Abstract
A retinal image has blood vessels, optic disc, fovea, and so forth as the main components of an image. Segmentation of these components has been investigated extensively. Principal component analysis (PCA) is one of the techniques that have been applied to segment the optic disc, but only a limited work has been reported. To our knowledge, fovea segmentation problem has not been reported in the literature using PCA. In this paper, we are presenting the segmentation of optic disc and fovea using PCA. The PCA was trained on optic discs and foveae using ten retinal images and then applied on seventy retinal images with a success rate of 97% in case of optic discs and 94.3% in case of fovea. Conventional algorithms feed one patch at a time from a test retinal image, and the next patch separated by one pixel part is fed. This process is continued till the full image area is covered. This is time consuming. We are suggesting techniques to cut down the processing time with the help of binary vessel tree of a given test image. Results are presented to validate our idea.
Highlights
This paper presents an extension of the application of principal component analysis (PCA) to retinal images
The application of PCA to determine the location of the fovea is a relatively difficult problem as compared with locating the optic disc because the fovea usually has lower contrast compared with the optic disc in a retinal image
When 70 retinal images were presented to PCA for evaluation, 68 of the images met the criterion of 10% accuracy leading to an efficiency of 97%, whereas in case of fovea localization, 66 images met the criterion of 14.2% accuracy leading to an efficiency of 94.3%
Summary
This paper presents an extension of the application of principal component analysis (PCA) to retinal images. Application of PCA is presented for two different cases: (1) to automatically locate the position of the optic disc in a retinal image, and (2) to automatically locate the position of the fovea in a retinal image. To our knowledge, the latter application is novel and has not been reported in the literature. For any number of unknown image patches from a retinal image each having dimension (1, n2) to resemble to one of the training images, the error as defined below must have a minimum value for the most similar patch: error = sum1 − sum2,. The algorithm that we used in localization of optic disc and fovea follows the steps explained by the mathematical modeling
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call