Research On OCTA Retinal Image Analysis Based On Deep Learning

Optical coherence tomography angiography(OCTA)is a novel imaging technology developed based on optical coherence tomography(OCT).OCTA can quickly and noninvasively generate three-dimensional blood flow images,and is an important tool for clinical detection of retinal blood flow function.As a newly developed imaging technology,OCTA has received extensive attention from the academic community.At present,OCTA has been applied to the clinical diagnosis of age-related macular disease,diabetes retinopathy,choroidal neovascularization,glaucoma,and other diseases.However,due to the late start and high price,the data resources of OCTA images are still scarce.The technical means of its quantitative analysis are still in the development stage,and its potential in clinical application has not been fully explored.In order to promote the development of OCTA image analysis technology,based on deep learning technology,this paper conducts OCTA research around data,methods,and applications.The main work of this paper is summarized as follows:(1)OCTA-500 dataset is proposed as the largest and most comprehensive publicly available OCTA dataset.The dataset contains OCTA imaging data from 500 subjects with rich information and annotation,including two imaging modalities(OCT/OCTA),six types of projection maps,four types of text labels(age/sex/eye/ disease),and seven types of segmentation labels(large vessels/arteries/veins/capillaries/2D foveal avascular zone/3D foveal avascular zone/layer segmentation).The OCTA-500 dataset provides the key data for the methodological study in this paper.(2)A novel method of foveal center localization is proposed which is based on foveal avascular zone segmentation.The foveal center is an important biomarker on the retina,and determining the location of the foveal center is one of the prerequisites for analyzing retinal images.In this paper,it is pointed out for the first time that the position of the foveal center can be determined utilizing the foveal avascular zone.A lightweight model for foveal avascular zone segmentation is designed,and different segmentation strategies are given for OCT and OCTA images.The location of the foveal center is determined by calculating the geometric center of the foveal avascular zone.Experiments on multimodal datasets containing various retinal diseases showed that the proposed framework can quickly and robustly locate the macular foveal center.(3)An image projection network(IPN)is proposed to segment the retinal vessel and the foveal avascular zone in OCTA images.The image projection network is a 3D-to-2D segmentation network,which takes a three-dimensional image as input and outputs twodimensional segmentation results end-to-end.It brings two benefits: no reliance on retinal layer segmentation;and better utilization of spatial information in 3D images.The proposed image projection network consists of several projection learning modules,which implement feature selection through convolutional operations and 3D-to-2D scale compression through unidirectional pooling operations.The image projection network enables high-performance vessel segmentation and foveal avascular zone segmentation.(4)A capillary segmentation framework based on label adversarial learning is proposed.We first implemented a preliminary capillary segmentation using an image magnification network,and then used label adversarial learning for further optimization of the capillary segmentation.The image magnification network uses an up-sampling encoder and downsampling decoder to capture relatively low-order image details,thus avoiding the loss of fine capillary structures.Label adversarial learning achieves adjustable vessel segmentation through the design of label adversarial loss and adjustment layers.The continuous segmentation process from label adversarial learning can recommend capillary segmentation results with clearer topology and denoising with the help of uncertainty maps.The proposed framework enables high-performance capillary segmentation.(5)A joint Capillary-Artery-Vein-FAZ segmentation task is proposed.In this paper,capillary segmentation,artery segmentation,vein segmentation and foveal avascular zone segmentation are integrated under one segmentation framework,which is called CapillaryArtery-Vein-FAZ segmentation.The proposed joint segmentation task brings convenience and new challenges to the quantitative analysis of OCTA images.Based on the joint segmentation task,we optimized the spatial structure of the image projection network,and we also explored the effects of different factors on the joint segmentation task,such as number of training samples,input modalities,baselines,retinal diseases.(6)We explored the retinal differences between autism spectrum disorder children and typical development children.Based on the above image analysis techniques,the retinal structure and function of autism spectrum disorder children and typical development children were analyzed and compared.Several new findings may become typical characteristics of autism spectrum disorder children: thicker ellipsoid zone thickness;fewer arteriovenous of the inner retina;and lateralization in the left eye.In addition,we preliminarily explored whether artificial intelligence can correctly classify the retinal images of children with autism spectrum disorder and typical development.