Automatic Detection Of Neovascularization Based On Color Retinal Image

Diabetic retinopathy,a common diabetic eye complication,is one of the important causes of blindness.Growing Neovascularization is a key indicator for being Proliferative Diabetic Retinopathy from Non-Proliferative.Fortunately,early detection and regular tracking control could help patients protect from severe vision loss or blindness.The retina is a particular place where is able to visually observe the structure of vascular non-invasively.Moreover,vessels is the "barometer" of human somatic function.There is a very far-reaching significance of researching on vascular abnormalities associated with the disease.General automatic detections of retinopathy are concentrated on hard exudates,hemorrhages,microaneurysms and cotton wool spot.However,for neovascularization,where new blood vessels grow due to extensive lack of oxygen in the retinal capillaries,few research has been done to detect it.This thesis presents an automatic framework for neovascularization detection,which including various combination of techniques such as vessel enhancement,morphology operator,vessels extraction,principle component analysis and machine leaning classifiers.In the preprocessing stage,uneven illuminations have been wiped off,new vessels were enhanced and most normal vessels were removed.Then a filter bank designed for distinguishing neovascularization from others are applied to the result of preprocessing,which including Entropy Filter,Differential Invariant Filters,Vessel Density,Anisotropic Gaussian Filters,Corner Filter,Vessel Curvature,etc.All of them are multi-scale,translation invariant and rotation invariant.There were three classifiers applied in this thesis for pixels classification:Support Vector Machine,Extreme Learning Machine and General Regression Neural Network.The automatic detection framework was learned and tested on images of two different public databases:DiaRetDB and MESSIDOR.Results show that this automatic detection framework could detect and mark the neovascularization regions in digital color fundus image with high veracity.The accuracies respectively up to 92.71%and 88.09%while features classified with GRNN and ELM.