Research On The Segmentation Method Of Bleeding Points In Fundus Images Based On Deep Learning

In recent years,with the improvement of people’s quality of life and the change of life patterns,diabetes has gradually become one of the important causes of human illness and death.The most common and serious microvascular complications of diabetes is Diabetic Retinopathy.DR).DR can cause people’s vision to decline,is one of the main blind eye diseases in people over 50 years old,ranking first in China’s blinding retinal vascular disease.Hemorrhage(HA)is an early symptom of DR.Therefore,accurate detection of HA can effectively solve the problem of automatic screening of DR,which is of great significance for the timely detection and treatment of diabetes.The traditional HA detection mostly adopts the image processing method.Because the characteristics of HA are more complicated,the traditional methods are often not perfect in the feature extraction of HA,which is easy to cause false detection and missed detection of HA.Aiming at the complexity of HA features,this paper proposes a method of segmentation of fundus image bleeding points based on deep learning.It mainly combines Dilated convolution and channel weighting module to improve PixelNet,which makes the fundus image HA segmentation better.The convolutional layer is the core layer of the convolutional neural network,which in turn is an informational aggregate of local receptive field features and channel features.In this paper,the shallow features of the network are cascaded with the deep features,and the Dilated convolution is used instead of the ordinary convolution form.When the computational load is saved,the local receptive field is increased,and the spatial relationship of more pixels is obtained.Image segmentation based on channel weighting.The weighting module first performs dimensionality reduction on the output of the convolutional layer,so that the shallow features also have a large receptive field;then the weight of the feature channel is obtained by learning;finally,the feature channel is weighted according to the learned weight.Channel weighting can enhance useful features to suppress useless features,and help deeper networks to achieve better learning results.The resulting model is more generalizable and more robust.The experimental results show that the Dilated convolution can obtain a wider receptive field,which enhances the feature extraction ability of the network.The channel weighting realizes the re-calibration of the channel weight,which improves the generalization ability of the network.In view of the segmentation problem of the bleeding point detection for fundus images,The improved PixelNet proposed in this paper can realize HA automatic detection with high accuracy.