Research On Segmentation And Classification Algorithms Of Fundus Images

Glaucoma and diabetic retinopathy have become the main diseases leading to impaired vision and blindness.Regular screening is particularly important for patients with fundus diseases.It can detect the condition early,delay the discovery of the disease,and avoid visual impairment and blindness.In recent years,with the rapid development of computer technology,the application of computer-aided diagnosis system is booming.Computer aided system can improve the accuracy of disease diagnosis and save manpower and material resources.Therefore,fundus image analysis is of great significance to clinical medicine.At present,in the task of segmentation of glaucoma fundus images,most methods use U-Net-based framework.However,these studies have ignored the following issues:(1)In the traditional coding layer,image features are often extracted through repeated convolution operations,which are too simple to extract finer features;(2)In decoding In the layers,the continuous deconvolution operation ignores the different information between different layers;(3)The segmentation performance of fundus images in small areas is not ideal.In the image classification task of diabetic retinopathy,the finegrained properties of the fundus lesion image are ignored;the traditional convolution operation cannot extract specific image features well;the diabetic retinopathy image data set generally has the problem of uneven sample distribution,which is limited the classification performance of the network is improved.Regarding the problems in the above two areas,this article proposes the following methods to deal with these challenges:1.This paper proposes a new aggregation channel attention network to make full use of the influence of contextual information on image segmentation.Different from the existing attention mechanism,this paper uses channel dependence and integrates information of different scales into the attention mechanism.At the same time,this paper improves the basic classification framework based on cross entropy,combines the dice coefficient and cross entropy,balances the contribution of dice coefficient and cross entropy loss to the segmentation task,and improves the performance of the small area network.segmentation.The network retains more image features,restores important features more accurately,and further improves the segmentation performance of medical images.2.This article uses a fine-grained image classification method to solve the problem of diabetic retinopathy classification.This paper uses compact bilinear pooling to achieve fine-grained image classification,and uses the channel attention mechanism to extract more specific image features.For the problem of sample imbalance,we used two loss functions to analyze the impact of sample imbalance on the experimental results from different perspectives.Various experiments have shown that the method used in this article has a positive effect on the classification of diabetic retinopathy.