Retinal Vessel Segmentation In OCTA Image Base On Deep Learning

Optical coherence tomography angiography(OCTA)technology has gradually been widely used in retinal examination,providing retinal and choroidal blood flow information.OCTA images reveal important details of eye diseases such as diabetic retinopathy(DR),glaucoma,and age-related macular degeneration(AMD).DR and AMD are the leading causes of blindness among these diseases.Accurate segmentation of retinal vessels in OCTA images is of great significance for subsequent diagnosis.In recent years,an automatic retinal vessel segmentation method based on deep learning has been proposed to assist clinical diagnosis.However,most segmentation methods are difficult to accurately segment retinal vessels with large scale changes,and are prone to loss of tiny vessels.In addition,retinopathy can cause deformation of retinal blood vessels,and some retinal blood vessels have low contrast with the background,which can easily cause missegmentation of retinal blood vessels.These issues pose challenges to the construction of retinal vessel segmentation algorithms.In this thesis,we propose two different deep learning-based retinal vessel segmentation methods to address these issues.(1)Aiming at the problem of low contrast between retinal vessels and background in OCTA images,this thesis constructs a new retinal vessel enhancement module based on direction consistency.This module aims to enhance blood vessel information at different scales and enrich blood vessel information of different sizes according to the directional consistency of blood vessels.Meanwhile,to guide the network to adapt to vessel scale changes,we propose a multi-scale vessel attention module.Furthermore,in order to learn the similarity of blood vessels and improve the ability of the network to recover high-level feature maps with low-level local high-resolution semantic information,we design a contrastive loss to learn the similarity of retinal vessels.Finally,the method is evaluated on the OCTA500 dataset and the ROSE-1 dataset,and the experimental results demonstrate the effectiveness of the proposed segmentation method.(2)Considering that OCTA volume data has richer blood vessel information and is only segmented on the projection map of OCTA,it is easily interfered by potential retinal diseases,resulting in inaccurate segmentation results.In this thesis,a new 3D and 2D combined OCTA retinal vessel segmentation method is designed.This method designs an adaptive projection learning module by which it undergoes dimension reduction operations while preserving as much vessel information as possible for the octa 3D data.In addition,a 2D concatenated vessel segmentation path is constructed to refine the results of coarse segmentation,leading to improved octa retinal vessel segmentation results.The method is evaluated on the OCTA500 dataset,and the experimental results demonstrate the effectiveness of the proposed segmentation method.