外文翻译图像特性的投影快速定位视神经盘

英文原文 Fast Localization of the Optic Disc UsingProjection of Image FeaturesAhmed E. Mahfouz and Ahmed S. Fahmy Abstract-Optic Disc (OD) localization is an important pre-processing step that signicantly simplies subsequent segmentation of the OD and other retinal structures. Current OD localization techniques suffer from impractically-high computation times (few minutes per image). In this work, we present a fast technique that requires less than a second to localize the OD. The technique is based upon obtaining two projections of certain image features that encode the x- and y- coordinates of the OD. The resulting 1-D projections are then searched to determine the location of the OD. This avoids searching the 2-D image space and, thus, enhances the speed of the OD localization process. Image features such as retinal vessels orientation and the OD brightness are used in the current method. Four publicly-avail-able databases, including STARE and DRIVE, are used to evaluate the proposed technique. The OD was successfully located in 330 images out of 340 images (97%) with an average computation time of 0.65 s. Index TermsImage features, localization, optic disc, projection.I. INTRODUCTIONWith the new advances in digital modalities for retinal imaging, there is a progressive need of image processing tools that provide fast and reliable segmentation of retinal anatomical structures. The optic disc(OD) is a major retinal structure that usually appears in retinal images as a circular bright object 1. It is the region where the optic nerve and the retinal and choroidal vessels emerge into the eye 2.A large number of algorithms have been proposed in literature to segment the OD; this includes the use of Hough Transform 35, active contour models 6, and gradient vector ow (GVF) 7 and 8.Nevertheless, the success and efciency of these algorithms depend mainly upon determining a seed point inside the OD, i.e., localization of the OD 68. Although manual localization of the OD is sufcient, the process can be prohibitively cumbersome when dealing with large number of images. This has stimulated several research groups to develop algorithms for automatic localization of the OD 1, 2 and 912. OD localization can also be useful for a number of applications. For example, the OD location can serve as a landmark for localizing and segmenting other anatomical structures such as the fovea.(where the distance between the OD center and the center of the fovea is roughly constant) 2. The location can also be used to classify left and right eyes in fovea-centered retinal images 13. In addition, the detection of OD location is sometimes necessary for computing some important diagnostic indices for hypertensive retinopathy based upon vasculature such as central retinal artery equivalent (CRAE) and central retinal vein equivalent (CRVE) 10. Also, since the OD can beeasily confounded with large exudates and lesions, the detection of its location is important to remove it from a set of candidate lesions 9.In normal eyes, automatic localization of the OD is simple because it has well-dened features. Nevertheless, developing fast and robust methods for automatic localization of the OD could be very challenging due to the presence of retinal pathologies that alter the appearance of the OD signicantly and/or have similar properties to the OD 3. OD localization methods can be classied into two main categories, appearance-basedmethods and model-basedmethods. Appearance-basedmethods identify the location of the OD as the location of the brightest round object within the retinal image. These methods include techniques such as intensity thresholding 4 and 5, highest average variation 14, matched spatial lter 12, and principle component analysis10. Although these methods are simple and have high success rates in normal images, they fail to correctly localize the OD in diseased retinal images where the pathologies have similar appearance properties to the OD.Model-based methods depend mainly upon extracting and analyzing the structure of the retinal vessels and dening the location of the OD as the point where all the retinal vessels originate 1, 2 and 9.Techniques such as geometrical models 9, convergence of vasculature 1, and Vessels direction matched lter 2 have a relatively high success rate in diseased images, but they are computationally very expensive because they require segmentation of the retinal vessels as an initial step of the localization process. For example, the geometrical model-based method proposed by Foracchia et al. 9 achieves a success rate of 97.5%, but it takes an average computation time of 2 min to localize the OD in a given image. The Vessels direction matched lter described by Youssif et al. 2 achieves an accuracy of 98.8%, but it takes an average computation time of 3.5 min per image to correctly locate the OD. The Hausdorff-based template matching technique proposed by Lalonde et al. 11 has a good computation time (1.6 0.3s), but the technique relies on three major assumptions: 1) the imaging protocol is known, 2) the OD represents a bright region, and 3) the OD is a circular object. The rst assumption requires that the graders mark each image as left, right, or OD centered. The other two assumptions are usually violated in retinal images with pathologies (e.g., 17, images 1, 46, 13, 17, 1921, 2527, 34, 36, 46, and 219) leading to low success rate of the OD localization process, i.e., 71.6%in STARE data-base, as reported by Youssef et al. 2.In this work, a novel fast technique for OD localization is proposed. The new method can be classied as a model-based method in which the OD is considered the region where the main retinal vessels originate in a vertical direction. The computation time of the localization process