Research On Joint Segmentation Method Of Optic Cup And Optic Disc Based On U-net

At present,glaucoma is the second largest blinding disease in the world.Nearly1.67 million people are diagnosed with glaucoma every year.It starts with peripheral vision loss and progresses to severe vision loss or blindness.Until now,glaucoma is incurable,and only early detection and treatment can halt its progression.Recognizing the signs of glaucoma requires specialized ophthalmologists with years of experience and practice,but the large patient population cannot be diagnosed in a timely manner.Therefore,the development of automatic glaucoma assessment algorithms based on fundus image analysis will help reduce the overall workload of ophthalmologists and make diagnosis more feasible and effective in smaller health units.In this paper,the optic cup（OC）and optic disc（OD）in fundus images are segmented based on deep learning to help doctors diagnose glaucoma quickly and accurately.The main work of this paper is as follows:The proportion of OD in fundus image is very small,and it is not satisfactory to input fundus image into the segmentation network directly.Through the method of OD localization,the region of OD image can be cropped,the area of image detection can be reduced,and the accuracy of OD and OC segmentation can be improved.Therefore,this paper proposes a YOLOv5 based network to effectively improve the accuracy of OD detection.Firstly,the position of OD is marked in the fundus images with labeling software,and the data set for OD detection is obtained.Then the data set is input to YOLOv5.It is found that YOLOv5 could not detect the location of OD exactly.Therefore,the Backbone of YOLOv5 is added with Multi-scale Convolution（MC）module to improve detection accuracy,which improves network learning ability.Secondly,a channel attention mechanism is added in the Neck part of YOLOv5 to suppress the expression of information irrelevant to the task.Finally,the detected OD region is cropped into 512×512 sub-image as the input image of OC and OD segmentation network.Experimental results show that the proposed Yolo V5-based network can achieve 100% accuracy on REFUGE and Drishti-GS1 data sets.Measuring cup disk ratio（CDR）by segmenting OD and OC is an effective method for diagnosis of glaucoma.In this paper,the deep learning architecture FA-Net is proposed to jointly segment the OC and OD in fundus images.FA-Net is an improved network structure based on U-Net.U-net is a U-shaped network composed of encoder and decoder.Due to the simple structure of U-Net,it failed to achieve good segmentation effect.By adding feature fusion module in U-NET,information loss in feature extraction can be compensated.The attention mechanism combining channel and space is added to the decoding path to highlight the important features related to the segmentation task,which effectively improves the accuracy of segmentation.The experimental results show that the performance of FA-Net on REFUGE and Drishti-GS1 data sets is better than the existing network of segment OC and OD.