Research On The Depth Segmentation Model Of Optic Cup And Disc In Fundus Image

Glaucoma is a common chronic eye disease,which is an irreversible blinding eye disease.Glaucoma is not accompanied by obvious symptoms in the early stage,and there is no obvious loss of vision.Many patients are already in the advanced stage when they notice the symptoms and then seek medical treatment.At this time,the visual damage is irreversible.Early screening and diagnosis of glaucoma is of great significance for disease prevention and vision protection of patients.The application of artificial intelligence technology in glaucoma screening has a good auxiliary effect on early diagnosis and early treatment of glaucoma.Glaucoma is optic nerve damage and visual field defect caused by primary factors such as pathological increased intraocular pressure and optic nerve insufficiency.Glaucoma alters the morphology of the optic nerve head,manifesting as a large cup-to-disk ratio in the optic nerve head region,graying of the optic disc,or hemorrhage of the optic disc.The Cup to Disc Ratio(CDR)is the ratio of the Optic Cup(OC)to the Optic Disc(OD)area in the fundus image,which is a key indicator for judging glaucoma.Using deep learning to accurately segment the optic cup and optic disc region in fundus images has become an important research content for assisting glaucoma screening.In the fundus image,the optic cup and optic disc will overlap,and their mutual influence restricts the accuracy of segmentation.Accurately segmenting the optic disc can calibrate the error of the optic disc area when the optic cup and optic disc are segmented synchronously.Therefore,this paper studies the depth segmentation model of the optic disc and Optic cup and optic disc synchronous segmentation model,the specific research contents are as follows:(1)In view of the multi-source and heterogeneous characteristics of fundus images,as well as the characteristics of various optic disc shapes,multi-scales,and blurred edges,this paper proposes an improved optic disc segmentation model ACE-Trans UNet(Trans UNet)based on the U-shaped model combined with context and attention with Attention and Context Extraction Mechanism,ACE-Trans UNet),including:(1)In the encoding stage,the IBN(Instance-batch Normalization,IBN)module and the CBAM(Convolutional Block Attention Module,CBAM)module are introduced to improve model generalization and image channels Feature extraction capability;(2)In the feature fusion stage,use the multi-level context information extraction module MCE(Multi-level Context Extraction,MCE)to process the features output by the backbone network,and enhance the model’s ability to extract the edge features of the optic disc;(3)In the decoding stage,use The Transformer mechanism improves the ability to extract multi-scale features of the optic disc and the global information of the fundus image.(2)In order to improve the accuracy and efficiency of the simultaneous segmentation of the optic cup and the optic disc,this paper proposes a simultaneous segmentation model of the optic cup and optic disc(Double Encoder Light Dense Network,DEL-Dense Net),including:(1)Introducing a lightweight convolution module and Mini-Inception,under the condition of controlling the weight of the model,realizes feature multiplexing and enhances the performance of model segmentation;(2)Design the architecture of dual decoders to process different channels of fundus images separately,reducing the impact of cup segmentation on the accuracy of optic disc segmentation;(3)Reduce the convolutional layer in the decoder part,lightweight the segmentation model,and improve the segmentation efficiency of the model.Based on the four public datasets of DRISHTI-GS1,i Challenge-PM,RIGA and RIM-ONE,1005 fundus images containing complete optic cup and optic disc were selected as experimental data,and the proposed optic disc segmentation model(ACE-Trans UNet),optic cup and optic disc synchronous segmentation model(DEL-Dense Net)performed ablation experiments and comparative experiments.Compared with FCN,UNet,UNet++,Deep Labv3+,and PSPNet,the optic disc segmentation model(ACE-Trans UNet)has higher segmentation efficiency and better stability.Its Dice,Mio U,MPA,and FPS reached 98.18% and 96.45% respectively.,98.11%,and 17.56.Compared with Seg Net,Fast FCN,UNet,PSPNet,Deep Labv3+,Refine Net,DEL-Dense Net has high segmentation efficiency and lighter weight.Its Disc Dice,Disc Io U,Cup Dice,Cup Io U and FPS respectively They are 96.91%,94.07%,84.40%,73.56% and 33.61.In summary,the optic disc segmentation model ACE-Trans UNet and optic cup and optic disc simultaneous segmentation model DEL-Dense Net proposed in this paper can provide technical support for early diagnosis of glaucoma diseases.