Analysis And Research Of Diabetic Retinopathy Based On Deep Learning

Diabetic retinopathy(DR)is one of the most common complications of diabetes and the eye disease with the highest blindness rate in the world.In the early stages of DR,patients are often unaware of its effects and easily miss the best time for treatment.In clinical diagnosis,the ophthalmologist makes a diagnosis by observing the retinal images of the patient.This method is very dependent on the doctor’s diagnostic experience,which not only takes a long time to diagnose,but also consumes a lot of energy,and is prone to missed diagnosis and misdiagnosis.In addition,due to the large number of patients in my country and the relatively tight medical resources,some patients cannot receive timely diagnosis and treatment.Regular screening can achieve early detection of the disease and prevent patients from missing the best treatment period.In recent years,deep learning has been widely used in the field of medical images and has achieved remarkable results.It has become an urgent need for ophthalmologists to diagnose patients based on DR images using computer-aided diagnosis technology.Based on this,this thesis uses the existing deep learning framework and learning model to build a DR classification and lesion segmentation model for diabetic retinopathy images,so as to achieve accurate classification of DR images and accurate segmentation of exudates at key lesions.Computeraided doctor’s efficient diagnosis plays a very important role.The main research work is as follows:(1)Aiming at the problem that the existing deep learning model has low classification accuracy due to the small difference between different DR lesions images,this thesis designs a classification network based on Mobile Net V2 and Efficient Net B0 model integration and attention mechanisms: ME-ANet.When designing the network,the idea of integration is adopted,and the core structures of the deep learning model Mobile Net V2 and Efficient Net B0 are integrated in stages to form the feature extractor of the model,which realizes the fusion of multi-scale feature information of retinal images.At the same time,a global attention mechanism(GAM)is designed and embedded into the ensemble model at different stages to obtain the weight information of lesion features and suppress the noise of non-lesion features.The network realizes the full extraction of image shallow feature information,reduces the convolution loss problem of micro-lesion feature information in the training process,and effectively improves the classification performance of the model.(2)Aiming at the problem that the existing deep model segmentation accuracy of exudate lesions in DR images is not high and affects the later diagnosis,this paper constructs a deep learning-based model for the segmentation of diabetic retinal exudates.The model uses U-Net encoder-decoder structure is the basic network architecture,and Res Net50 is used as the feature extraction network at the encoding end,which not only realizes the further deepening of the feature extraction network,but also avoids the problem of gradient disappearance during model training;at the decoding end,design attention Finally,a multi-feature scale fusion layer is constructed at the decoding end of the model to realize the learning of multi-scale semantic features of the model.During the experiment,the Focal loss focal loss function is used to further strengthen the model’s learning and optimization of the characteristics of small target objects in the sample.Compared with other exudate segmentation models,this model has better segmentation performance and has universal applicability.In summary,this paper builds a deep learning classification model to complete the task of classifying diabetic retinal lesions,and then builds a deep segmentation model to complete the task of segmenting diabetic retinal exudates,so as to achieve the purpose of computer-aided medical diagnosis more targeted,to further improve the screening and diagnosis efficiency of diabetic retinopathy.