RETINAL IMAGE ANALYSIS
This project is conducted as a summer project for the MSc Computer Science, University of Birmingham during 2015-2016. Supervisor of this project is Prof. Ela Claridge. This project is about processing Retina Images, which are photos taken from the back of the eye. The aim of this research is to extract the vein shape of these images and correlate the thickness statistics with various diseases. The basic process includes the following steps: segmentation, classification, and export of statistics.
During segmentation, we try to threshold the image by hand and by Otsu thresholding. We evaluate the results by making use of Jaccard Index. Following, we apply oriented Difference of Gaussian (DoG) filters to get better results of extracting the veins. Then, we reduce noise in the image. At the next step, we classify the veins into classes according to their thickness, with Digital Sieve and we measure their length of each class. At the final step, we apply this process to STARE database and get interesting results about the statistics of �diseases.
The results we get with thresholding by hand is 11.30% and for Otsu thresholding is 31.65%. So we choose Otsu but it is not sufficient enough. We improve the success by using DoG and we have an average of 47.78% which is 50.94% better than OtsuBlurred images or lesions on retina images do not let us better results.
Moreover, the results from STARE database include the classification of the veins. For all the images in STARE database, we get 69.7% for the thinnest veins of 1-4 pixels width, 24.5% for medium veins of 5-8 pixels width and 5.8% for thicker veins of 9 pixels to the maximum with.
