| The eye is an essential visual organ which can sense light.The eye diseases which can cause blindness seriously threaten human visual health.Glaucoma is the world’s first chronic eye disease that causes blindness.The visual damage caused by glaucoma is irreversible.So far,glaucoma is incurable,only early detection and early treatment can prevent glaucoma damage to visual health,so the early screening and diagnosis of glaucoma are critical.Examining the optic disc and optic cup is very important for the early screening and treatment of glaucoma,and the ratio of the optic disc and optic cup is one of the crucial indicators for diagnosing glaucoma in clinical practice.The optic disc and optic cup ratio can be quickly obtained by automatically and accurately dividing the optic disc and optic cup with a computer to assist doctors in diagnosing glaucoma.The existing methods have achieved satisfactory results in optic disc segmentation.However,due to the small area of the optic cup and the low contrast with the edge of the optic disc,the result of the optic cup segmentation is poor,and the task of separating the optic disc and optic cup is still very challenging.To effectively improve the segmentation performance of the optic cup and optic disc segmentation model,this paper first explores the impact of the channel attention mechanism on the performance of the semantic segmentation model.Based on the SE-Net channel attention module,an improved channel attention mechanism based on non-global pooling is proposed.Then the attention mechanism is applied to the optic cup and optic disc segmentation.Secondly,to fully fuse the feature information of the model and enhance the feature reuse performance of the model,this paper improves the feature extraction module and jump connection structure of the U-Net network model and proposes an AFFM-Unet network model based on multi-feature fusion,which can effectively improve the accuracy of the optic and optic cup segmentation.The main research work of this paper is as follows:(1)The SE-Net channel attention module is improved to change the original global pooling into non-global pooling.In addition,a channel attention module based on non-global pooling is proposed to obtain richer local features.Finally,the proposed non-global pooling attention module is applied to the semantic segmentation task,and the effectiveness of this method is verified on the PASCAL VOC 2012 enhanced image segmentation dataset.(2)The proposed non-global pooled attention module is applied to the optic cup and optic disc segmentation experiment to guide the network to learn vital information about the image independently to improve the accuracy of the optic disc and optic cup segmentation.(3)The feature extraction module and jump connection module of the U-Net network model is improved to make the model fully extract image features and fuse multi-scale image feature information.(4)Multi-scale input and deep supervision structures are introduced to assist network training and accelerate convergence.AFFM-Unet shows advanced segmentation performance on the REFUGE data set.Among them,the Acc of the optic disc is 97.46%,the Dice of the optic disc is 98.14%,the IoU of the optic disc is 96.38%,the Acc of the optic cup is 95.89%,the Dice of the optic cup is 94.28%,and the IoU of the optic cup is 89.97%.The experimental results show that AFFM-Unet is superior to other advanced optic disc and optic cup segmentation methods. |
