A RULE - BASED SYSTEM FOR PROCESSING RETINAL IMAGES

Mohamad  J.A. ZEMERLY

This thesis reports work carried out in connection with the use of computer vision for analyzing and processing of retinal images of persons suffering from diabetic retinopathy.

The aim of the project is twofold. The first concerns the segmentation and interpretation of retinal  images to quantify lesions associated with the disease. These lesions are haemorrhages, soft exudates, and hard exudates. Another lesion of interest, which also has to be quantified, is the optic disc. The importance of detecting the optic disc will be discussed later. For this purpose, a rule‑based system for segmentation and interpretation was implemented and applied to 8 digitized retinal images. In order to assess the results of this system a simple evaluation technique based on comparison between a manually segmented image by the aid of a digitizing tablet and the output of the system was applied to each image tested. The rule‑based system and the results obtained are presented. These results show a major improvement on previous work carried out in this area. The average percentage total error was found to be 22.88%. Suggestions for further improvement of the results are made.

The second concerns the development of two image compression methods using Tchebyshev polynomials to allow more efficient storage of digitized images. The methods proposed were applied to 15 retinal images and 5 images of diversity of subjects (2 natural scenes, a typeset text, a finger print, and human face). The compression algorithms and the results obtained are described. The results of these methods were superior to a previously reported method. A reduction in the amount of data of 68% and 74% on average for the two methods respectively was achieved, while resulting in no noticeable distortion.