Development and Performance Analysis of a Novel Algorithm Feature Extarction for Detection of Glaucoma from Fundus Images

Abstract

newline At present, the grading of retinal diseases caused by diabetes is carried out manually, in a newlinesemi automated environment by the ophthalmologist. The engineering research work is newlineprogressing by considering single aspect such as either pre-processing, feature extraction or newlineclassification. The proposed work aims at providing a complete solution for grading of newlinediabetic glaucoma. To accomplish this task the work targets each phase exclusively. Firstly, newlinein the pre-processing phase a novel approach called Reaction Diffusion Level Set newlineSegmentation for effective segmentation of optic disc and optic cup is been designed and newlineimplemented on 175 retinal images. The technique marked the optic disc region with an newlineaverage accuracy of 98.95% over 87.34% of previous work. Secondly, a 3and#120590; technique in newlinecombination with active contour is implemented to mark the optic disc and optic cup newlineboundaries. The technique first extracts the region of interest containing optic disc using newlineConnected Component Analysis. The blood vessel network is eliminated using newlinemorphological operations in order to have a break free continuous optic disc boundary. newlineFinally 3and#120590; , a statistical measure is used as threshold to check the intensity of the pixel under newlinetest to belong to either optic disc or optic cup. The results obtained are encouraging and have newlineyielded approximately 100% results with the fundus images containing complete optic disc. newlineTo achieve robust detection of optic disc and optic cup, a novel frame work is proposed. The newlineproposed architecture embeds stacked Auto Encoder with deep learning technique to extract newlineoptimal features for grading glaucoma. The proposed technique consists of three newlineconvolutional layers, one max pool layer and two fully connected layer. The convolution newlinelayer works towards feature extraction and learning. Max pool layer optimizes the features, newlineand two Fully Connected layers work towards classifying the input image as glaucomatous newlineor non glaucomatous image. The technique has yielded the result