Differential artery-vein analysis in OCT angiography for objective classification of diabetic retinopathy

摘要：本计画旨在建立光学相干断层扫描之差分动脉-静脉分析 血管造影（OCTA），并验证全面的OCTA功能，用于糖尿病患者的自动分类。 视网膜病变（DR）。早期发现，及时干预和可靠的治疗结果评估是 这对于防止DR造成的不可逆视力丧失至关重要。众所周知，DR可以不同地靶向动脉和静脉。 因此，差分动脉-静脉分析可以提供更好的DR检测和分类性能。 然而，临床OCTA仪器缺乏区分动脉-静脉的能力。在这个项目中，我们 建议使用OCT的定量特征分析（与OCTA同时采集）来引导动脉- OCTA的静脉分化。第一个目标是在OCTA中建立自动化的动脉-静脉区分。在 与我们最近证明的血管跟踪技术、OCT强度/几何特征相协调 将用于自动引导OCTA中的动静脉区分。血液的动脉-静脉鉴别分析 血管迂曲度（BVT）、血管口径（BVC）、血管密度（BVD）、血管周长指数（VPI）， 血管分支系数（VBC）、血管分支角（θ）、分支宽度比（BWR）、无血管中心凹 区域面积（FAZ-A）和FAZ轮廓不规则性（FAZ-CI）。目标的关键成功标准 1项研究旨在证明OCTA中存在稳健的动静脉分化，并建立客观的OCTA特征 第二个目的是验证DR的自动OCTA分类。我们 建议采用集成机器学习来集成多个分类器以实现鲁棒的OCTA 目的2研究的关键成功标准是识别OCTA特征和最佳特征， 联合检测早期DR，并建立OCTA特征与临床 生物标志物。第三个目的是验证OCTA对DR治疗的预测和评价。我们初步的OCTA 抗血管内皮生长因子（抗VEGF）治疗糖尿病性黄斑水肿（DME）的研究 表明BVD可以作为预测视力改善生物标志物。在这个项目中，我们计划测试 用于DME治疗评价的差分动脉-静脉分析。目标3研究的关键成功标准是 识别动脉-静脉特征以提供DME治疗结果的稳健预测和评估。作为 作为另一种方法，我们提出了一种基于深度机器学习的全卷积神经网络（FCNN）。 动-静脉和DR分类。FCNN中的早期层将产生简单的特征，这些特征将被卷积 并被过滤到更深层以产生用于动脉-静脉和DR分类的复杂特征。进一步 研究通过机器学习过程学习的新特征之间的关系， 临床生物标志物将使我们能够优化设计，以实现更好的DR分类。该项目的成功将 为使用定量OCTA特征进行早期DR检测、客观预测和 评估治疗结果。

Abstract: This project aims to establish differential artery-vein analysis in optical coherence tomography angiography (OCTA), and to validate comprehensive OCTA features for automated classification of diabetic retinopathy (DR). Early detection, prompt intervention, and reliable assessment of treatment outcomes are essential to prevent irreversible visual loss from DR. It is known that DR can target arteries and veins differently. Therefore, differential artery-vein analysis can provide better performance of DR detection and classification. However, clinical OCTA instruments lack the capability of artery-vein differentiation. During this project, we propose to use quantitative feature analysis of OCT, which is concurrently captured with OCTA, to guide artery- vein differentiation in OCTA. The first aim is to establish automated artery-vein differentiation in OCTA. In coordination with our recently demonstrated blood vessel tracking technique, OCT intensity/geometry features will be used to guide artery-vein differentiation in OCTA automatically. Differential artery-vein analysis of blood vessel tortuosity (BVT), blood vessel caliber (BVC), blood vessel density (BVD), vessel perimeter index (VPI), vessel branching coefficient (VBC), vessel branching angle (VBA), branching width ratio (BWR), fovea avascular zone area (FAZ-A) and FAZ contour irregularity (FAZ-CI) will be implemented. Key success criterion of the aim 1 study is to demonstrate robust artery-vein differentiation in OCTA, and to establish OCTA features for objective detection and classification of DR. The second aim is to validate automated OCTA classification of DR. We propose to employ ensemble machine learning to integrate multiple classifiers to achieve robust OCTA classification of DR. Key success criterion of the aim 2 study is to identify OCTA features and optimal-feature- combination to detect early DR, and to establish the correlations between the OCTA features and clinical biomarkers. The third aim is to verify OCTA prediction and evaluation of DR treatment. Our preliminary OCTA study of diabetic macular edema (DME) with anti-vascular endothelial growth factor (anti-VEGF) treatment has shown that BVD can serve as a biomarker predictive of visual improvement. During this project, we plan to test differential artery-vein analysis for DME treatment evaluation. Key success criterion of the aim 3 study is to identify artery-vein features to provide robust prediction and evaluation of DME treatment outcomes. As an alternative approach, we propose a fully convolutional neural network (FCNN) for deep machine leaning based artery-vein and DR classification. Early layers in the FCNN will produce simple features, which will be convolved and filtered into deeper layers to produce complex features for artery-vein and DR classification. Further investigation of the relationship between the new features learned through the machine learning process and clinical biomarkers will allow us to optimize the design for better DR classification. Success of this project will pave the way towards using quantitative OCTA features for early DR detection, objective prediction and assessment of treatment outcomes.