Research On Fundus Image Segmentation Algorithm Based On Multi-scale Feature Fusion

According to relevant research surveys,chronic diseases such as hypertension and diabetes have shown a rapid growth trend in recent years.Clinical medical research has found that the symptoms of these diseases are closely related to the retina.Diagnosis provides an important basis,and computer image processing technology provides new ideas for diagnosis.This paper aims to study a fundus image segmentation method to assist doctors in diagnosis.In recent years,convolutional neural networks have been widely used in image processing tasks,and designing network frameworks according to specific segmentation tasks can achieve good segmentation results.Using the U-Net network architecture,this paper conducts research on fundus image segmentation based on deep learning,and proposes a fundus image segmentation algorithm based on multi-scale feature fusion.In order to obtain multi-scale feature information and improve the segmentation effect of the model,the inception module is introduced to replace the partial convolution operation of the basic U-Net encoder part.In order to increase the receptive field of the convolution operation and obtain a wider range of information,this paper uses dilated convolutions with different expansion coefficients in the inception module to process data in parallel.At the same time,this paper adds a PPM module based on the idea of multi-scale average pooling to capture rich feature information.In order to optimize the network performance and prevent the model from overfitting,the drop block technology is introduced after the 3*3 convolution in the network to close part of the continuous units of the neuron to improve the generalization ability of the model.In this paper,the public datasets DRIVE and CHASE are used for evaluation and testing.In order to obtain better segmentation results on a small amount of data,this paper performs data preprocessing on the dataset during the experiment to expand the number of samples and improve the quality of samples.The results of the effective ablation experiments show that the improved U-Net model in this paper has an improvement of 1.04% and 4.31% in accuracy and sensitivity,respectively,compared with the basic U-Net,reaching 95.72% and 81.94%,respectively.At the same time,in order to verify the rationality of the inception module,this paper designs a variety of network variants,and introduces this module in different scale stages in U-Net.The experimental results show that the segmentation performance has been improved to varying degrees.In this paper,a horizontal comparison experiment is carried out to evaluate the performance of the model.The accuracy of the algorithm in this paper is better than that of the three algorithms,which are improved by 1%,0.16% and 0.34%respectively;the sensitivity is better than that of the two algorithms,which is improved by 4.09% respectively.% and 2.62%,further verifying the effectiveness of the algorithm in this paper.