Research On The Segmentation Method Of Color Fundus Image Cup Based On Deep Learning

Glaucoma is a chronic eye disease in which the optic nerve is progressively damaged.It is usually associated with changes in the optic cup,optic disc and retinal nerve fiber.The segmentation of optic cup can be applied for further parameters measurement,which are important for glaucoma assistant diagnosis.Therefore,the accurate segmentation of optic cup has clinical significance for automatic screening and inhibiting the development of glaucoma.In this thesis,a method of cup segmentation based on FCM clustering is proposed,which combines the feature of the blood vessels bending at the cup boundary to correct the position of the cup.Although good segmentation results are obtained,traditional cup segmentation methods need manual design features,which have complex processes and low robustness.The deep learning offsets the shortcomings of traditional cup segmentation methods.Deep learning is essentially an extension of neural networks,and has made a breakthrough in speech recognition,target detection,semantic segmentation,and many other artificial intelligence fields.The method of semantic segmentation by deep learning is using pixel-level annotated images.Some special layers such as upsampling and deconvolution are used to restore the original image size,so as to achieve an end-to-end learning.In this thesis,a new network architecture called Seg-ResNet is presented.In the new architecture,residual network structure is the main body,and squeeze-and-excitation structure is introduced to automatically adjust the dependency relationship of feature channels through the way of learning,thus realizing the re-calibration of the weight of feature channels.Then the weighted low-level features and high-level features are fused to improve network performance.Also,we combine with the transfer learning to accelerate the speed of convergence and improve the segmentation accuracy.In this thesis,GlaucomaRepo and Drishti-GS fundus image datasets are tested and compared from many aspects,which prove the robustness and effectiveness of optic cup segmentation based on deep learning.