الفهرس | Only 14 pages are availabe for public view |
Abstract Optic disc (OD) detection and segmentation are important steps in many computer-aided diagnosis systems. It is a prerequisite for effective segmentation of many other anatomical features in retinal images. Ocular fundus images, which are acquired using a fundus camera, are used in this study. In this thesis, two methods are proposed for segmentation of the OD from color retinal images in addition to a pre-processing step based on the Gaussian filters. In the pre-processing stage, a method for automatic binary mask generation is proposed. Intensities from red and green channels are combined to obtain a modified intensity channel based on image histogram and P-tile threshold values. This adaptive method allows different percentages to be taken from red or green channels based on each channel histogram. This results in a visually enhanced version of the intensity image that will help in the OD segmentation task. In the OD segmentation stage, two new automatic methods for OD segmentation from digital color fundus images are proposed. One used the multi-scale Laplacian of Gaussian (LOG) filters. The LOG filter is applied at different scale values related to ~O’s radius, followed by local thresholding and then selecting objects with the largest circularity. The performance has been evaluated using 1441 images from four publicly available datasets. Experimental results show a 100% success rate for both DRIVE and ARIA datasets. For the STARE and MESSIDOR, success rate of 98.8% and 99.83% are achieved respectively which are better that the state-of-the-art methods. The second proposed method applied the Difference of Gaussian (DOG) filter by using two values of sigma related to OD’s radius, then thresholding, and selecting regions with circular objects. The performance has been evaluated on 1660 images representing six publicly available datasets; STARE, DRIVE, ARIA, DIARETDB1, DIARETDBO, and MESSIDOR datasets. Experimental results show that a 100% success rate for images from DRIVE, ARIA, DIARETDB1, and DIARETDBO datasets ; which is better than the of state-of-the-art methods with accuracy less than 99% for ARIA, DIARETDB1, and DIARETDBO datasets. While achieving 98.8% and 99.83% for STARE and MESSIDOR data sets respectively. The algorithms run with an average computational time of 1.9 and 1.2 seconds for first and second algorithm respectively. Keywords: Binary mask generation, intensity component, Optic disc segmentation, Fundus images. |