Eye Image Segmentation Based On Deep Learning

Eye image processing technology can help doctors diagnose many eye diseases,such as glaucoma and cataracts.The early disease discovery usually allows patients to receive timely treatment,thereby avoiding permanent vision loss.In recent years,deep learning has achieved fruitful results in tasks such as image classification and segmentation.How to use deep learning to further improve the performance of eye image processing has become a research hotspot.This thesis studies the segmentation of eye images based on deep learning,including optic disc and cup segmentation,retinal blood vessel segmentation,and the anterior segment optical coherence tomography(AS-OCT)image segmentation.However,eye image segmentation faces many challenges: 1)the fundus image contains complex information that makes the optic disc and cup segmentation difficult;2)the prior knowledge in fundus image is hard to fully utilized;3)the low contrast and weak edge information decrease the segmentation performance;4)the segmentation result of the deep learning-based method may not conform to medical cognition.To address the above problems,the innovations and contributions of the thesis are summarized as follows:1)To accurately segment optic disc and cup,this thesis proposes a Mask-RCNN based network to localize and jointly segment the optic cup and optic disc.The network utilizes a segmentation-based region proposal network(RPN)to locate the optic disc and cup accurately.Furthermore,a pyramid Roi Align module is proposed to learn more discriminative feature representation by fusing multi-level information.The optic disc and cup are segmented simultaneously by considering the positional relationship.Experimental results show the effectiveness in localization and the improvement in segmentation.2)To make full use of the prior knowledge in fundus images,this thesis proposes to combine deep learning with level set methods for the optic disc and cup segmentation.By treating the output of the neural network as a level set,the segmentation result of the neural network can be restricted to satisfy a specific prior shape.To make the network predict the required level set,a level set loss function is proposed and consists of the length constraint and the region constraint.The length constraint makes the segmentation boundary smooth,and the region constraint can make the response within the predicted region tend to be consistent.Experimental results show that the proposed method can effectively utilize prior knowledge to improve the segmentation of the optic disc and optic cup.3)To achieve accurate segmentation in fundus images with low-contrast and weak edge information,this thesis adopts multi-information fusion strategies,including multi-scale fusion,feature fusion,and classifier fusion to perform retinal vessel segmentation.The multi-scale fusion strategy can segment retinal vessels of different sizes.The feature fusion strategy combines deep features with retinal vessel enhancement filter response to restore the edge information loss caused by the down-sampling operation.The classifier fusion strategy provides extra supervision to the network.This thesis also proposes a neural network combined with guided filters for vessel segmentation.The network can obtain the edge information from the guidance image to recover the edge information loss caused by the down-sampling operation.In addition,the guided filter module also reduces the impact of noise for segmentation.Experimental results show that the proposed method can effectively segment small blood vessels with weak edge information in low-contrast regions.4)To make the segmentation result of AS-OCT image meets the medical cognition,this thesis designs a shape template to finetune the segmentation boundary of a deep neural network.A U-shaped network is proposed to predict the initial mask of the cortex and lens nucleus in the AS-OCT image.Furthermore,a shape template is designed based on the physiological structure of the nucleus to finetune the predicted nucleus boundary.Experimental results show that the refinement can improve the accuracy while making the segmentation results in line with the physiological structure.