Research On High-precision Slice Method For Retinal OCT 3D Images

OCT examination of the retina is a routine diagnostic method of ophthalmology,the main purpose is to diagnose and early prevention of fundus diseases.The doctor observes the retinal morphology of the patient through OCT images and evaluates the retinal morphology.In the process,it needs to involve the segmentation and quantitative analysis of the retinal layer structure.This is a time-consuming,tedious,and highly repetitive work,and will be subject to The influence of subjective factors such as experience,attention,and fatigue.At present,the retinal interlayer structure marking function attached to the OCT system used in hospitals cannot meet the standards of clinical application,and the existing retinal OCT segmentation theory is backward in theory,the method is not intelligent,and the segmentation is inaccurate,so there is an urgent need for a reliable High-precision layer cutting method to assist doctors in diagnosis.In this paper,the high-precision layer cutting method of retinal OCT threedimensional images is studied.For two different data types,two corresponding retinal OCT image segmentation methods based on deep learning are proposed.The main research content of this article has the following four aspects:1.Established a retina OCT imaging standard database,which has a total of 11,200 images in the 2D OCT data set and 11,520 images in the 3D OCT data set,and includes an auxiliary calibration software,which provides a reliable data source for subsequent research in this article.2.Implemented a deep segmentation network for retinal OCT two-dimensional images,using an end-to-end structure and a lightweight design,including a Dense Block module with a bottleneck,and integrating multi-scale interactive information through Skip Connection.The accuracy,precision and sensitivity of segmentation of each layer of 2D OCT samples have reached 93.1%,88.5% and 90.3% respectively,and the overall segmentation accuracy,precision and sensitivity have reached 94.7%,90.3% and 92.2%,respectively,when using GPU The average prediction speed is 0.45 seconds / sheet,and various performance indicators fully surpass traditional image processing methods and classic deep learning methods.3.Implemented a 3D segmentation network for retinal OCT 3D images,used 3D convolution to extract volume data spatial information,adopted Encoder-Decoder structure and Skip Connection mode,while realizing low computing resource occupation,the 4-layer segmentation was average and accurate The rate and accuracy reached 0.955 and 0.977;the average accuracy and accuracy of the 9-layer segmentation reached 0.932 and 0.979.And the prediction speed of less than 0.152 seconds per slice is achieved,which is better than the twodimensional segmentation network.4.Explore the initial application of retinal OCT image segmentation.By using the 3D reconstruction technology and taking advantage of the spatial context of the 3D OCT image to realize volume data reconstruction,we have found the foothold of the OCT image segmentation technology of the retina.