Predicting the future development of advanced age-related macular degeneration (AMD) using multi-modal imaging and genetics

1)研究描述和潜在影响老年性黄斑变性(AMD)是发达国家最常见的致盲原因。在英国，每天有200多人患上晚期致盲新生血管(“湿”)型疾病。湿性AMD通常先影响一只眼睛，使患者依赖未受影响的“好”眼来进行日常生活活动。不幸的是，在许多(但不是所有)患者中，良好的视力随后会受到影响，患者变得严重视力受损。因此，一些研究已经开始探索预防AMD进展的治疗方法，包括湿性和干性形式。在许多情况下，这些治疗是侵入性的，有可能产生不良影响。因此，预测AMD未来进展的可靠方法将使这些治疗更有针对性--这种风险分层可以识别那些面临即将转换的风险(即，在六个月内发展为晚期AMD)的患者以及那些可以放心的患者(即，不太可能在未来两年内发展为晚期AMD)。2)目标-开发机器学习系统，用于预测即将到来的AMD进展，定义为在6个月内进展到脉络膜新生血管(CNV)或地理萎缩(GA)，使用人口学和临床元数据加上来自以下方面的多模式成像(彩色眼底摄影(CFP)/眼底自发荧光(FAF)和高分辨率3D光学相干断层扫描(OCT))：1)单一时间点，和2)纵向数据。-使用相同的数据，开发AMD“患者放心”模型，定义为在两年内不进展为脉络膜新生血管(CNV)或地理萎缩(GA)。-将有关遗传变异和/或其他诊断测试的信息纳入AMD预测模型，并评估其在每种临床情况下对模型性能的递增影响。-对照人类专家(Moorfield眼科医院在视网膜疾病方面具有专科专长的眼科医生)的表现作为AMD预测模型的基准。-通过在前瞻性、非干预性临床研究中验证模型的性能来探索初步的临床翻译。3)研究方法的新颖性机器学习的使用已经显示出利用成像方式对视网膜疾病进行分类的巨大潜力。许多研究已经证明，通过视网膜CFP和/或OCT扫描，深度学习可以预测未来AMD的进展。然而，这些研究通常在单个时间点采用单一模式，产生良好的结果，但不是惊人的结果。我们将开发结合纵向、多模式成像数据和遗传数据的AMD进展模型，然后展示其潜在的临床适用性。该项目将涉及卷积神经网络等现有技术的应用，以及图形神经网络等较新方法的应用。它还将涉及更高级的建模技术，如神经常微分方程式。最后，它将涉及监督和半监督学习。4)与EPSRC的战略保持一致与EPSRC的“医学成像”研究领域和EPSRC的医疗技术挑战密切相关，以“通过有效的诊断、特定患者的预测和基于证据的干预来优化治疗和护理”。5)合作本项目将使用来自Moorfield眼科医院的成像数据。Moorfield隶属于伦敦大学学院眼科研究所，拥有世界上最大的单中心眼科成像数据库(包括来自近10,000名晚期AMD患者的200,000对CFP和OCT扫描)。自2019年以来，他们已经开始收集基因，他们也开始收集这些患者的基因数据。

1) Description and potential impact of the research Age-related macular degeneration (AMD) is the most common cause of blindness in the developed world. In the UK, more than 200 people develop the advanced blinding neovascular ("wet") form of the disease daily. Wet AMD typically affects one eye first, leaving patients reliant upon the unaffected "good" eye to allow the activities of daily living. Unfortunately, in many - but not all - patients, the good eye subsequently becomes affected, and the patient becomes severely sight impaired. As a result, a number of studies have begun to explore preventative therapies for the development of AMD progression, both in its wet and dry forms. In many cases, these treatments are invasive, with the potential for adverse effects. Robust methods for predicting future progression of AMD would thus allow better targeting of these therapies - such risk stratification could allow identification of those patients at risk of imminent conversion (i.e., development of advanced AMD within a six-month period) as well as those patients that could be reassured (i.e., unlikely to develop advanced AMD within the next two years). 2) Objectives- Develop machine learning systems for prediction of imminent AMD progression, defined as progression to choroidal neovascularization (CNV) or geographic atrophy (GA) within a 6-month period, using demographic and clinical metadata plus multi-modality imaging (colour fundus photography (CFP) / fundus autofluorescence (FAF), and high-resolution 3D optical coherence tomography (OCT)) from: 1) single time-point, and 2) longitudinal data. - Develop AMD "patient reassurance" models, defined as NO progression to choroidal neovascularization (CNV) or geographic atrophy (GA) within a 2-year period, using the same data. - Incorporate information on genetic variants and/or other diagnostic tests into AMD prediction models and evaluate its incremental effects on model performance in each clinical scenario.- Benchmark AMD prediction models against performance of human experts (ophthalmologists with subspecialty expertise in retinal disease at Moorfields Eye Hospital).- Explore preliminary clinical translation by validating model performance in prospective, non-interventional clinical studies.3) Novelty of Research MethodologyThe use of machine learning has shown great potential for retinal disease classification using imaging modalities. A number of studies have demonstrated the potential of deep learning to predict future AMD progression using retinal CFP and/or OCT scans. However, these studies have typically employed a single modality at a single time-point, producing good - but not spectacular - results. We will develop AMD progression models that incorporate longitudinal, multi-modal imaging data, and genetic data, and then demonstrate their potential clinical applicability. This project will involve the application of established technologies such as convolutional neural networks, as well as newer approaches such as graph neural networks. It will also involve more advanced modelling techniques such as neural ordinary differential equations. Lastly, it will involve both supervised and semi-supervised learning.4) Alignment to EPSRC's strategiesStrongly related to EPSRC's "Medical Imaging" research area and EPSRC's Healthcare Technology challenge to "Optimise Treatment and Care through effective diagnosis, patient-specific prediction and evidence-based intervention."5) CollaborationsThis project will use imaging data from Moorfields Eye Hospital. Affiliated with the UCL Institute of Ophthalmology, Moorfields has the world's largest single-centre ophthalmic imaging database (including >200,000 paired CFP and OCT scans from nearly 10,000 patients with advanced AMD). Since 2019, they have begun collecting gene they have begun collecting genetic data on these patients also.