Discovery and Validation of AMD Biomarkers for Progression Using Deep Learning

项目摘要 视网膜相关性黄斑变性（AMD）是老年人失明的主要原因。 目前还没有被证实有效的治疗晚期非新生血管性AMD的疗法，称为“非血管性AMD”。 地图样萎缩（GA）。较早的干预可能是可取的，但这将需要确定那些 具有进展为萎缩的最高风险的个体。在过去的二十年里，各种研究 包括我们的研究已经确定了几个基于光学相干断层扫描（OCT）的因素，这些因素似乎 与AMD进展的高风险相关。 该提案的中心假设是，深度学习-人工智能（AI）结构可以 客观自动地学习和量化最重要的风险因素，从而更好地预测 AMD进展风险高于当前主观指定的特征。在这个计划中，我们将首先开发一个AI- 基于系统自动识别“主观指定”的高风险因素，基于个人光谱 域（SD）OCT 2D扫描，并自动分割GA（AMD的终末期结局变量）， OCT 2D en面图。随后，作为我们假设的概念验证研究，我们将应用基于AI的 “反向学习”方法客观地学习和识别纵向OCT数据中的AMD高危因素。 为了实现这些目标，我们将努力实现以下具体目标： 目标1：开发并验证一种AI方法，将单个OCT 2D扫描分类为包含或不包含 包含预先规定的风险因素。在我们之前的工作中，我们手动识别了 不存在预先特定的高风险因素，并根据整个OCT量分配风险评分。等 这种方法既费时又不精确。在这个提议中，将应用AI算法来检测 个体OCT扫描的高风险因素。因此，评分系统的精度可以大大提高。 增强了高计算复杂度。 目标2：开发并验证一种AI方法，从2D OCT en face map中分割GA病变。用于OCT 体积有萎缩，我们将执行GA分割从脉络膜高透射导致的 使用多尺度CNN的en脸地图。 目标3：开发和验证人工智能“反向学习”方法，使用 纵向OCT数据。“反向学习”将基于多个CNN，然后是去 卷积网络来客观地识别高风险因素。我们之前的评分系统 通过潜在的包含（或替代）来自我们的新风险因素进行改进和优化。 客观的AI方法。 本提案中的工作将在来自图像数据池的SD-OCT图像中回顾性执行， 多PI萨达已聚集多年的Doheny图像阅读中心主任。

Project Abstract Age-related macular degeneration (AMD) is the leading cause of blindness among elderly individuals. Currently there are no proven effective therapies for treatment of advanced non-neovascular AMD, termed geographic atrophy (GA). Earlier intervention may be preferable, but this would require identification of those individuals with the highest risk for progression to atrophy. Over the last two decades, various studies including ours have identified several optical coherence tomography (OCT)-based factors that appear to associate with a higher risk for AMD progression. The central hypothesis of this proposal is that a deep learning - artificial intelligence (AI) construct can objectively and automatically learn and quantify the most important risk factors, yielding a better prediction of AMD progression risk than current subjectively specified features. In this proposal, we will first develop an AI- based system to automatically identify the “subjectively-specified” high risk factors based on individual spectral domain (SD) OCT 2D scans, and to automatically segment GA (the end-stage outcome variable of AMD) in OCT 2D en face maps. Subsequently, as a proof-of-concept study of our hypothesis, we will apply an AI-based “reverse learning” approach to objectively learn and identify AMD high risk factors in longitudinal OCT data. To achieve these objectives, we will pursue the following specific aims: Aim 1: Develop and validate an AI approach to classify individual OCT 2D scans as containing or not containing the pre-specified risk factor(s). In our previous work, we manually identified the presence or absence of the pre-specific high-risk factors and assigned to a risk score based on the entire OCT volume. Such approach was time consuming and not precise. In this proposal, an AI algorithm will be applied to detect the high risk factors from individual OCT scans. Hence, the precision of the scoring system can be greatly enhanced with high computational complexity. Aim 2: Develop and validate an AI approach to segment GA lesions from 2D OCT en face maps. For the OCT volumes having atrophy, we will perform the GA segmentation from the choroidal hypertransmission-resulted en face map using the multi-scale CNNs. Aim 3: Develop and validate an AI “reverse learning” approach to objectively identify the high risk factors using longitudinal OCT data. The “reverse learning” will be based on the multiple CNNs, followed by de- convolutional networks to identify the high risk factors objectively. Our previous scoring system will possibly be refined and optimized by the potential inclusion (or substitution) of novel risk factors derived from our objective AI approaches. The work in this proposal will be performed retrospectively in SD-OCT images from the image data pool that the multi-PI Sadda has aggregated over years as the director of the Doheny image reading center.