Design And Implementation Of A Lightweight Model Based On Convolutional Neural Networks To Segment Optic Cup And Optic Disc

The eyes are the most important sense for the human body to obtain information from the surroundings,which play a very important role in daily life.Glaucoma is the second largest ocular disease causing blindness worldwide,and it is difficult to cure glaucoma once diagnosed.Therefore,early screening for glaucoma is of great significance.The optic cup and optic disc in color fundus photography are the important basis for the diagnosis of glaucoma.Because the subjective estimation error of the optic cup and optic disc of different doctors is large,it is very important to use the computer aided diagnosis technology to help doctors to segment them quickly and accurately.Optic cup is completely inside the optic disc and the edge of the optic cup is not clear enough,so the accurate segmentation of optic cup and optic disc is a challenging task.At present,there have been many methods to study the segmentation of optic cup and optic disc.Although the performance of deep learning method is better than that of traditional methods,the number of model parameters is large,requires high video memory and multi-scale information is not considered more.In this paper,the region of optic disc is intercepted from fundus photography as the region of interest to reduce the amount of calculation during training,and a lightweight semantic segmentation model based on convolutional neural network,Eye Net,is proposed to achieve automatic,rapid and accurate segmentation of optic cup and optic disc.The main research work of this paper is summarized as follows:(1)Aiming at the problem that the target size of optic cup and optic disc varies greatly in different fundus photography,this paper fully introduces multi-scale information in the coding stage to improve the receptive field of feature map without increasing the network depth,so as to realize the accurate recognition of the edge of optic cup and optic disc.(2)In view of the problem that the edge of optic cup and optic disc is not clear enough,this paper proposes an edge module,which can enlarge the differences between features and help improve the segmentation ability of the target edge.(3)Coordinate attention mechanism is introduced in this paper,and model is carried out in channel dimension and spatial dimension to enhance important information,so as to achieve the purpose of accurate segmentation.The final weight of Eye Net model is 2.71 M,and the accuracy of dice segmentation on the test set reaches 89%,which are superior to the lightweight Deeplab V3+ model based on Mobile Net V2 in terms of both the number of parameters and accuracy.