Research On Multi-Scale Optic Disc And Cup Segmentation Method Based On Optimized Features

Glaucoma is the main cause of irreversible blindness,and early screening and treatment is the key to reduce the rate of blindness.In color fundus images,the shape change of the optic disc and cup is a significant feature of glaucoma,and the calculation of the cup-disc ratio(CDR)has also become an important diagnostic basis in glaucoma screening.The greater the CDR,the higher the risk of glaucoma.China has a vast territory and a large population,and the medical level between regions is not sufficient and unbalanced.Only relying on doctors to diagnose glaucoma can not complete large-scale screening,but also makes it difficult to promote the screening and prevention of glaucoma in remote areas.Using the Deep Learning algorithm to analyze the fundus image for automatic segmentation of the optic disc and cup will help improve the efficiency of glaucoma screening.Based on the improved Deep Learning algorithm,this paper realizes efficient and accurate automatic segmentation of a large number of fundus images in a short time to help doctors diagnose glaucoma.The main work is as follows:1.The encoder-decoder model is often used as the backbone network in the automatic segmentation method to capture more context information.However,due to the semantic differences between the two,the direct fusion of shallow features and deep features will produce noise and affect the segmentation results.To solve the above problems,this paper proposes CSPM-Net to be used for the automatic segmentation of optic disc and cup.In order to improve training efficiency,CSPM-Net uses a lightweight U-Net as the backbone network for feature extraction.CSPM-Net adds a muti-scale pooling module —— CSPM module(Channel-Spacial Pyramid Pooling Module)based on attention mechanism in the early layer of the network,which can enhance the shallow features and reduce the impact of the semantic gap.The CSPM module extracts rich multi-scale features through pooling cores of different sizes,and transmits the spatial context information lost in the feature extraction process to the later stage of the network.The attention mechanism suppresses irrelevant semantic information to reduce the semantic differences between shallow features and deep features.In this paper,CSPM-Net was trained and tested on three datasets: Drion-DB,Drishti-GS and RIM-ONE v.3.Compared with the existing algorithms,CSPM-NET was improved.2.Most of the existing optic disc and cup segmentation algorithm networks are trained in a fully supervised manner,relying on a large number of annotation data,which is a very scarce data resource.In order to reduce the dependence of network training on labeled data and make full use of unlabeled data,this paper proposes a semi-supervised algorithm based on boundary perception for cup and disk segmentation.This method based on the Transformation Consistent Self-ensembling Model.There are two models with the same structure but different weight updating methods in the method: one uses the gradient descent method to update the weight,which is called the student model;Another is the average value of the student model,which uses the moving average method to update the weight,called the teacher model.By minimizing the difference between the teacher model and the student model in the same target prediction results,we can use the unlabeled data to calculate the consistency constraints of the two models for training the network.In the training,it was found that the commonly used Dice Loss can solve the imbalance of the front background in the cup and disk segmentation,but it is not sensitive to the wrong segmentation of the target boundary.To solve this problem,this paper constructs a compound loss function based on boundary loss for network training to improve the sensitivity of the network to boundary segmentation.In addition,this paper introduces uncertainty awareness into the consistency constraint to filter out the uncertain part of the network’s prediction of unlabeled data and ensure that the network can learn from reliable targets.The experimental results on the Refuge dataset and the existing semi-supervised algorithm show that the semi-supervised algorithm based on boundary perception effectively utilizes the unlabeled data and improves the network segmentation accuracy.