Automatic detection ofdiabetic retinopathy using anartificial neural network: ascreening tool

Aits-Todetermine ifneuralnetworks candetectdiabetic features infundus imagesandcomparethenetwork against anophthalmologist screening asetoffundusimages. Methods-147diabetic and 32 normal images werecaptured fromafundus camera,storedon computer, andanalysed usingabackpropagation neuralnetwork. Thenetwork wastrained torecognise featuresintheretinal image. Theeffects of digital filtering techniques anddifferent network variables wereassessed. 200diabeticand101normalimageswerethen randomisedand usedto evaluate the network's performance forthedetection ofdiabetic retinopathy against an ophthalmologist. Results-Detection ratesfortherecognitionofvessels, exudates, andhaemorrhageswere91.7%,93.1%,and 73.8% respectively. When comparedwiththe results oftheophthalmologist, thenetworkachieved asensitivity of88.4%anda specificity of83.5%forthedetection of diabetic retinopathy. Conclusions-Detection ofvessels, exudates, andhaemorrhages was possible, withsuccess ratesdependent uponpreprocessing andthenumberofimages used intraining. Whencomparedwiththeophthalmologist, thenetworkachieved good accuracy forthedetection ofdiabetic retinopathy. Thesystemcouldbeusedas anaidtothescreening ofdiabetic patients forretinopathy. (BrJOphthalmol 1996;80:940-944)