Automatic Detecting Glaucoma Based On Lightweight Network

Glaucoma is the second leading cause of blindness and the first cause of irreversible blindness.During the disease,the visual acuity level of patients is often a gradual and irreversible slow damage,and most patients have no obvious symptoms until the visual impairment is severely damaged.Therefore,glaucoma is also known as the " silent thief of sight".Glaucoma cannot be prevented.Only through early detection and early treatment can prevent the disease from continuing to deteriorate.This measure can effectively reduce the blindness rate of patients.However,the number of ophthalmologists is still insufficient to support large-scale glaucoma screening.In clinical practice,medical images can effectively help doctors make judgments.Color fundus photography is currently the lowest cost of various non-invasive retinal examination methods.The optic cup and optic disc diameters were obtained from color fundus photography,and the vertical cup-to-disc ratio was calculated.This is a commonly used clinical indicator to judge glaucoma.It also takes about 8 minutes for a senior ophthalmologist to segment a color fundus photo.So many researchers focus on semantic segmentation techniques,especially those based on deep learning.So far,many algorithms have been studied on the glaucoma recognition problem,but they are generally computationally expensive.These algorithms are not suitable for large-scale glaucoma screening.At present,a lightweight network is urgently needed to solve this problem.This paper proposed the Lightweight Glaucoma Network（LGNet）with high segmentation accuracy and low computational cost.By controlling the network width and introducing a channel splitting mechanism,the model capacity is simplified.A basic module that effectively improves the receptive field is formed by parallel stacking of atrous convolution,and the ability to acquire and fuse context information is optimized.By introducing the attention mechanism,the feature extraction ability of the encoder is improved.Through the design of loss function and the strategy of data augmentation,the problem of sample imbalance is alleviated.Simplifies the commonly used two-stage solution of localization-then-segmentation to an end-to-end one-stage segmentation.While other researchers commonly use multi-model prediction,this paper uses a single model to solve the multi-class segmentation problem.In this way,resource overhead can be reduced.In this paper,the network model is verified on the public dataset.Compared with the advanced methods in recent years,it has achieved good results in the evaluation indicators such as Dice coefficient,Io U,parameters and FLOPs.In this paper,the effectiveness of the network model,training strategy and loss function design is verified through ablation experiments.