–1512. [25] Research Section, Digital Retinal Image for Vessel Extraction (DRIVE) Database, Univ. Med. Center Utrecht, Image Sci. Inst., Utrecht, The Netherlands, 2014 [Online]. Available: http://www.isi.uu.nl/Re-search/Databases/DRIVE [26] R.C. Gonzalez, R.E. Woods, Digital Image Processing, 229–237, Addison-Wesley, Reading, 1993, pp. 583–586. [27] D.L. Toulson, J.F. Boyce, Segmentation of MR images using neural nets, in: Proc. Br. Mach. Vision Conf., 1991, pp. 284–292. [28] D.L. Toulson, J.F. Boyce, Segmentation of MR images using neural nets (expanded version), Image Vis. Comput. 10 (1992) 324–328. [29] S.C. Lo, M.T. Freedman, J.S. Lin, et al., Automatic lung nodule detection using profile matching and back-propagation neural network techniques, J. Digit. Imaging 6 (1993) 48–54. [30] M.L. Astion, P. Wilding, The application of backpropagation neural networks to problems in pathology and laboratory medicine, Arch. Pathol. Lab. Med. 116 (1992) 995–1001.