Fundus Retinal Image Segmentation Based On Convolutional Neural Network

The distribution of capillary structure in fundus retinal images can be used as an important basis for the diagnosis of some diseases.Fast and accurate segmentation of fundus retinal images by computer can avoid the problem of artificial segmentation errors.This method can not only improve the accuracy of segmentation but also reduce the workload of doctors.Therefore,this technology has profound implications for assisting disease diagnosis.In this paper,convolutional neural networks are used to design two models with different structures.The experimental results showed that the network model proposed in this paper was effective and could accurately segment fundus retinal images.The main contents of this paper are as follows:(1)This paper introduces the principle and theoretical basis of convolutional networks and common convolutional neural network models.Image enhancement methods such as gamma transform and histogram equalization were used to preprocess fundus retinal images to enhance the contrast of blood vessels in the images.In addition to the conventional data expansion methods such as rotation and clipping,this paper also expands the data by generating the retina image by generating the adversal network.(2)According to the characteristics of blood vessels in retinal images,this paper proposes a method of segmenting retinal blood vessels in fundus based on residual convolutional network.The structure of codec network is constructed by connecting the low-level feature map and the high-level feature map by skip connection.The hollow convolution pyramid module is added to the model to increase the network receptive field and reduce the model training parameters.The loss function is optimized by the deep monitoring mechanism to enhance the performance of the model.(3)The structure of the traditional U-Net network model is improved.Res Net18 is used as the down-sampling part of U-Net to extract image features.Then,the downsampling part of U-Net is connected with the up-sampling part through SENet,and the deep features and shallow features are input into the fusion attention module to enhance the feature representation.In the up-sampling part of the network,dense convolution blocks are used to establish dense connections between the front and back layers,which can realize the multiplexing of features in the channel dimension,so as to slow down the occurrence of gradient disappearance.Finally,in order to make it easier to use the model,the trained model is deployed to the Android terminal.