Early Detection of Progressive Visual Loss in Glaucoma Using Deep Learning

青光眼是导致不可逆失明的主要原因，对退伍军人的影响不成比例。虽然经常进步， 缓慢地，青光眼也可以迅速进展，特别是考虑到标准视野（VF）测试的可变性 为了监测进展，目前确定那些需要更积极治疗的个体可能具有挑战性。 治疗方案退伍军人可能会经历永久性视力丧失（以及相应的视觉相关质量）。 生活），同时等待后续测试以显示VF丧失进展（并因此指示治疗变化 需要）。结构光学相干断层扫描（OCT）测量，如黄斑厚度 神经节细胞层（GCL）、视网膜神经纤维层（RNFL）和视盘形态也可用于帮助 监测进展。然而，现有的临床使用全局参数来评估青光眼进展可能 对焦点缺陷的恶化不敏感。也不知道不同的空间模式的进展 影响生活质量。对于简单易用的方法来更准确地估计 未来的进展和相应的生活质量措施。我们将使用特定类型的深度学习 方法，称为深度变分自动编码器（VAE），以提供一种新的标准化和灵敏的方法 监测青光眼的进展，堪比青光眼专家。我们的具体目标如下： 1.评估基于图像的深度学习变分自动编码器（VAE）模型的使用效果 来监测病人目前的昏迷进展这一目标首先涉及培训， 评估每个基于图像的感兴趣结构的单独的深度VAE模型以及深度VAE 24-2视野阈值数据的模型。一旦经过训练，每个VAE模型将允许提取 给定输入图像的所谓潜在变量值。这些潜在变量值监控 将（在一个独立的测试组中）将随时间变化的数据与标准的全球和区域参数进行比较。 由于潜变量方法能够自然地捕获全局和局部变化， 与当前的临床报告相比，能够更好地检测随时间的变化。 2.评估基于图像的深度学习变分自动编码器（VAE）模型的使用效果 来预测病人未来的昏迷进展为此，我们将首先制定一种方法， 用于基于从先前时间的学习来预测结构/功能的未来潜在变量表示 价值观系列。一旦确定，未来的潜在值将映射回其原始值 使用VAE的训练的“解码器”部分的结构/功能表示。这种做法将 为临床医生提供了具有未来结构的视觉空间表示的明显优势， 功能轨迹，以优化早期治疗决策。 3.评估新型双眼VAE模型的潜在变量与视觉生活质量的关系 措施在这个目标中，我们将首先开发一个额外的VAE模型，以考虑双目视觉 (what患者睁开双眼观看），然后将每个模型的潜在因素与 生活以横截面的方式进行测量。我们假设，双眼VAE模型的结构和 功能将比目前的方法更能预测视觉生活质量指标，有助于 优先考虑并指导治疗。 这些目标的成功完成预计将对帮助退伍军人昏迷产生积极影响。 患者在早期疾病阶段避免了永久性视力丧失，并保持了与视力相关的生活质量。

Glaucoma, a leading cause of irreversible blindness, disproportionately affects veterans. While often progressing slowly, glaucoma can also progress rapidly, and especially given the variability of standard visual-field (VF) tests to monitor progression, it currently can be challenging to determine those individuals needing a more aggressive treatment plan. Veterans may experience permanent loss of vision (and corresponding vision-related quality of life) while waiting for subsequent tests to show VF loss progression (and thus indicating a change in treatment is needed). Structural optical coherence tomography (OCT) measures, such as the thickness of the macular ganglion cell layer (GCL), retinal nerve fiber layer (RNFL) and optic disc morphology can also be used to help monitor progression. However, existing clinical use of global parameters to assess glaucoma progression may be insensitive to worsening of focal defects. It is also not known how differing spatial patterns of progression affects quality of life. There is an unmet clinical need for simple-to-use approaches to more accurately estimate future progression and corresponding quality-of-life measures. We will use a specific type of deep-learning approach, called deep variational autoencoders (VAEs) to provide a novel standardized and sensitive approach to monitoring glaucomatous progression, comparable to a glaucoma expert. Our specific aims are as follows: 1. Evaluate how well image-based deep-learning variational autoencoder (VAE) models can be used to monitor a patient’s current glaucomatous progression. This aim will first involve training and evaluating a separate deep VAE model for each image-based structure of interest as well as a deep VAE model for 24-2 visual field threshold data. Once trained, each VAE model will allow for the extraction of the so-called latent variable values given the input image. The ability of these latent variable values to monitor change over time will be compared (in an independent test set) to standard global and regional parameters. Because of their ability to naturally capture both global and local changes, the latent-variable approach will be able to better detect changes over time compared to current clinical reports. 2. Evaluate how well image-based deep-learning variational autoencoder (VAE) models can be used to predict a patient’s future glaucomatous progression. In this aim, we will first develop an approach for predicting future latent-variable representations of structure/function based on learning from a prior time series of values. Once determined, future latent values will be mapped back to their original structure/function representations using the trained “decoder” part of the VAE. Such an approach will provide a clear advantage for a clinician in having visual spatial representations of future structure and function trajectories to optimize early treatment decisions. 3. Evaluate how latent variables from a novel binocular VAE model relate to visual quality-of-life measures. In this aim, we will first develop an additional VAE model to take into account binocular vision (what the patient sees with both eyes open) and then relate latent factors from each model to quality-of- life measures in a cross-sectional fashion. We hypothesize that binocular VAE models of structure and function will be more predictive of visual quality-of-life measures than current methods, helping to prioritize and guide treatment. Successful completion of these aims is expected to have positive impact to help veteran glaucomatous patients avoid permanent vision loss at an early disease stage and maintain vision-related quality of life.