Research On Fundus Image Segmentation Algorithm Based On Improved Encoder-Decoder Structure

In the medical field,medical images are an important basis for diagnosis and treatment of diseases,so the analysis of the visualization results of human internal tissues and organs has great clinical significance.Among them,the segmentation of medical images is even more important,because the purpose of assisting medical treatment can be further achieved through the quantitative calculation of regional lesions or target tissues.Fundus images are now an important auxiliary examination method in medical diagnosis.Since the cornea,lens,and vitreous in the eyeball are transparent,we can usually obtain clearer fundus images.This is very helpful for further processing and analysis.The blood vessels on the retina are the visible blood vessels in the human body.Doctors usually analyze the state of these blood vessels to understand the blood vessels in other parts of the body.Nowadays,many diseases can get some symptom responses from fundus images,such as diabetes,hypertension,etc.Therefore,segmentation of fundus images is an important task in the medical field.This article analyzes the characteristics and difficulties of fundus image segmentation,and introduces the relevant technical knowledge about fundus image segmentation.With the rapid development of convolutional neural networks in the field of computer vision,many deep learning algorithms have begun to be combined with the medical field,and gradually focus on the clinical diagnosis of various diseases.At present,there is a large amount of research work that has applied different image segmentation models to various medical image data sets,and the final experimental results show that the segmentation effect of many models has reached a level equivalent to that of professional doctors and can be truly applied to the clinical diagnosis.Therefore,the use of convolutional neural network to segment the fundus image is a very meaningful research work.Since medical images have the characteristics of less data and high complexity,the encoder-decoder structure has achieved good results in processing medical image tasks.In this article,the neural network and fundus image segmentation of the encoder-decoder structure are deeply studied.Based on the open source fundus image data set DRIVE,our experiments are carried out and an improved encoder-decoder network structure is proposed.The network structure introduces multiple atrous convolutions in the encoder part to meet the extraction requirements of different sizes of features,and at the same time adds the SE block to improve channel attention,and finally uses different sizes of average pooling to obtain more context information;The decoder part uses deconvolution to perform up-sampling to restore the size of the feature map.After a series of experiments,the network structure proposed in this article has achieved good results on the DRIVE data set,with sensitivity(Sen)reaching 0.8070,accuracy(Acc)reaching 0.9563,and AUC reaching 0.9779,which is better than many network structures including U-Net,and proves that the improved encoder-decoder structure can effectively segment the fundus image.In addition,we set up many variants of the network proposed in this article in the later stage,and conducted a series of comparative experiments to verify the influence of the location and number of SE blocks on the network performance.As a result,the performance of other variants is slightly lower than the network of this article.We also conducted ablation experiments to further prove the effectiveness of the improved encoder-decoder structure proposed in this paper in the task of fundus image segmentation.