Research On OCT Image Retinal Layer Boundary Segmentation Based On Embedded Residual Networks

Optical coherence tomography(OCT)has come to the attention of the public and researchers with the increasing attention to eye health.The researchers usually measure the thickness change of the retinal layer by dividing the boundary of the retinal layer,to better assist the medical staff in further analysis.However,artificial segmentation of retinal layer boundaries in OCT images is a time-consuming and subjective task,which seriously affects the efficiency and accuracy of clinical diagnosis.Due to the influence of retinopathy and image noise in the research process,OCT image retinal layer boundary segmentation is still very challenging.In this paper,two different methods based on the embedded residual network are designed for retinal layer boundary segmentation.The main work is as follows:1)This paper proposes a method of retinal layer boundary segmentation based on Recurrent Residual network(RR-Net).It combines the advantages of residual network and recurrent network effectively.Specifically,RR-Net uses the residual network as the backbone network to protect the integrity of the underlying information transmission.In addition,recurrent blocks are embedded at each stage of the backbone network to make full use of the relationship between local regions of the image and extract more distinctive retinal layer boundary features.Secondly,because of the continuity and smoothness of the graph search algorithm,the results of retinal layer boundary segmentation can be further optimized.In this paper,a graph search algorithm is used to refine the results of layer boundary segmentation.2)This paper proposes a retinal layer boundary segmentation result based on the Attention Global Residual Network(AGR-Net).Compared with the previous model,this method takes into account the global information features of the image more comprehensively and redesigns the original recurrent block into a global feature module.Secondly,to make the network more focused on finding significant useful information related to segmentation results in retinal OCT images,to improve the performance of retinal layer boundary segmentation,the channel attention module is embedded into the network.Finally,the network results of AGR-Net are further optimized by graph search,to remove the isolated regions in the network segmentation results.3)This paper evaluates the proposed method from different aspects and compare the results with those of the state-of-the-art approaches on three publicly available datasets,which have different sizes,characteristics,and segmentation difficulties.The quantitative results and visual effects show that our method not only achieves the best segmentation accuracy but also has good stability.