Research On Optic Cup And Disc Segmentation Algorithm Of Fundus Image Based On Improved U-Shaped Network

Glaucoma is a fundus disease with a high prevalence,and it has been paid more and more attention because of its irreversible characteristics,the number of patients increasing year by year,and the younger and younger age of disease.Manual diagnosis of glaucoma disease is time-consuming and labor-intensive.it is difficult to conduct a large-scale population census.Therefore,it is particularly necessary to carry out automatic identification and auxiliary diagnosis of glaucoma based on fundus images.Among them,the glaucoma diagnosis method based on the optic nerve head is the most feasible method.This method only needs to segment and extract the optic cup area and optic disc area of the fundus image.By calculating the value of the cup to disc ratio,it can assist the doctor in making diagnosis of glaucoma.Most of the existing optic cup and optic disc extraction methods are based on Morphological Image Processing,but they rely too much on manual operations and have great limitations.In order to solve the problem of accurate extraction of optic cup and optic disc from fundus images,this paper trains the network through deep learning,and studies the automatic extraction and segmentation of image features.This paper constructs a multi-task learning model based on U-shaped network by introducing contour detection task and distance map estimation task.Aiming at the problem that the U-Net network has relatively rough edges in the segmentation of retinal optic cup and optic disc,this paper introduces the contour detection task branch.Through the experimental verification,the introduction of the contour detection task makes the final segmentation map have a smoother boundary,which is consistent with the physiological structure of the optic cup and optic disc.Aiming at the problem of some misclassified pixels in the segmentation results,this paper introduces the task branch of distance map estimation.Through experimental verification,the distance map estimation task can help the model to better reduce the occurrence of misclassification problems.The introduction of two auxiliary tasks can help the model to retain more information about the shape and boundary of the target area in the image;this paper highlights the advantages of the joint segmentation of retinal optic cup and optic disc.The position information of the optic disc is introduced in the process of cup segmentation,which improves the accuracy of segmentation.Compared with existing methods,the method proposed in this paper has better segmentation performance,among which RIM-ONE r3 dataset OD（Dice=0.970;Jaccard=0.896）,OC（Dice=0.908;Jaccard=0.823）;Drishti-GS1 dataset OD（Dice=0.968;Jaccard=0.903）,OC（Dice=0.909;Jaccard=0.791）.In this paper,comparative experiments and ablation experiments are also designed for different auxiliary tasks.The experimental results show that the multi-task structure proposed in this paper has good segmentation performance.