Article Title

Diyabetik retinopati tespitinde yeni bir algoritma kullanılarak optik disk yerinin kestirimi


In most of the automated retinal image analysis systems Optic Disc(OD) localization is a main step. The position and region of OD is significantly important in terms of a few points. First of all, the location of macula can be detected using the location of OD. Since the fact that the blood vessels originate from the OD region in the embryonic period, the OD location is also used as the seed point for vessel extraction algorithms. More and more, the color range of the exudates, especially the hard ones, is so similar to the color range of the OD. Thus, the localization of the OD is a main step to be able to differentiate between the exudates and OD in this yellowish color range. In the color fundus images the OD can be observed as circle or ellipse like yellowish region where red blood vessels and optic nerves originate from inside of it. The other name of the OD is blind spot because of the fact that it contains no photoreceptor. In a normal fundus image the diameter of the OD is in the range of 80 and 100 pixels. The main disadvantages of OD localization are the inhomogeneous light distribution of the light in the color fundus images and the fact that the red blood vessels locate on this yellow region as well as extremely irregular OD shapes (Kaur and Sinha, 2012). The DRIVE image database has been used for the evaluation of the implemented algorithms. In order to get rid of the inhomogeneous light distribution over the fundus images, the images are converted from Red Green Blue (RGB) color space to Hue Saturation Intensity (HSI) color space and then the intensity channel has been equalised using Contrast Limited Adaptive Histogram Equalization (CLAHE). After CLAHE algortihm has been applied the HIS color space has been converted back to the RGB color space. This new RGB image has been converted to Grayscale format. The Grayscale image has been applied morphological closing operation with a disk structuring element of diameter 10. Afterwards, the Canny Edge Detection (CED) algorithm has been applied to the closed image with a threshold of 0.1 and the resulted edges has been applied Morphological Closing Operation (MCO) with a disk structruing element of a diameter value within 3 and 10. Finally, the Circular Hough Transform (CHT) algorithm has been applied over these edges and all circular patterns as an OD candidate has been localised. Two problem specific features are extracted for each circle to be tested whether it is in OD area or not during the classification phase. One of these features is the multiplication of two extracted features. In order to get these extracted features, a threshold representing the yellowish region in green channel histogram is iteratively calculated by a novel algorithm. The first extracted feature is the ratio of the region whose pixel values are above this yellowish threshold to the whole masked region. The second extracted feature is the count of the pixels whose pixel values are above this yellowish threshold. The first feature is calculated as the multiply of these two extracted features. The other feature is a flag which is set as true only for the circle which has the maximum value of the first feature. Each detected circle has been classified by applyinng its features to a Multi Layer Perceptron (MLP). The success ratio is 95.00 % for 20 training images and 20 testing images. This is a novel method for OD localization without contour detection which may be a basic step for the other retinal lesion detection systems.