Using Genetic Algorithm for Identification of Diabetic Retinal Exudates in Digital Color Images ()
1. Introduction
The eye is an organ associated with vision. The eye is housed in a socket of bone called the orbit and is protected from the external air by the eyelids [1]. A cross section of the eye is shown in Figure 1. Light entering the eye through the pupil is focused on the retina. The retina is a multi-layered sensory tissue that lines the back of the eye. It contains millions of photoreceptors that capture light rays and convert them into electrical impulses [2]. These impulses travel along the optic nerve to the brain where they are turned into images. In a normal FI, the optic disk is brighter than any part of the retina and is normally circular in shape. It is also the entry and exit point for nerves entering and leaving the retina to and from the brain. A typical retina fundus image looks like the one shown in Figure 2. The bright optic disc and the vascular network can clearly be seen in the image.
Neovascularization can be described as the abnormal growth of blood vessels in some areas of the eye including the retina [3,4]. This occurs in response to ischemia, or diminished blood flow to ocular tissues. These new blood vessels have weaker walls and can rupture which subsequently leads to bleeding or causes scar tissue growth. An increase in the number of such veins can lead to lifting the retina away from the back of the eye. When the retina is lifted away, this is called retinal detachment. If left untreated, this can cause severe vision loss, including blindness [1]. Neovascularization and bleeding in FIs are associated with bifurcation point (BP) and crossover point (CP) changes [3-5]. We proposed a new technique which is named the Genetic Algorithm and Vertex Chain Code to detected cross-point number method (GAVCC) that can be used for vascular bifurcation and intersection detection. The fundus image undergoes preprocessing and segmentation before the GAVCC technique is applied during the last stage which is named the bifurcation and intersection detection stage. The preprocessing involves color space conversion, noise removal, and illumination equalization.
Figure1. Cross sectional diagram of human eye [1].
The segmentation stage separates the background from the vascular network using the skeletons algorithm. During the bifurcation and intersection detection stage, as the name suggests, vascular bifurcation and crossover points are detected by applying the GAVCC technique. In addition, we compared this method with others techniques and analyzed by applying them to more images taken from two FI databases. The results of all the techniques have been analyzed and compared with each other.
2. Related Work
The eye is a window to the retinal vascular system which is uniquely accessible for the non-invasive, in vivo study of a continuous vascular bed in humans. The detection and measurement of blood vessels can be used to quantify the severity of disease, as part of the process of automated diagnosis of disease or in the assessment of the effect of therapy. Retinal blood vessels have been shown to change in diameter, branching angles or tortuosity, as a result of a disease, such as hypertension [6], diabetes mellitus or retinopathy of prematurity (ROP) [7].
Furthermore, retinal arteriolar or venular changes predict development of hypertension [6,8], new onset diabetes [6], progression to diabetic retinopathy [9] and development of diabetic renal disease [10]. Thus a reliable method of vessel segmentation would be valuable for the early detection and characterization of morphological changes.
A first approximation to the segmentation of retinal blood vessel using this approach was previously presented [11], where segmentation method was tested on a small image sample without any validation. An extension of this work is presented here and the method is now tested on two local databases and two public databases of complete manually labeled images [12,13]. We evaluate our segmentation using the public databases that have been also used by other authors for the same purpose [13,14]. Validation of segmented vessel diameters and branching angles measurements are also made: between red-free against fluoresce in images and between our algorithm and one of the public databases.
This paper is organized as follows: Section 2 presents some of the reported work in relation to this research; Sections 3 and 4 present a vertex chain code method to extraction the features of retina; In Section 5 applied genetic Algorithm to archival the accuracy and all the stages that an image has to go through to allow bifurcation and intersection detection are discussed. The proposed technique for bifurcation and crossover detection is introduced briefly; and Section 6 presents results obtained from experiments on FIs.
3. Proposed Method
A three-stage framework is used for detecting BPs and CPs from FIs. The stages involved are: preprocessing followed by segmentation, and finally detection of bifurcations and crossovers. The fundus images which have been collected for different hypertension and DR patients are not of good quality and are badly or unevenly illuminated. The main reasons why these images are of bad quality are poor lighting conditions and motion of the eye or the camera during photography. Also, the fundus images may be of different sizes, color resolutions (gray scale as well as color images) and different formats, etc.
3.1. Fundus Image Enhancement
The quality of the blood vessels structures in a retinal image is an important characteristic, as the ridges carry the information of characteristic features required for minutiae extraction.
