Detection and Quantification of Drusen in Retinal Fundus Images

Abstract

Drusen are common features of the aging macula, caused by accumulation of extra cellular materials and fatty deposits beneath the retinal surface, visible in retinal fundus images as yellow-white spots. In the ophthalmologists’ opinion, the evaluation of total number and area of drusen in a sequence of images taken during a treatment will help to understand the disease progression and the extent of its effect on the eye. However, this evaluation is fastidious, tiresome and difficult to reproduce when performed manually. A literature review on automated drusen detection showed that the works already published were limited to drusen segmentation and quantification using limited information and also few of them provided grading of Age-related macular degeneration (AMD). The purpose of this work is to propose two automated method to quantify and grade the severity of AMD using advanced digital image processing techniques. Method 1 presents an automated technique for segmentation and quantitative analysis of drusen based on: First, region-based statistical analysis which corrects the non-uniform illumination of background, enhances local intensity, minimizes image noise, segments image through Otsu’s threshold in addition to morphological operation and hence computes area and edge of the detected drusen. Second, pixel-wise feature are examined that helps to extract the features of overlapped components through weighted centroid and standard deviation. It makes counting of number of drusen easy. Method 2 presents an automated method to detect and segment drusen in retinal fundus images using (i) gradient based segmentation to find true edges of drusen, (ii) connected component labeling to remove suspicious pixels from drusen region and (iii) edge linking to connect all labeled pixels into a meaningful boundary. Both methodologies have been applied and tested on two publicly available retinal image datasets i.e. Structured Analysis of retina (STARE) and Automated Retinal Image Analysis (ARIA), acquired with the aid of a digital fundus camera. These methodologies are evaluated to improve the pixel-to-pixel analysis by comparing the segmented results with ground truth and validated by using different statistical analysis. From the results of these studies it can be concluded that the methodologies proposed are capable to automatically measure drusen area, size and number in an accurate and reproducible process. Also, the thesis proposes grading of AMD to evaluate the stages of macular degeneration so that it can be treated at time to stop the progression of AMD.