The Research On Segmentation And Classification Of Diabetic Retinopathy Based On Deep Learning

When detecting diabetic retinopathy,deep learning can perform efficient and accurate analysis.It makes significant effects in its related diagnosis,including the classification of lesions,the segmentation of lesions,and the segmentation of retinal blood vessels in the fundus.Deep learning can not only improve the diagnosis efficiency of doctors,but also help doctors detect this lesion in time.It makes it possible to implement large-scale accurate diagnosis,and also promotes the development of automatic diagnosis in the medical field.In this paper,the improved U-shaped network is used to achieve the segmentation of the fundus retinal blood vessels and diabetic retinopathy,and the multi-fusion classification network is used to achieve the classification of diabetic retinopathy.The research of this article is as follows:(1)Aiming at the problems of the bifurcation of microvessels and the poor segmentation at the endpoints,the inconspicuous exudate boundary,and the small and scattered distribution of bleeding points that are difficult to segment,an improved Ushaped network is proposed,and the context feature coding module is improved to extract richer High-level features.In the feature coding stage,a hybrid attention mechanism is added to highlight the characteristics of microvessels and lesions,and reduce the impact of background and noise.Experiments have proved that our method can distinguish different information well.It also improves the segmentation effect of microvessels at the bifurcation and endpoints,and small and scattered lesions.It can help the doctor diagnose effectively.(2)To address the problem of similarity between grades of diabetic retinopathy.Moreover,the feature information of retinal vessels and lesions plays a great role in the grade classification of diabetic retinopathy.The segmentation results of the fundus retinal vessels and lesions of the diabetic retinopathy classification data set are obtained by the proposed improved U-shaped segmentation network,and feature fusion is performed.The obtained fusion feature image and the original fundus image are input as the classification network.It is helpful to guide the network training and make the network pay more attention to the feature information that has a beneficial impact on the classification effect.The classification network model adopts the multi-network fusion model.It can effectively combine the advantages of different networks and make comprehensive use of their classification results.Experiments verify that the method in this paper can effectively overcome the misclassification caused by the unobvious difference between classes and the serious imbalance of the data set,and obtain good classification results.