Deep-Learning-Derived Endophenotypes from Retina Images

摘要 该方案的目标是建立一种新的基于人工智能的基因组策略 广谱关联研究用于识别与普通人相关的新的遗传基因座 精神错乱。对这些疾病背后的遗传因素的识别不仅将提供 对疾病的机械性洞察也形成了开发新的预防方法的基础， 诊断性和针对性的治疗方法。虽然全球气候变化网络在以下方面取得了巨大成功 在过去的1.5年里，目前只有一小部分常见疾病的遗传性可以 这一现象可由传统的GWA型基因座来解释。利用公共基因和 针对临床影像数据，我们提出了一种新的方法，称为基于图像的GWAS(IGWAS)， 其中，将对来自图像的内表型进行GWAS，使用切割边缘 深度学习(DL)算法。通过创建更客观、更定量的产出 在信息损失较少的情况下，许多新的与疾病相关的遗传基因座有望 被指认出来。为了测试这种新方法的有效性，将使用人类的视觉系统 作为一个例子。一种DL-表现型、多通道多视点自监督深度学习 视网膜图像编码器(Mussler)，将开发用于从 正常眼和患者的光学相干断层扫描和眼底图像 患有糖尿病视网膜病变(DR)。将对正常的眼内表型进行GWA以 阐明与视网膜发育和生理相关的基因。此外，还将执行全球气候变化分析 DR特异性内表型以确定DR相关遗传基因座和候选基因。如果 成功，我们的方法代表了一个通用的框架，可以很容易地扩展到 用影像资料研究其他视网膜疾病和其他常见病。此外， 本项目建立的表型神经网络体系结构可以很容易地应用于 开发一个针对不同临床数据的自动评分工具，并将其应用于许多其他数据 常见病。

Abstract The goal of this proposal is to establish a novel artificial intelligence-based strategy of genome wide association studies (GWAS) to identify new genetic loci associated with common human disorders. Identification of the genetic factors underlying these diseases will provide not only mechanistic insights into the diseases but also form the basis for developing novel preventative, diagnostic, and targeted therapeutic methods. While GWAS has achieved great successes in the past 1.5 decades, only a small portion of common diseases' heritability can be currently explained by loci identified from traditional GWAS. Leveraging on the public genetics and clinical imaging data, we propose a novel approach, termed image based GWAS (iGWAS), where GWAS will be performed on endophenotypes derived from images using cutting edge deep learning (DL) algorithms. By creating a more objective and quantitative output from the images, with less information loss, many new disease-associated genetic loci are expected to be identified. To test the efficacy of this novel approach, the human visual system will be used as an example. A DL-phenotyper, Multi-modal multi-view Self-Supervised deep Learning Encoder for Retinal images (MuSSLER), will be developed to extract quantitative output from optical coherence tomography scans and fundus images from both normal eyes and patients with diabetic retinopathy (DR). GWAS will be performed on normal eye endophenotypes to elucidate genes relevant to retina development and physiology. Also, GWAS will be performed on DR-specific endophenotypes to identify DR associated genetic loci and candidate genes. If successful, our approach represents a general framework that can be readily extended to the study of other retinal diseases and other common diseases with imaging data. Furthermore, the phenotyping neural network architecture established in this project can be readily adopted to develop an automated grading tool for heterogeneous clinical data and applied to many other common diseases.