Research On Localization And Segmentation Of Optic Disc In Fundus Images Based On Deep Convolutional Neural Networks

With the development of medical imaging technology,the processing and analysis of ocular images has become an important way to diagnose ocular diseases,which results in the continuous development of computer aided diagnosis techniques based on ocular image.The traditional ocular image processing methods rely on manual design of task-related features.The process of feature design is relatively complex and the extracted features are less robust to different images.In recent years,in many fields,such as computer vision and speech processing,deep learning has made great performance breakthroughs by its advantage of ability to automatically learn task-related high-level features from the data.Therefore,more and more researchers applied deep learning to the processing and analysis of medical images,and achieved successful application results.In this paper,deep learning method is used to learn task-related features from retinal fundus images automatically,which contributes to localization and segmentation of optic disc in fundus images.The main works of this paper are as follows:First,we propose a cascading method based on deep convolutional neural network to automatically localize optic disc in retinal fundus images.By combining saliency map and deep convolution neural network,the accuracy of optic disc localization is effectively improved.Specifically,the corresponding saliency map is computed from the fundus image first and the most salient region is found in the saliency map.Then the corresponding region is extracted from fundus image as a candidate region for optic disc localization,and the deep convolution neural network is used to extract the image features of the candidate region for determining if the region contains optic disc.If it is classified as a non-optic-disc area,we return to the saliency map to find the salient areas and input it into convolution neural network to classify again.The loop ends when the network finds a region with optic disc,and the corresponding area is output as the final localization result.The experimental results show that compared with the traditional optic disc localization method,the proposed method has higher localization accuracy and faster localization speed.And the proposed deep convolution neural network can learn image features more conducive to the classification of optic disc than traditional feature extraction methods.In addition,for the glaucoma disease,this paper implements an end-to-end optic disc and cup segmentation method based on deep convolutional neural networks.The optic disc and cup features are extracted from input images through the deep network and the corresponding images of segmentation results are output.The vertical cup to disc ratio is calculated as an important reference for diagnosis of glaucoma.Specifically,the blood vessel is removed from the patch of optic disc localization to reduce the influence on segmentation result,and then the blood vessel removed image is input to deep convolutional network for optic disc segmentation to obtain the corresponding optic disc segmentation image.The distance is calculated from each pixel to optic disc center of segmentation result,and the distance information is fused into input image to obtain segmentation result of optic cup.Segmentation edge is smoothed by ellipse fitting and finally the vertical cup to disc ratio is calculated for glaucoma diagnosis according to the segmentation results.The experimental results show that compared with the traditional segmentation method for optic disc and cup,in order to improve segmentation accuracy,the implemented deep convolution neural network can effectively learn the features of optic disc and cup which is advantageous to segmentation.The implemented method reached performance close to the artificial segmentation for glaucoma diagnosis.