Deep Learning with Neuroimaging Genetic Data for Alzheimer's Disease

摘要 阿尔茨海默病(AD)影响着全球超过4400万人，预计这一数字将增加两倍 到2050年。然而，目前还没有治愈阿尔茨海默病的方法。该项目旨在开发和应用新的统计学 方法，特别是深度学习，促进阿尔茨海默病的神经影像遗传学。它涉及到新的方法论 AIMS 1-4的发展，对大规模英国生物库神经成像遗传数据的成本效益应用 广告(目标5)和软件开发(目标6)。方法开发的所有四个目标都解决了新出现的问题- 深度学习的新主题及其在阿尔茨海默病神经成像遗传学中的应用；尽管其他三个目标 用它们自己的其他广泛应用来处理独立的主题，它们反过来服务于Aim 1：1)Aim 1适用 人工搜索用于从神经图像中自动提取特征/表型的深度学习模型，通过该模型 后续全基因组关联研究的统计能力和生物学解释(GWAS) 2)Aim 2使用(自动)神经体系结构搜索(NAS)来更科学地执行fi 确定更好的深度学习模型，然后将其应用于目标1，以增强特征提取/表型分析 3)目标3侧重于可解释的深度学习，提供生物学见解 通过本地化和突出显示深度学习模型提取的最重要特征，这些特征可用于 目标1；4)目标4)为深度学习开发了一种新的推理理论，并将其应用于严格的测试 在目标1中使用的任何选定/突出显示的特征的统计意义。在目标5中，这些新方法将是fi 应用于英国生物库神经成像和GWAS数据以识别新的遗传位点和神经成像特征 对于AD。作为副产品，我们将开发和分发实现目标6中建议的方法的软件。

Summary Alzheimer's disease (AD) affects over 44 million individuals worldwide, and the number is projected to triple by 2050. However, currently there is no cure for AD. This project aims to develop and apply novel statistical methods, especially deep learning, to advance neuroimaging genetics for AD. It involves novel methodological developments in Aims 1-4, cost-effective applications to the large-scale UK Biobank neuroimaging genetic data for AD (Aim 5), and software development (Aim 6). All four Aims for the methods development tackle emerging impor- tant topics in deep learning with their applications to neuroimaging genetics for AD; although the other three Aims deal with independent topics with their own other broad applications, they in turn serve for Aim 1: 1) Aim 1 applies manually searched deep learning models for automatic feature extraction/phenotyping from neuroimages, by which both the statistical power and biological interpretation of subsequent genome-wide association studies (GWAS) are expected to be enhanced; 2) Aim 2 employs (automatic) neural architecture search (NAS) to more efficiently identify better deep learning models, which are then applied to Aim 1 for enhancing feature extraction/phenotyping and thus boosting the power of GWAS; 3) Aim 3 focuses on explainable deep learning, offering biological insights by localizing and highlighting the most important features extracted by deep learning models that can be used for Aim 1; 4) Aim 4 develops a novel inferential theory for deep learning, which is then applied to rigorously test for the statistical significance of any selected/highlighted features used in Aim 1. In Aim 5, these new methods will be applied to the UK Biobank neuroimaging and GWAS data to identify novel genetic loci and neuroimaging features for AD. As a byproduct, we will develop and distribute software implementing the proposed methods in Aim 6.