Learning-based 3D modeling of AMD to assess disease progression and response to treatment

摘要： 视网膜相关性黄斑变性（AMD）是发达国家不可逆视力丧失的主要原因。 随着人口老龄化的加剧，AMD的负担将继续上升。尽管进行了大量的研究， 我们对是什么推动了疾病的进展以及为什么有些患者会进展到 最后阶段，对可用的治疗没有反应，并有严重的视力丧失。为了改善视力， 对于这些患者，开发能够识别具有早期无症状变化的个体的方法至关重要 他们有患上晚期疾病的风险。因此，通过利用最近的发展， 在深度学习和自动图像和视频分析方面，该项目的目标是开发计算 用于自动图像分析和生物标志物识别的方法和模型，以提高我们的理解 并帮助我们预测其进程。为了实现这一目标，我们将开发基于深度神经网络（DNN）的 用于从临床常用的不同成像模式中识别AMD的有用图像生物标志物的模型 实践-光学相干断层扫描（OCT）、眼底自体荧光（FAF）和眼底血管造影 （FA）。我们假设，增加FAF和FA图像将显着提高定位和分类的， 因此，有助于更好地了解疾病的自然史和对治疗的反应。 为了执行这一创新、高风险/高回报的项目，我们将使用来自杜克眼科的数据 Registry是全球最大的眼科单机构临床记录数据库。我们可以访问 一个下载的和专家注释的大型图像数据集，超过6400名患者符合我们的研究 入选标准-AMD从早期进展到晚期（干性和湿性）。对于每例患者，OCT， FAF和FA图像在单次访视期间或指定的随访间隔内（长达10年或 more）。我们将开发基于学习的方法和模型，用于单时间点图像融合和分析; 具体地，用于AMD分类的DNN模型（早期对中期对晚期干性AMD对湿性AMD）， 病变组织的分割和受影响区域的3D重建。最后，利用纵向 我们将开发深度学习方法和模型，以分析疾病随时间的演变。临界 将纵向数据集视为2D和3D病理模型序列， 允许使用卷积神经网络（CNN）和基于长短期记忆（LSTM）的神经网络。 网络，这是最近推出的，并采用时间相关的视频帧预测的上下文中， 自动驾驶汽车和人类行为识别。这项研究将导致早期AMD患者 分层，这将允许个性化定制的随访时间表，并导致及时治疗;这是 这一点非常重要，因为湿性AMD的早期治疗可以获得更好的视力结果。

Abstract: Age-related macular degeneration (AMD) is the leading cause of irreversible vision loss in developed countries. With the growing aging population, the burden of AMD will continue to rise. Despite significant research efforts, we still have limited understanding of what drives the disease progression and why some patients advance to final stages, not respond to available treatments, and have profound vision loss. To improve visual outcomes for these patients, it is critical to develop methods that can identify individuals with early, asymptomatic changes who are at the risk of developing advanced forms of disease. Consequently, by employing recent developments in deep learning and automated image and video analysis, the goal of this project is to develop computational methods and models for automated image analysis and biomarker identification to improve our understanding of AMD and help us to predict its course. To achieve this, we will develop Deep Neural Network (DNN)-based models to identify useful image biomarkers for AMD from different imaging modalities commonly used in clinical practice – Optical-Coherence Tomography (OCT), Fundus Autofluorescence (FAF), and Fundus Angiography (FA). We hypothesize that adding FAF and FA images will significantly improve localization and classification of pathology; thus, facilitate a better understanding of the disease's natural history and response to treatment. To execute this innovative, high-risk/high-reward project, we will use the data from The Duke Ophthalmic Registry, the largest single-institution clinical database for ophthalmic records in the world. We have access to a downloaded and expert-annotated large image dataset numbering over 6400 patients that meet our study's inclusion criteria – progression of AMD from early to advanced stages (dry and wet). For each patient, OCT, FAF, and FA images were captured during a single visit or within specified follow-up intervals (up to ten years or more). We will develop learning-based methods and models for one-time-point image fusion and analysis; specifically, DNN models for AMD classification (early vs. intermediate vs. advanced dry AMD vs. wet AMD), segmentation of the diseased tissue, and 3D reconstruction of the affected area. Finally, using the longitudinal datasets, we will develop deep-learning methods and models to analyze disease evolution over time. The critical insight and novelty are to consider longitudinal datasets as sequences of 2D and 3D pathology models, which allows for the use of Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM)-based neural networks, which were recently introduced and employed for time-dependent video frame prediction in the context of autonomous vehicles and human action recognition. The proposed research will result in early AMD patient stratification, which will allow for individually tailored follow-up schedules and lead to timely treatments; this is very important since the early treatment of wet AMD results in better visual outcomes.