is signicantly enhanced by reducing the problem from one 2-D localization problem to two 1-D problems that does not require segmentation of the retinal vessels. The remaining sections of this manuscript are organized as follows; Section II-A describes the “easy-to-compute” image features that can be used to decompose the image into two 1-D signals. Section II-B contains the methodologies of determining the horizontal and the vertical locations of the OD from the Fig. 1. Feature_map_1 showing a horizontally sliding window at two differentlocations, the sliding direction and the projection direction.resulting two 1-D signals. Section II-C proposes a geometry-based method that can be used to enhance the robustness of the localization process. Section II-D contains the detailed algorithm that can be used to implement the proposed technique and reproduce the results which are displayed in Section III. Section IV contains a discussion of the results and the conclusion, respectively.II. THEORY AND METHODSA. Projection of Image FeaturesSearching for the OD location in a 2-D space (image space) renders any localization algorithm highly expensive in terms of computation time. The main objective of this work is to propose a localization algorithm with signicantly enhanced speed by converting the typical 2-D localization problem into two 1-D localization problems. This reduction of dimensionality is achieved by projecting certain features from the retinal image onto two orthogonal axes (horizontal and vertical).The resulting two 1-D signals are then searched to determine the horizontal and vertical coordinates of the OD location. The key factor needed for the success of the dimensionality reduction process is to obtain two meaningful 1-D signals that can be used to determine the coordinates of the OD location. In order to produce such 1-D signals, the set of retinal image features to be projected on either axis should be carefully determined.Two feature maps are used to produce the two 1-Dprojection signals. The rst map feature_map_1 is constructed by calculating the difference between the vertical and horizontal edge maps of the retinal image and dividing the result by the intensity map, see Fig. 1. The second map feature_map_2 is constructed by calculating the summation of the vertical and horizontal edge maps and multiplying the result by the intensity map, see Fig. 2. These maps can be used to accurately localize the OD based upon the simple observation that the central retinal artery and vein emerge from the OD mainly in the vertical direction and then progressively branch into the main horizontal vessels, see Fig. 3(a). This vasculature structure of the retina suggests that a vertical window (with height equal to the image height and a proper width) would always be dominated by vertical edges and dark pixels (vertical vessels) when centered at the OD, see location 1 in Fig. 1. Although the window may contain vertical edges at other locations, i.e., corresponding to small vascular branches or lesions, it will always be populated by strong horizontal edges as well, i.e., the edges of the two main horizontal branches of the retinal vessels, see location 2 in Fig. 1. The vasculature structure suggests also that a horizontal window (with height and width equal to the OD diameter) centered at the determined horizontal location would always enclose a large number of strong edges and a large number of bright pixels when centered at the OD, see Fig. 2. This follows the fact that the possibility of having lesions in the regions above or below the OD is very small, because no retinal vessels are present in these regions. It is worth noting that simple gradient operators (the kernel 1 0 1 and its transpose) are used to produce the vertical and horizontal edge maps of the image. Fig. 2. Feature_map_2 showing a vertically sliding window centered at the OD, the sliding direction and the projection direction.Fig. 3. (a) Example of a retinal image. (b) Plot of the 1-D signal resulting from projecting feature_map_2 onto the vertical axis. (c) Plot of the 1-D signal resulting from projecting feature_map_1 onto the horizontal axis.B. OD LocalizationIn order to localize the OD, the process is split into two steps. In the rst step, the image features are projected onto the horizontal axis to determine the horizontal location of the OD. In the second step, the horizontal location, determined from step 1, is searched for the correct vertical location of the OD. The following two sections show these two steps in detail.It is worth noting that the areas outside the camera aperture (circular region) are excluded using a binary mask generated by thresholding the red component of the image based upon the method described in 3. Fig. 3(c) shows the 1-D signal resulting from projecting the features encoded in feature_map_1 on the horizontal axis. The value of the signal at each horizontal location is the summation of the values of feature_map_1 inside the vertical window, shown in Fig. 2, when centered at this horizontal location. Notice that the horizontal location of the optic disc is easily identied as the location of the maximum peak of the resulting 1-D signal.Fig. 4. (a) Vertical localization signals corresponding to Peak 1. (b) Retinal image showing the two candidate OD locations. (c) Vertical localization signals corresponding to Peak 2. (d) Horizontal localization signal.It is worth Noting that the vessels thickness and the OD diameter are calculated automatically from the image resolution; assuming that the average OD diameter in adults is 1.5mm and the main vessels thickness is 15% of the OD diameter 15.Assuming that the horizontal location of the OD is successfully identied, the objective now is to determine the vertical location of the