Research On Joint Segmentation Algorithm Of Optic Cup And Optic Disc In Fundus Image Based On Deep Learning

Glaucoma is the second leading cause of blindness worldwide,after cataracts.If left untreated,glaucoma gradually damages the retina and optic nerve,causing irreversible vision loss and eventually blindness.Clinically,fundus examination is one of the main methods for diagnosing glaucoma.Ophthalmologists diagnose glaucoma by observing the morphological features of the optic cup and optic disc in fundus images and calculating relevant clinical parameters.This work requires professional medical knowledge and experience,and is very time-consuming and laborious,so it is not suitable for large-scale screening.In order to improve the efficiency of early screening of glaucoma,this thesis explores the method of automatic segmentation of optic cup and optic disc in fundus images based on deep learning.Compared to traditional artificial diagnosis methods,deep learning methods use deep neural networks to train a large number of fundus images to automatically identify the position and contour of the optic cup and optic disc,thereby helping doctors quickly and accurately diagnose glaucoma and adapt to the needs of large-scale screening.Therefore,this thesis conducts in-depth research on the segmentation of optic cup and optic disc in fundus images based on deep learning.The main research contents are as follows:(1)A hybrid network named CTH-Net(CNN-Transformer Hybrid Network)based on CNN and Transformer is proposed for joint segmentation of optic cup and optic disc.CTHNet combines the advantages of convolutional neural network in capturing local features.And Transformer’s advantages in handling global context and long-distance dependencies.By efficiently mixing the two model architectures,local information and global contextual information can be effectively utilized simultaneously,thereby improving the accuracy and robustness of the model.In addition,the hybrid model also introduces the ASPP(Atrous Spatial Pyramid Pooling)module.The ASPP module can effectively expand the receptive field,improve the model’s ability to extract features of different scales,and enable the model to better handle complex features of fundus images.(2)A hybrid network named CSTH-Net(CNN-Swin Transformer Hybrid Network)based on CNN and Swin Transformer is proposed for joint segmentation of optic cup and optic disc.CSTH-Net uses a dual encoding structure,feature fusion module and skip connection technology.The dual encoding structure enhances the global contextual semantic information extraction ability of the network by introducing the Swin Transformer encoder branch,and makes up for the deficiency of the convolutional neural network encoder branch.Therefore,the dual-encoding structure can better deal with the long-distance dependence problem and different-scale information capture problem in fundus image segmentation.The feature fusion module can fuse the global information of the image with the local information without losing the details of the image,so as to improve the accuracy and efficiency of segmentation.At the same time,CSTH-Net also uses the skip connection technology to fuse the multi-scale features of the encoder with the up-sampling features,which can better capture the different scale information of the image,and further improve the accuracy and efficiency of the segmentation.