Research On Disease Detection In Ophthalmic Medical Images

With the rapid increase of diabetic patients,the prevalence of diabetic retinopathy blinding rate is also increasing year by year.Diabetic nephropathy has now become the first blinding disease in working-age population.There are no obvious symptoms in early stage so it is difficult to find.But it is also difficult to treat in the late stage.Therefore,it is of great significance to detect the lesions as early as possible.Diagnosis of diabetic retinopathy require careful examination by a professional ophthalmologists,usually need one or two days delay.Therefore,the use of image analysis and deep learning technology automatically detect lesions has becoming an effective solution.This paper focuses on the segmentation of diabetic retinopathy in retinal images and using convolutional neural network to identify lesions rapidly and accurately.The main contents are as follows:(1)We proposed a cascade detection method of diabetic retinopathy based on multi-scale region blocks.It is to detect whether there is a lesion area in the retinal image in the first step.And then segment the lesion area pixel by pixel to simultaneously detect micro aneurysms,hemorrhages,hard exudations and soft exudations.High detection rate and faster speed was got in the DIARETDB1 database,which proved the effectiveness of this method.In the detection rate of four lesions,microaneurysms with 88.62%,hemorrhages with 94.91%,hard exudates with 98.91%and soft exudation with 92.91%.(2)Based on the idea of semantic segmentation of natural images,this paper designs a fully convoluted neural network with multi-level features based on the varying thickness,varying sizes and irregular positions of lesions in the retinal image.The validity of the method is demonstrated with private data sets.(3)Construction of diabetic retinopathy labeling system and automatic detection system.The Diabetic Retinopathy Labeling System is designed specifically to get the data needed for diabetic retinopathy segmentation,to help ophthalmologists mark the retinal lesion quickly and accurately.Diabetic Retinopathy Automatic Detection System is based on the multi-level feature fusion algorithm.It can find locations of four kinds of diabetic retinopathy at the same time and assist the ophthalmologists in their clinical diagnosis.The system has already been deployed in the hospital.