Detection Of Diabetic Retinopathy Based On Multi-Layer Cascadefusion Network

As living conditions get better,diabetes poses a very serious hazard to many people.A series of retinopathy caused by diabetes,commonly referred to as Diabetic Retinopathy,referred to as DR,can lead to complete blindness in severe cases.Current important imaging methods include two types,Optical Coherence Tomography and image visualization of the fundus,and Optical Coherence Tomography images are generally referred to as oct images.There are some differences between the two images.The usual problems with fundus images are subjective limitations and different opinions of experts on the severity of the disease.The characteristics of oct images are more obvious,especially in the early stage of the disease,the technical characteristics determine its high sensitivity,and doctors can easily make judgments.However,in practice,the number of oct images is large,and doctors need to look through these images to make judgments,which both consumes doctor’s energy and delays the patient’s treatment time.Therefore,it is necessary to use the machine to automatically judge the images.In this way,the image can be detected at an early stage,and the lesion of the patient’s eyeball can be found,and the timely treatment can prevent the patient’s disease from worsening and save the treatment cost.Firstly,based on the method constructed by VGG networks model,the basic classification net is established by referring to the model structure of KAGGLE competition.The networks layer is deep,similar to the VGG networks,using a structure in which multiple convolution layers are followed by a pooling layer.And a special convolution kernel is used to solve the problem that the medical image input size is large and contains a large amount of information.The model is not complicated,but it is very effective.Secondly,the idea of Holistically-Nested Edge Detection(HED)algorithm for multi-scale feature learning is introduced.The HED algorithm has a very good effect on natural image segmentation.It uses a multi-layer fusion structure,which can be used in the process of neural network sampling.The sampling features that may be lost in the middle are merged at the end.We introduce this structure into the basic networks structure and propose a multi-layer cascade fusion classification networks.In the process of model establishment,in order to solve the problem of normalization of feature graphs,a spatial pyramid pooling structure was introduced.The networks can effectively collect medical image information and improve the accuracy of oct image recognition.Finally,the built network is tested on the oct dataset.The performance of the network model constructed in this paper and the classical network model can be compared by various evaluation criteria.It is proved that the structure of multiple cascaded fusions is better than the classical network in oct images.The feasibility of using multi-layer cascaded network structure to detect glycocalyx disease was verified.