A deep lstm neural network based rfo algorithm for early detection of diabetic retinopathy

Abstract

Diabetes has become a widespread disease found in adults over forty newlineyears old. the prolonged persistence of diabetes causes health issues such as newlinediabetic retinopathy-vision loss, diabetic neuropathy-nerve impairment, newlinediabetic nephropathy-kidney malfunctioning. among these, dr is drastically newlineincreasing in asian countries and makes many people blind. simultaneously newlinepatient ophthalmologist ratio is approximately 1:100000. dr is a chronic newlinecomplication that damages the retina. fifty percent of people become blind newlinebefore diagnosis. the ophthalmologist may use three physical tests to identify newlinethe severity of dr. they are the visual acuity test, pupil dilation test, and newlineoptical coherence tomography. thus, an automatic diagnosis system is needed newlineto avoid severe lesions in the retina and late treatment. nowadays, many deep learning algorithms are in application to recognize objects or patterns. any classification algorithm that classifies the newlinelesion based on a severity scale is need of the hour. convolutional neural newlinenetworks do not have a memory cell to store previous data with a timestamp, newlinewhereas recurrent neural networks have the same. the proposed work newlinedeveloped an algorithm that combines classification and optimization newlinetechniques to perform well and help the eye doctor diagnose dr earlier. newlinefundus dr image is selected as the input image. learning networks newlinecannot train the fundus images as it is. thus, retinal fundus images come across newlinefour stages: preprocessing, segmentation, feature extraction, and combined newlineoptimization and classification. in preprocessing rgb image is converted into newlinea greyscale image. then contrast is improved using histogram equalization newline(he) and contrast limited adaptive histogram equalization (clahe). newlinefinally, regression is done with a generalized linear model (glm). the image newlinesize is reduced to manage the memory needed to store the data. in the second newlinestage, an adaptive watershed segmentation algorithm is used for segmenting the newlineoptic disc. here the region of interest is the optic disc. when eye