Research On Retinal Layer Segmentation In OCT Images Based On Convolutional Neural Network And Graph Search

Retinopathy is a common eye disease that seriously jeopardizes the human visual sensory system.With the increase of people's age,the incidence of retinopathy is gradually increasing,which has attracted more and more attention of researchers.Clinical studies have shown that many diseases can cause changes in the thickness of the retina,and even cause a certain cell layer of the retina to disappear.Optical coherence tomography(OCT)imaging technology can clearly display the various cell layers of the retina,and doctors can diagnose the disease by observing these subtle changes in the retinal layer.The precise segmentation of retinal layer in OCT images is an important prerequisite and guarantee for disease diagnosis.Retinal layer segmentation still poses significant challenges due to the low boundary contrast of certain retinal layers and the fact that the disease may cause deformation of the retinal layer.In order to more accurately segment the retinal layer,this paper uses advanced deep learning network model to carry out in-depth research on OCT image retinal layer segmentation.Two retinal layer segmentation methods based on convolutional neural network and graph search are proposed.The main work is as follows:1)In order to extract more discriminative comprehensive features to identify retinal layer boundaries and get a more accurate probability map,this paper proposes a multiscale convolutional neural network(MCNN),which incorporates multiscale feature information into the traditional CIFAR network.At the same time,in order to reduce the probability that the background pixels which are similar to the boundary pixels are discriminated as boundary pixels,this paper introduces position information into the background pixels during sampling.In addition,considering the horizontal characteristics of the retinal layer,this paper improves the graph model of the graph search,and change the original eight-neighbor undirected graph to the three-neighbor directed graph,effectively removing a large amount of redundant information and improving the detection efficiency of graph search.2)This paper proposes a spatial pyramid network(SP-Net)to extract feature information at multiple scales and obtain coarse segmentation results.Due to the imbalance in the ratio of retinal pixels to background pixels in OCT images,the network tends to learn about a larger number of samples.In order to enhance the learning of retinal layer by the network,this paper increases the weight of loss in retinal layer area to emphasize the learning of retinal layer by the network.In addition,considering that the segmentation result directly obtained by the network has defects such as boundary faults or holes,this paper uses the graph search algorithm to smooth the segmentation results.3)This paper uses a variety of evaluation indicators to evaluate the proposed methods from multiple perspectives on two publicly available datasets,and this paper compares the experimental results of the two algorithms with various state-of-the-art methods.The experimental results show that the proposed methods are superior to most state-of-the-art methods,and not only obtain the highest numerical results,but also obtain clear visual segmentation results.