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

摘要：这是R01项目的行政补充应用程序，标题为‘区分动-静脉 OCT血管造影对糖尿病视网膜病变客观分类的分析目的 本补充应用是为增强视网膜图像增加人工智能(AI)主题研究 构建和自动化动静脉分析。拟议的人工智能主题研究的潜在影响是 两方面：1)在眼睛健康方面推进人工智能技术；2)培养对人工智能技术感兴趣的研究生。 一名代表不足的美国出生的越南学生和两名来自欠发达地区的非洲/黑人学生 各国将参与这一项目。众所周知，糖尿病视网膜病变(DR)可以针对视网膜动脉和 不同的静脉。因此，动-静脉差值分析可以提供更好的DR检测和 分类。有了这种主动的R01支持，我们已经开发了算法来实现动脉-静脉的差异化 在OCTA分析中，为了更好地对DR进行临床管理，我们建议扩大这一补充应用程序 基于深度学习(DL)的人工智能方法，以促进差异动静脉分析的临床部署。这个 提出的人工智能研究课题自然会提供一个有用的平台，促进 生物医学工程和人工智能眼科专业的学生人数不足。第一个人工智能话题是转移 学习OCTA构建和DL动静脉分析。在传统OCTA机中，多幅OCT图像 需要采集，并使用后续的相关算法。然而，由于要求， 在多个图像采集中，OCTA的成像速度和分辨率/视场之间存在权衡。 我们最近已经证明了从单个体积的OCTA构造转移学习的可行性。 扫描动物(鼠)10月。我们建议验证基于DL的OCTA构建和动静脉分析 人视网膜的临床OCT。第二个人工智能主题是验证一种便携式高保真动脉眼部成像仪- 静脉成像和人工智能眼科。使用临床OCTA，我们已经演示了不同的动-静脉分析 用于改进DR检测。然而，临床OCTA机器通常体积庞大且价格昂贵，限制了其 在农村和服务不足地区申请远程医疗。此外，目前临床上可用的OCTA具有 视野(FOV)通常在10-20度内，相当于3-6毫米的视网膜区域。我们建议验证一个 具有AI迁移学习功能的便携式眼科成像仪，可用于毛细血管水平的高保真动静脉分析。这款便携的 低成本眼科成像仪基于我们最近展示的高动态范围(HDR)眼底相机和 AI主题1研究中的AI迁移学习结构。我们预计这款便携式眼科成像仪将提供 临床OCTA水平分辨率用于区分动-静脉分析，快照视野高达67o，对应 至20 mm的视网膜区域。便携式、宽视场、高保真眼科成像仪的成功验证将提供 经济实惠的解决方案，在家庭和家庭护理环境中促进人工智能远程医疗，从而减少 农村和服务不足地区眼科护理的差距。

Summary: This is an administrative supplement application for the R01 project, entitled ‘Differential artery-vein analysis in OCT angiography for objective classification of diabetic retinopathy’ (R01EY030842). The purpose of this supplement application is to add artificial intelligence (AI) topic research for enhanced retinal image construction and automated artery-vein analysis. Potential impact of the proposed AI topic research is twofold: 1) to advance AI technology in eye health; 2) to training graduate students interested in AI technology. One underrepresented American-born-Vietnamese student and two African/black students from underdeveloped countries will be involved in this project. It is known that diabetic retinopathy (DR) can target retinal arteries and veins differently. Therefore, differential artery-vein analysis can provide better performance of DR detection and classification. With this active R01 support, we have developed algorithms to achieve differential artery-vein analysis in OCTA for better clinical management of DR. For this supplement application, we propose to expand deep learning (DL) based AI approaches to foster clinical deployments of differential artery-vein analysis. The proposed AI research topics will naturally provide a useful platform to foster the education training of underrepresented students in biomedical engineering and AI ophthalmology. The first AI topic is transfer learning OCTA construction and DL artery-vein analysis. In traditional OCTA machine, multiple OCT image acquisitions are required, and subsequent correlation algorithms are employed. However, due to the requirement of multiple image acquisitions, there is a tradeoff between imaging speed and resolution/field-of-view in OCTA. We have recently demonstrated the feasibility of transfer learning OCTA construction from a single-volumetric- scan animal (mouse) OCT. We propose to validate DL based OCTA construction and artery-vein analysis using clinical OCT of human retina. The second AI topic is to validate a portable eye imager for high-fidelity artery- vein imaging and AI ophthalmology. Using clinical OCTA, we have demonstrated differential artery-vein analysis for improved DR detection. However, clinical OCTA machine is typically bulky and expensive, limiting their application for telemedicine in rural and underserved areas. Moreover, currently available clinical OCTA has a field of view (FOV) typically within 10-20o, corresponding to 3-6 mm retinal region. We propose to validate a portable eye imager with AI transfer learning for high-fidelity artery-vein analysis at capillary level. This portable, low-cost eye imager is based on our recently demonstrated high dynamic range (HDR) fundus camera and the AI transfer learning construction in AI topic 1 study. We anticipate that this portable eye imager will provide clinical OCTA level resolution for differential artery-vein analysis with a snapshot FOV up to 67o, corresponding to 20 mm retinal region. Successful validation of the portable, wide field, high-fidelity eye imager will provide an affordable solution to foster AI telemedicine at home and family care environments, and thus to reduce the disparity of eye care in rural and underserved areas.