The morphology of the retinal blood vessels can be an important indicator for diseases like diabetes, hypertension and retinopathy of prematurely (ROP). Thus, the measurement of changes in morphology of arterioles and venules can be of diagnostic value. Here we present a method to automatically segment retinal blood vessels based upon multiscale feature extraction. This method overcomes the problem of variations in contrast inherent in these images of the magnitude of the gradient and the maximum by using the first and second spatial derivatives of the intensity image that gives information about vessel topology.
This approach also enables the detection of blood vessels of different widths, lengths and orientations. The local maxima over scales principal curvature of the Hessian tensor are used in a multiple pass region growing procedure. The growth progressively segments the blood vessels using feature information together with spatial information. The algorithm is tested on red-free and fluoresces in retinal images, taken from two local and two public databases.
3.2. Segmentation
The first step of the retinal blood vessels enhancement algorithm is image segmentation. Segmentation is the process of separating the foreground regions in the image from the background regions. The foreground regions correspond to the clear retina area containing the vessels, which is the area of interest. The background corresponds to the regions outside the borders of the retina area, which do not contain any valid information. When minutiae extraction algorithms are applied to the background regions of an image, it results in the extraction of noisy and false minutiae. Thus, segmentation is employed to discard these background regions, which facilitates the reliable extraction of minutiae.
In a retina image, the background regions generally exhibit a very low grey-scale variance value, whereas the foreground regions have a very high variance. Hence, a method Directional Fourier Filtering (DFF) based on variance threshold can be used to perform the segmentation. Firstly, the image is divided into blocks and the grey-scale variance is calculated for each block in the image. If the variance is less than the global threshold, then the block is assigned to be a background region; otherwise, it is assigned to be part of the foreground. The grey-level variance for a block of size WxW is defined as:
where V(k) is the variance for block k, I(i, j) is the grey-level value at pixel (i, j), and M(k) is the mean grey-level value for the block k.
3.3. Normalization
The next step in the retinal blood vessels enhancement process is image normalization. Normalization is used to standardize the intensity values in an image by adjusting the range of grey-level values so that it lies within a desired range of values. Let I(i, j) represent the grey-level value at pixel (i, j), and N(i, j) represent the normalized grey-level value at pixel (i, j). The normalized image is defined as:
where M and V are the estimated mean and variance of I(i, j) respectively, and M0 and V0 are the desired mean and variance values, respectively. Normalization does not change the vessels structures in retinal; it is performed to standardize the dynamic levels of variation in grey-level values, which facilitates the processing of subsequent image enhancement stages.
3.4. Binarization
Most minutiae extraction algorithms operate on binary images where there are only two levels of interest: the black pixels that represent ridges, and the white pixels that represent valleys. Binarization is the process that converts a grey level image into a binary image. This improves the contrast between the ridges and valleys in a fingerprint image, and consequently facilitates the extraction of minutiae.
3.5. Thinning
The final image enhancement step typically performed prior to minutiae extraction is thinning. Thinning is a morphological operation that successively erodes away the foreground pixels until they are one pixel wide. A standard thinning algorithm [15] is employed, which performs the thinning operation using two sub iterations.
This algorithm is accessible in MATLAB via the “thin” operation under the bimorph function. Each subiteration begins by examining the neighborhood of each pixel in the binary image, and based on a particular set of pixel-deletion criteria, it checks whether the pixel can be deleted or not. These subiterations continue until no more pixels can be deleted. The application of the thinning algorithm to a retinal image preserves the connectivity of the ridge structures while forming a skeletonised version of the binary image. This skeleton image is then used in the subsequent extraction of minutiae.
The skeleton of the vascular tree is obtained from the segmented binary image by a thinning process where pixels are eliminated from the boundaries toward the center without destroying connectivity in an eight connected scheme [16]. A pruning process is applied to eliminate short, false spurs, due to small undulations in the vessel boundary. False spurs are deleted if they are smaller or equal to the largest vessel diameter expected in a particular image. Figure 3 shows the stages of the segmentation method. The extraction of minutiae from a skeleton image will be discussed in the next section.
4. Feature Extraction Using VCC Contours
The most commonly employed method of feature extraction is the Crossing Number (CN), with comparison with other methods which based on Fourier transform of retinal images and Two-dimensional Wavelet Transform [17] more efficient. The concept this method involves the use of the skeleton image where the ridge flow pattern is eight-connected.
The landmarks are extracted by scanning the local neighborhood of each ridge pixel in the image using a 3 × 3 window. The CN value is then computed, which is defined as half the sum of the differences between pairs of adjacent pixels in the eight-neighborhoods. Using the
Figure 3. The segmentation method and enhancement (a) Original image; (b) Enhanced image; (c) Binary image; (d) Thinned and separated vassals image.
properties of the CN, shown in Table 1, the ridge pixel can then be classified as a ridge ending, bifurcation or non-minutiae point. For example, a vascular pixel with a CN of one corresponds to a ridge ending, and a CN of three corresponds to a bifurcation.
