Three-dimensional Segmentation And Application Of Foveal Avascular Zone

The foveal avascular zone(FAZ)is highly sensitive to ischemic phenomena,and thus its measurement parameters are often used as landmark indicators of retinal vascular pathology.The application of optical coherence tomography angiography(OCTA)in clinical ophthalmology facilitates the study of FAZ morphology in three dimensions.Quantitative analysis and applied research on 3D FAZ using image processing and machine learning methods are of great clinical and scientific significance.In this paper,the FAZ volume is firstly proposed as a clinical index for FAZ measurement as well as its calculation method.Then the statistical significance is analysed.After that,this paper presents an automatic segmentation algorithm for the FAZ in a 3D manner for the first time.The contents of this paper is summarized as follows:(1)In order to achieve FAZ 3D quantization in OCTA data,SD-OCT layer segmentation and OCTA projection artefacts suppression are performed.To address the problems of the raw OCTA data,a 3D U-Net layer segmentation model based on multi-modal is adopted to obtain the layer information,so as to limit the region of interest of OCTA;The principles of OCTA artefacts suppression are explored and the mainstream projection artefacts suppression algorithm is reproduced for the subsequent FAZ segmentation work in advance.(2)A new clinical assessment index,FAZ volume,is proposed.The metric aims to measure FAZ in three dimensions and thus uncover more comprehensive and objective information about FAZ.Firstly,a FAZ quantification method in three dimensions is introduced;To clarify the statistical significance of the new index and its clinical significance,correlation analysis and difference significance analysis are performed on the datasets: for normal eyes,the correlation of FAZ volume with central macular thickness(CMT)and FAZ area are explored.For High myopia(HM)and diabetic retinopathy(DR)eyes,FAZ volume is compared with control values,and FAZ volume is found to be more sensitive to vascular alterations than the conventional parameter FAZ area.Finally,it is concluded that the FAZ volume has advantages over two-dimensional metrics in characterizing the size and shape of the FAZ.(3)An automatic 3D segmentation algorithm for FAZ guided by priors is proposed.The algorithm incorporates priors in the training phase of the network according to the data characteristics and adopts tailor-made designs in data augmentation,model construction,and loss function design: the input image is trained in a 3D atten-UNet containing an attention gating module with the non-local blocks to obtain a FAZ segmentation probability map.A persistence homotopy barcode is constructed for the probability map,and the topological structure of the segmentation result is constrained by using the sum of bar lengths in each dimension as a metric and calculating topological loss with the topology prior.The combination of region and topological loss is used as the network loss function to supervise the network training.Finally,the network output segmentation results are post-processed to adapt to the original layer structure to obtain the final segmentation results of FAZ.The multi-fold cross-validation experiments show that the proposed algorithm can segment the 3D FAZ accurately and robustly,and the proposed algorithm has obvious accuracy advantages compared with the current mainstream 3D segmentation algorithms.The ablation experiments show that the three proposed designs of random center cropping,non-local attention gating module and topological consistency loss have synergistic and positive effects on improving both the region similarity and boundary accuracy of segmentation results.