Deep-learning based phenotyping and prediction of disease progression in geographic atrophy secondaryto age-related macular degeneration

视网膜相关性黄斑变性（AMD）是当今发达国家中心视力丧失的最常见原因。对于非新生血管的晚期表现，“地理萎缩”（GA），尚未建立治疗方法。作为临床试验和个性化医疗的先决条件，我们的目标是（i.）改进的成像生物标志物的定量，（ii.）无偏见的、数据驱动的表型鉴定，以及（iii.）功能丧失的预测。计划的分析将使用DFG资助的研究地理性萎缩中的定向扩散的最新成像数据进行。定量成像实验室（PI：Prof.丹尼尔L。鲁宾，斯坦福大学）将提供最新的基于深度学习的方法的应用的专业知识。拟议的项目分为三个工作包（WP）：在WP 1中，我们的目标是基于深度学习的GA分割和GA进展的预测。我们的目标是在光学相干断层扫描（OCT），包括微分玻璃疣表型，外视网膜管（ORT）和弗里德曼脂质球，这通常只被定性描述的情况下，GA的高阶结构特征的全自动识别和定量。在WP 2中，我们的目标是建立一个无偏见的，数据驱动的方法，以表型发现的基础上，使用自动编码器架构的高分辨率OCT图像数据。通过与已知发病机制的单基因疾病的临床数据进行系统比较，我们打算确定AMD谱中的离散亚型。与继发于Bruch膜改变的已知单基因疾病的特征重叠（弹性假黄瘤、Sorsby眼底营养不良、迟发性视网膜变性）和/或光感受器/色素上皮水平的氧化应激（Stargardt病，中央晕状脉络膜营养不良）可以促进未来基因型特异性治疗方法的发展。在WP 3中，目标是预测随着时间的推移由于GA导致的视力丧失，基于结构成像数据。GA患者的视觉功能评估往往是耗时的-特别是眼底控制视野检查（FCP;也称为“微视野检查”）。视网膜功能的空间分辨预测和映射将使得能够实现基于SD-OCT的“准功能”临床研究终点，并促进关于驾驶执照条例和临床“低视力”康复的临床评估。

Age-related macular degeneration (AMD) is the most common cause of central vision loss in developed countries today. For the non-neovascular late-stage manifestation, "geographical atrophy" (GA), no therapy has been established yet. As a prerequisite for clinical trials and personalized medicine we aim at (i.) an improved quantification of imaging biomarkers, (ii.) an unbiased, data-driven identification of phenotypes and (iii.) the prediction of functional loss. The planned analyses will be performed with state-of-the-art imaging data from the DFG-funded study Directional Spread in Geographic Atrophy. The Laboratory of Quantitative Imaging (PI: Prof. Daniel L. Rubin, Stanford University) will provide the expertise for the application of the latest deep learning based methods. The proposed project is divided into three work packages (WP):In WP1 we aim at deep-learning-based segmentation of GA and prediction of GA progression. Our goal is the fully automated identification and quantification of higher order structural features in optical coherence tomography (OCT), including differential drusen phenotypes, outer retinal tubulations (ORT) and Friedman-lipid globules, which have usually only been described qualitatively in the context of GA. This would enable a stratified selection of patients for future therapeutic trials.In WP2, we aim to establish an unbiased, data-driven approach to phenotype discovery based on high-resolution OCT image data using an auto-encoder architecture. Through systematic comparison with clinical data of monogenic diseases with known pathogenesis, we intend to identify discrete subtypes in the spectrum of AMD. Feature overlap with known monogenic diseases secondary to Bruch’s membrane alterations (Pseudoxanthoma elasticum, Sorsby Fundus Dystrophy, Late-onset retinal degeneration) and/or oxidative stress at the photoreceptor/pigment epithelium level (Stargardt disease, Central areolar choroidal dystrophy) could facilitate the development of future genotype-specific therapeutic approaches.In WP3, the goal is to predict the loss of vision due to GA over time, based on structural imaging data. The assessment of visual function in patients with GA tends to be time-consuming - especially fundus-controlled perimetry (FCP; also called "microperimetry"). The spatially resolved prediction and mapping of retinal function would enable the implementation of SD-OCT-based "quasi-functional" clinical study endpoints and facilitate clinical assessment with regard to the Driver's License Ordinance and clinical "low vision" rehabilitation.