Microaneurysm and Hemorrhage Detection in Retinal Images

Abstract

Incidences of diabetes are increasing worldwide. Inefficiency of human newlinebody to produce and consume insulin leads to diabetes. Diabetes over newlinea period of time starts showing adverse effects on different organs. If it newlineaffects eye it is termed as diabetic retinopathy. newlineIf left untreated, diabetic retinopathy can cause vision loss. Red lesions newlineare the foremost clinically observable signs of DR. Their early detection newlinecan help ophthalmologists in treating abnormalities efficiently newlineand limit the disease severity. So the detection of red lesions at an early newlinestage has become an indispensable task today. In this thesis, we give newlinean overview of earlier proposed algorithms and methods by comparing newlinethese algorithms based on their performance, for supporting the researchers newlineby providing the gist of these algorithms. The standard retinal newlineimage databases are also compared and discussed. newlineMicroaneurysm (MA) detection is accomplished with three proposed newlinealgorithms: First, we propose a rule based system for microaneurysms newlinedetection in digital fundus images. The proposed system is a three step newlinealgorithm. Initially image pre-processing is done, secondly candidate newlinemicroaneurysms are segmented, thirdly features are extracted from these newlinecandidates and true microaneurysms are recognized using rule based newlineexpert system. The system performance is evaluated on publicly available newlinedatabase DIARETDB1, which consists of evaluation protocol and newlineground truth collected from experts. newlineSecond, we propose a novel method which is simple, efficient and newlinereal-time for segmenting and detectingMAin color fundus images (CFI). newlineA novel set of features based on statistics of geometrical properties of newlineconnected regions, that can easily discriminate lesion and non-lesion newlinepixels are used. For large scale evaluation proposed method is validated newlineon DIARETDB1, ROC, STARE and MESSIDOR dataset. newlineThird, we propose a novel two-level method: coarse level use morphology newlinefor MA candidate detection and fine level use classification and newlineregression tree(CART) for true MA classificat