Deep Learning Approaches to Detect Glaucoma and Predict Progression from Spectral Domain Optical Coherence Tomography

项目摘要/摘要 原发性开角型青光眼(POAG)是美国和世界范围内导致失明的主要原因。它是 据估计，超过220万美国人患有POAG，超过13万人在法律上因 疾病。随着人口老龄化，美国的POAG人数将增加到3.3人以上 到2020年，全球将达到1.118亿人，到2040年，估计将达到1.118亿人。POAG是一种进行性疾病，与 具有特征性的功能和结构变化，临床医生用来诊断和监测疾病。 在过去的几年里，光谱域光学相干层析成像(SDOCT)已经成为标准 用于测量POAG中结构的工具。这种3D成像模式提供了关于视网膜的丰富信息 结构和与POAG相关的视网膜层。如此大量的数据对于临床医生来说很难解释和使用 有效地帮助指导治疗决策。取而代之的是诸如平均层厚度等汇总指标 用于将SDOCT图像缩减为少数几个值。虽然这些指标很有用，但它们可能很难 解释并丢弃有关体素强度和纹理的重要信息， 视网膜层，以及视网膜的整体3D结构。过于依赖这些指标限制了我们的能力 更深入地了解结构对POAG的贡献，结构与视觉之间的关系 功能，以及结构(和功能)如何改变POAG的进展。人工智能的最新进展 然而，智能和深度学习为解释3D SDOCT提供了新的数据驱动工具和技术 并从世界各地诊所收集的大型SDOCT数据集中学习。这项提议将 将最先进的深度学习技术应用于3D SDOCT数据，以便(1)开发更准确的 POAG检测工具，(2)揭示结构-功能关系，以及(3)预测结构和 开角型青光眼的功能进展。 该建议还详细说明了一项培训计划，以帮助PI从博士后学者过渡到独立学者 研究员。该奖项的指导阶段将由主要导师Linda Zangwill博士监督，以及 包括Robert Weinreb博士(眼科)、David Kriegman博士在内的多学科指导团队 (计算机科学与工程)和Armin Schwartzman博士(生物统计学)。执行建议的 研究、正式的课程工作和有指导的职业发展将为PI提供备受欢迎的- 在技能和经验之后，帮助确保成功过渡到独立。

Project Abstract / Summary Primary open angle glaucoma (POAG) is a leading cause of blindness in the United States and worldwide. It is estimated that over 2.2 million Americans suffer from POAG and that over 130,000 are legally blind from the disease. As the population ages, the number of people with POAG in the United States will increase to over 3.3 million in 2020 and worldwide to an estimated 111.8 million by 2040. POAG is a progressive disease associated with characteristic functional and structural changes that clinicians use to diagnose and monitor the disease. Over the past several years, spectral domain optical coherent tomography (SDOCT) has become the standard tool for measuring structure in POAG. This 3D imaging modality provides a wealth of information about retinal structure and POAG-related retinal layers. This large amount of data is hard for clinicians to interpret and use effectively to help guide treatment decisions. Instead, summary metrics such as average layer thicknesses are used to reduce SDOCT images to a handful of values. While these metrics are useful, they can be difficult to interpret and they throwaway important information regarding voxel intensity and texture, relationships across retinal layers, and the overall 3D structure of the retina. Relying too heavily on these metrics limits our ability to gain a deeper understanding structural contributions to POAG, the relationship between structure and visual function, and how structural (and functional) changes progress in POAG. Recent advances in artificial intelligence and deep learning, however, offer new data-driven tools and techniques to interpret 3D SDOCT images and learn from the large SDOCT datasets being collected in clinics around the world. This proposal will apply state-of-the-art deep learning techniques to 3D SDOCT data in order to (1) develop more accurate POAG detection tools, (2) reveal structure-function relationships, and (3) predict structural and functional progression in POAG. This proposal also details a training plan to help the PI transition from a postdoctoral scholar to an independent researcher. The mentored phase of this award will be supervised by the primary mentor, Dr. Linda Zangwill, and a multidisciplinary mentoring team including Dr. Robert Weinreb (Ophthalmology), Dr. David Kriegman (Computer Science and Engineering), and Dr. Armin Schwartzman (Biostatistics). Performing the proposed research, formal coursework, and mentored career development will the provide the PI with highly sought- after skills and experience to help ensure a successful transition into independence.