OD. This is done by projecting the features encoded in feature_map_2 on the vertical axis, as shown in Fig. 3(b). The value of the signal at each vertical location is the summation of the values of feature_map_2 inside the horizontal window, shown in Fig. 2, when centered at this vertical location. Notice that the vertical location of the optic disc is easily identied as the location of the maximum peak of the resulting 1-D signal.C. Improving the RobustnessConsider the horizontal signal of the image shown in Fig. 4(b). It can be shown that the true peak corresponding to the OD horizontal location, peak 2 in Fig. 4(d), is not the maximum peak. This is due to the image artifact that appears as a bright spot to the left of the image. If we follow the algorithm described previously, the estimated OD location will be at a point that, by intuition, cannot belong to an optic disc. On the other hand, if the second peak of the horizontal signal is considered a candidate horizontal location for the OD, the estimated OD location will correspond to the correct location of the OD.This observation can be used to enhance the success rate of the technique. That is, instead of considering the maximum peak of the horizontal signal only, a candidate list of possible horizontal OD locations is used. The candidate list contains the locations of the maximum peaks and the vertical localization step is repeated for each candidate horizontal location. This results in possible (2-D) candidate locations of the OD. In order to determine the nal location, a set of image features is used to score each candidate. Then, the nal location of the OD is taken as the candidate with the maximum scoring index.In this work, the candidate list contains two locations. The scoring index is calculated as the peak strength of the horizontal signal at the candidate location multiplied by a weighing factor. The weighting factor incorporates some a priori knowledge of the typical geometric and appearance properties of the OD.To calculate the weighting factor, a square window (with edge equal to twice the OD diameter) is centered at the candidate OD location. Then, 10% of the brightest pixels within this window are segmented. If an object (large cluster of bright pixels) exists at the candidate location, the eccentricity, dened as the ratio of the objects minor axis length to the objects major axis length 16, of this object is calculated. If no object exists, the eccentricity of the candidate location is set to a very small value (e.g., 0.1). Then, the weighting factor of this location is set equal to the calculated eccentricity.D. AlgorithmStep 1) Get image features Construct feature_map_1 and feature_map_2:feature_map_1feature_map_2 Where, and are the absolute vertical & horizontal edge images, respectively; and is the image intensity.Step 2) Project feature_map_1 on the horizontal axis1) Define as a rectangular window of size (image height, 2main vessel width) centered at the horizontal location x2) Slide over feature_map_1 and calculate:sum of feature_map_1 values inside 3) The horizontal location of the OD, , is the location of the maximum peak ofStep 3) Project feature_map_2 on the vertical axis1) Define as a rectangular window of size (OD diameter, OD diameter) centered at the horizontal location and the vertical location y2) Slide over feature_map_2 and calculate:sum of feature_map_2 inside 3) The vertical location of the OD, , is the location of the maximum peak of Step 4) Improving the robustness1) For each candidate (,), select a square region of interest ( ROI ) with edge size of 2OD diameter2) Segment the brightest 10% pixels within each ROI3) Group neighboring pixels into objects4) Calculate the eccentricity of the largest object:eccen (,) = ( minor axis length / major axis length )5) Calculate the Scoring Index ( SI ) of each candidate: SI (,)=6) Select the final OD location as the location with the largest SIIII. RESULTSFour publicly available databases are used to evaluate the accuracy and the computation time of the proposed technique. The four databases are: 1) STARE database (81 images, 605 X700 pixels) 17, 2) DRIVE database (40 images, 565 X 584 pixels) 18,3) Standard Diabetic Retinopathy Database “Calibration Level 0” (DIARETDB0) (130 images, 1500, X1152 pixels) 19. 4) Standard Diabetic Retinopathy Database “Calibration Level 1” (DIARETDB1) (89 images, 1500 X1152 pixels) 19. The diseased images in the four databases contain signs of DR, such as hard exudates, soft exudates, hemorrhages, and neo-vascularization (NV). The accuracy and computation time results of evaluating the proposed method using these databases are summarized in Table I. The table includes the results of applying the method without constructing the candidate list (the maximum peak of the horizontal localization signal is selected as the correct location) and also the results with the list containing two candidates.Fig. 5. Selected OD localization results. (a)(l) Show successful OD localization examples. (m)(o) Show examples of failure in OD localization. The white “X” indicates the location of the OD as detected by the proposed method.TABLE IRESULTS FOR EVALUATING THE PERFORMANCE OF THE PROPOSED METHODINTERMS OF SUCCESS RATE AND COMPUTATION TIME USING FOUR DATABASESDatabaseSTAREDRIVEDIARETDB0DIARETDB1TotalNormal Images313320589Diseased Images50711084251Number of Images814013089340Success Rate ( % )89.010094.696.694.4Success Rate ( % )( with improvement)92.610098.597.897.0Computation Time( seconds )0.460.320.980.98The proposed method was implemented using Matlab (The MathWorks, Inc.) and the results shown in Table I are acquired by running the developed code on a PC (2.66 Intel Core 2 Due and 4 GB RAM). The detected location of the OD is considered correct if it falls within 60 p