We propose a new feature extraction algorithm based on chain code contour following. Chain codes have been used in computer vision to describe the shapes of object boundaries (see Figure 4). Chain codes are translation invariant and can be made rotation invariant if we use relative directions. This is one of the primary reasons why it is used for object recognition applications in machine vision. However, we are interested in this representation since it provides us with a loss-less representation of ridge contours at the same time yielding a wide range of information about the contour such as curvature, direction, length etc. [18]. The chain code representation has several advantages mentioned below.
(1) By directly processing the contour we obviate the need for thinning the binary image. This reduces the computational complexity of the feature extraction process.
(2) Several post processing steps previously used to eliminate spurious features can now be integrated into the contour processing stage. For instance, holes within the ridges can be filled by considering only exterior contours. Also small artifacts can be easily removed by neglecting contours whose length is beneath a pre-defined threshold.
(3) Detecting minutiae becomes very straightforward as will be seen in the following sections.
Most retina minutia extraction methods are thinningbased where the skeletonization process converts each ridge to one pixel wide. Minutia points are detected by locating the end points and bifurcation points on the thinned ridge skeleton based on the number of neighboring pixels. The end points are selected if they have a single neighbor and the bifurcation points are selected if they have more than two neighbors. However, methods based on thinning are sensitive to noise and the skeleton structure does not conform to intuitive expectation. Our chain code based method is obtained by scanning the image from top to bottom and right to left. The transitions from white (background) to black (foreground) are detected.
The contour is then traced counterclockwise and expressed as an array of contour elements. Each contour element represents a pixel on the contour. It contains fields for the x, y coordinates of the pixel, the slope or direction of the contour into the pixel, and auxiliary
Table 1. Properties of the crossing number.
Figure 4. Chain code contour representation: (a) An example of an object boundary; (b) Direction convention used for chain code representation; (c) Resulting chain code representation.
information such as curvature. In a binary fingerprint image, ridge lines are more than one pixel wide. Tracing a ridge line along its boundary in counterclockwise direction, a termination minutia (ridge ending) is detected when the trace makes a significant left turn. Similarly, a bifurcation minutia (a fork) is detected when the trace makes a significant right turn (Figure 5(a)).
Figure 6(a) shows the candidate points marked with circles over the skeleton and Figure 6(b) summarizes the three possible cases; where two intersections with the window are spurs, three intersections are bifurcation points and four intersections are crossing points.
Once all of the landmarks are identified, bifurcation points are labeled as –r, where r is the radius of the maximum circle centered on that point that fits inside the boundary of the bifurcation Figure 7. The sign is used to distinguish between the radius and chain code numbers that are described in the following section. For crossing points, the skeleton is corrected and marked to define a single crossing point see the last row of Figure 6(b). Figure 7 shows the corrected skeleton with all of the landmarks marked. Until this stage the process is fully automatic. The process fails when:
(1) Two true bifurcation points are very close and are merged into a crossing point.
(2) Or when the two candidate bifurcation points fall outside the circular window and are, thus, defined as two bifurcation points.
An example is shown in the top-left corner in Figure 7(a) where two branching points are found, an inspection of the full image shows that it should be a crossing point. These cases must be corrected manually the complete image is labeled in this way and it normally contains several independent vascular trees.
Figure 5. Chain code contour representation: (a) The contour is extracted by tracing the ridge boundaries in a counter clockwise direction. Each contour element along with its co-ordinates, curvature slope and other information is stored in a list; (b) The slope convention.
Figure 6. (a) Candidate points marked with circles over the skeleton; (b) Types of landmarks.
Figure 7. (a) Corrected skeleton with landmarks marked: Bifurcation points are marked with the maximum circle that fits inside the vessel boundaries, terminal points with triangle, crossing points with crosses, and the root point with a black circle; (b) The area A is the number of vessel pixels bounded by the two lines tangents to the circles and perpendicular to the skeleton and La is the true length of the of the skeleton.
Figure 8(a) shows the chain code directions in a 3 × 3 neighborhood. When the first bifurcation point is reached the chain of the current branch is ended and the coordinates of the starting points of the two daughters are found and saved. In order to keep the network information about the relationships between branches along the tree, a binary ordering scheme is adopted by labeling the root as one and, thereafter, the daughters of parent k0 are labeled with key numbers