Three-dimensional Reconstruction And Analysis Of Retinal Vessels Via OCTA 2D Depth Map Prediction

Optical Coherence Tomography Angiography(OCTA)is an emerging ophthalmic imaging technique that is capable of obtaining high-resolution three-dimensional(3D)retinal blood flow and structural information in a non-invasive way,and is now used in the diagnosis of many ocular diseases.Currently,most of the research work related to the morphological analyses of blood vessels on OCTA image is performed on its 2D en face image,ignoring the spatial information of blood vessels.Compared to 2D projection image analysis,three-dimensional analysis and visualization of blood vessels can provide additional spatial information to the physician.Therefore,3D reconstruction of the retinal vessels in the fundus is critical for the observation of morphological changes in the vessels.Existing OCTA image-based vascular reconstruction work achieves 3D reconstruction of retinal vessels by directly processing the raw 3D OCTA volume data.However,this method faces great challenges: in terms of technology,there are difficulties such as high complexity of 3D calculation and difficult removal of vascular artifacts;in terms of data,due to the manufacturer’s limitation,some OCTA imaging devices cannot export the original 3D data,which directly leads to the difficulty of 3D reconstruction of retinal vessels.And with the development of imaging technology,some OCTA devices have developed a depth map imaging function.Each pixel of the depth map indicates the distance between a point in the scene and the camera,which is an effective vehicle for recovering the spatial structure of blood vessels.Therefore,to address the difficulties at the technical level and data acquisition level of the above reconstruction methods,this thesis aims to recover the 3D structure of retinal vessels by using intelligent image processing algorithms combined with depth map prediction through 2D en face images of OCTA,and to assist subsequent clinical diagnosis by 3D structural analysis and feature quantification of the vessels.The main contents of this paper are as follows:(1)In this paper,a framework for 3D reconstruction of retinal vessels from 2D to3 D is constructed.Firstly,this paper proposes a structure-guided multi-scale OCTA depth prediction method SDE-Net,which combines pixel-and overall structure-level loss functions,and uses machine-taken depth map labels for supervised constraints,to perform depth estimation of the input OCTA en face images.By introducing structural constraints and a multi-scale depth aware module to fuse the structural prior of blood vessels and capture the multi-scale global information without losing feature resolution,this method shows superior performance on the OCTA depth map prediction task,which is significantly improved compared with other widely used methods.Next,based on the vessel segmentation information provided by the OCTA en face images and the spatial location information of the vessels obtained from the predicted depth maps,this paper successfully acquires the 3D centerline point clouds of the vessels and obtains the final 3D surface reconstruction results of the vessels by specific sampling and optimization algorithms.(2)A cross-domain adaptive OCTA image depth prediction model SDE+ is proposed in this work.In practical applications,the depth map is not yet available for OCTA devices produced by most manufactures,which makes it impossible to achieve3 D reconstruction of blood vessels directly from the depth map.In order to make SDE-Net applicable to OCTA images taken by different imaging devices(different domains),SDE+ enhances the generalization performance of the network by designing the domain adaptive module DAM based on SDE-Net for unsupervised learning characteristics.DAM effectively reduces the differences(noise,contrast,etc.)between different domain images and is able to self-correct for predicted outliers,to achieve unsupervised depth map prediction.SDE+ shows excellent results on the cross-domain OCTA depth map prediction task,with significant improvement compared to the comparison methods.(3)In this paper,feature extraction and clinical disease correlation of reconstructed retinal vessels are performed.Based on the OCTA image datasets containing diabetic retinopathy(DR)and healthy controls(HC),this paper first reconstructs the 3D structure of retinal vessels in OCTA images taken by different machines through the proposed SDE+ method,and extracts the structural features of vessels in 2D and 3D space,including vessel density,fractal dimension,vessel curvature,and carry out the correlation analysis between each feature and disease by statistical analysis.Finally,based on the extracted 2D and 3D vascular features,a support vector machine(SVM)classification experiment is conducted between the DR and HC groups.The experimental results of feature analysis and SVM classification show that compared with 2D features,3D vascular features significantly improve the accuracy of DR classification,reflecting the superiority of 3D structural features of retinal vessels in representing vessel specificity,which further verifies the importance of the proposed 3D vascular reconstruction method,and also reflects the clinical value of 3D retinal vascular analysis.