Research On Joint Optic Disc And Cup Segmentation Based On Convolutional Neural Networks

Glaucoma is an eye disease that causes irreversible loss of vision.Therefore,early screening can detect diseases as early as possible,which is essential for early treatment to maintain healthy vision and maintain quality of life.However,manual diagnosis performed by ophthalmologists is time-consuming and expensive,and is not suitable for screening large-scale populations.The automatic segmentation of the optic disc and cup can obtain some clinical data needed for the evaluation of glaucoma,alleviating the time-consuming and expensive problem of manual evaluation.Traditional methods are mainly based on artificially extracted features such as colors,stripes,etc.,lack sufficient distinguishing and representation capabilities,and are easily affected by the lesion area and low-quality contrast.Deep learning methods can automatically learn and extract features in fundus images,and are relatively less affected by lesions.In recent years,deep learning methods have been applied to segment the optic disc and optic cups,and some good results have been obtained,but there is a lack of research on the relationship between optic disc and optic cup.Therefore,it is necessary to explore new deep learning methods to overcome these shortcomings.Deep learning methods are used to extract more visual feature information of retinal images than other methods,and use these visual features to segment the optic disc and optic cup in order to achieve a better automatic glaucoma screening effect.In view of the above problems,this paper designs an improved automatic segmentation framework for optic disc and optic cup based on full convolutional neural network.The main research contents of the paper are as follows:(1)This research proposes a network model Recurrent Fully Connection Network(RFC-Net)based on recurrent full convolution and polar transformation.RFC-Net rebuilds and fine-tunes the fully convolutional network,adding recurrent convolution to it,making the network richer,and is committed to improving the segmentation accuracy of the disc and the cup.By adopting a multi-scale input layer,a substantial increase in parameters is effectively avoided,the sensitivity of the network to global context information is improved,and the network width of the decoder path is increased.The side output layer is used to generate accompanying local prediction maps of different scale layers,and the side output layer is used to make full use of the feature information at different stages to improve the network learning ability without adding additional parameters and calculations.Verification through comparative experiments shows that both the multi-scale input layer and the side output layer can make the model perform better in the segmentation task.(2)A model Skip-Link Attention Guided Network(SLAG-CNN)based on a new type of Skip-Link attention guidance network and polar transformation is designed for joint segmentation of optic disc and optic cup.First,in SLAG-CNN,the attention gate SLAG module is designed.It is used as a sensitive extension path to transfer the semantic information and position information of the previous feature map to eliminate the noisy components introduced from the complex background by the original feature map.SLAG-CNN with novel network structure and attention gate has played a great role in the improvement of the whole model.It filters low-resolution feature maps and high-resolution feature maps to recover spatial information from different resolution levels and merge structural information.This can effectively learn better attention weights to the subtle boundaries of the optic disc and optic cup.Secondly,combined with the idea of polar transformation,the model is trained using the feature map in the polar coordinate system,which improves the segmentation performance of the model.(3)Finally,the framework proposed in this paper was trained and tested on the Drishti-GS1 retinal image standard dataset.In the Drishti-GS1 standard dataset,the desired results are obtained by applying RFC-Net to joint optic disc and cup segmentation.The F1 and BLE of optic disc segmentation are 0.9787 and 3.96,respectively,and the F1 and BLE of optic cup segmentation are 0.9058 and15.40,respectively.Secondly,applying SLAG-CNN to the joint optic disc and optic cup segmentation to obtain the desired result,the F1 and BLE of optic disc segmentation are0.9891 and 3.20,respectively,and the F1 and BLE of optic cup segmentation are 0.9029 and 14.01,respectively.Through comparison and analysis with the segmentation results of RFC-Net,the SLAG-CNN model further improves the accuracy of optic disc and optic cup segmentation.