The Machine and the Brain: Early Prediction of Dementia using Retinal Biomarkers by Artificial Intelligence

背景：痴呆引起的结构和化学改变导致神经元丢失和脑萎缩。痴呆症患者的认知能力会下降，这损害了他们在没有帮助的情况下完成日常任务的能力，因此需要依靠照顾者的帮助。诊断痴呆症的传统方法虽然有争议是可靠的；然而，它们可能是昂贵的、耗时的，并且具有中度侵入性。然而，由于视网膜和大脑在解剖学上有相当大的相似性，通过光学相干断层扫描（OCT）和OCT血管造影（OCTA）识别视网膜异常已被探索作为一种可行的非侵入性方法来诊断神经退行性疾病，特别是阿尔茨海默病（AD）和轻度认知障碍（MCI）。为了实现这些视网膜变化，从OCT/OCTA中提取可靠的视网膜生物标志物，必须考虑合适的分割方法。目的：本研究的主要目的是探讨认知能力下降与OCT/OCTA图像中提取的视网膜生物标志物之间的关系，并评估生物标志物在AD和MCI早期检测中的有效性。在提取视网膜生物标志物之前，为了解决分割方法的不足，另一个研究目标是开发基于深度学习（DL）的分割工具，使分割过程高效/准确地自动化，从而提取更可靠的视网膜参数。然而，通过特征选择算法确定哪些参数在分类任务中更占主导地位是本项目的另一个目标。方法：在PubMed、Web of Science和Scopus上进行系统检索，直到2022年12月，使用商定的搜索关键词和纳入/排除标准获得64篇论文。此外，意大利国家研究所IRCCS“Saverio De Bellis”研究医院（Castellana Grotte）收集了一个专门的数据集，其中包含：155名hc， 168名MCI， 66名AD和3名早期AD （eAD）个体，这些个体主要通过MMSE和FAB评分进行认知评估。所有参与者使用RTVue XR 100 Avanti SD-OCT系统（Optovue, Inc.）进行OCT检查，使用商用软件提供的AngioVue模块进行OCT分割。获取参与者的OCT b扫描、OCTA面部扫描和Optovue RTVue XR软件自动提取的参数。通过结构良好的统计计划，对机器提取的参数进行统计检查，以发现认知衰退阶段的关联，主要是AD和MCI对hc的影响。此外，使用公开可用的数据集（ROSE和OCT-500），训练各种深度学习模型，以更好地从OCTA面部扫描中分割血管和中央凹无血管区（FAZ），然后进行以前没有商业解决方案提供的新参数提取。此外，另一个由OCT图像组成的复杂数据集来自三个不同的OCT系统，由专家注释者手动描绘视网膜层。该OCT数据集用于训练另一个DL模型，以自动分割以中央凹为中心的视网膜层。结果：该系统综述论文成功发表，指出了各种神经退行性疾病与OCT/OCTA成像方式提取的特异性视网膜生物标志物之间的关联。另一方面，对提取的机器参数进行的统计分析表明，与hc相比，MCI和Dem在中央凹周围不同区域的血管密度明显降低。相反，一些研究的机器参数提供了与文献不同的结果，因此，这些发现以另一份手稿草稿的形式写成，目的是发表另一篇论文。另一方面，开发的深度学习模型在中央凹周围视网膜层分割（OCT）的性能

Background: Structural and chemical alterations induced by dementia lead to neuronal loss and brain atrophy. Individuals with dementia experience a decline in cognitive abilities, which impairs their ability to carry out usual daily tasks unaided, hereby relying on caregivers for assistance. Conventional approaches to diagnose dementia although arguably reliable; however, they can be expensive, time-intensive, and moderately invasive. Nevertheless, due to the considerable anatomical resemblances between the retina and the brain, abnormalities in the retina identified through optical coherence tomography (OCT) and OCT angiography (OCTA) have been explored as a viable non-invasive method for diagnosing neurodegenerative conditions, especially Alzheimer's Disease (AD) and Mild Cognitive Impairment (MCI). To realize these retinal changes and extracting reliable retinal biomarkers from OCT/OCTA, an appropriate segmentation method must be considered. Objectives: The main objective of this study is to explore the relationship between cognitive decline and retinal biomarkers extracted from OCT/OCTA images as well as evaluating biomarkers' effectiveness in early detection of AD and MCI. Prior extracting retinal biomarkers, and to address the shortcomings of segmentation methods, another research objective is to develop deep learning (DL) based segmentation tools to automate the segmentation procedure efficiently/accurately, and hence, extracting more reliable retinal parameters. However, determining which parameters are more dominant in the classification task via feature selection algorithms is another objective of this project. Methods: A systematic search was conducted on PubMed, Web of Science, and Scopus until December 2022, resulted in 64 papers using agreed search keywords, and inclusion/exclusion criteria. Moreover, a dedicated dataset was collected by the National Institute IRCCS "Saverio De Bellis" Research Hospital - Castellana Grotte - Italy that contained: 155 HCs, 168 MCI, 66 AD, and 3 early AD (eAD) individuals that have been cognitively assessed mainly via MMSE and FAB scores. All participants underwent OCTA examination by RTVue XR 100 Avanti SD-OCT system (Optovue, Inc.) where OCT segmentation was performed using the AngioVue module present by commercial software. The OCT b-scans, OCTA en-face scans, and parameters automatically extracted by Optovue RTVue XR software were acquired for participants. The machine extracted parameters were statistically examined, via a well-structured statistical plan, to find associations with cognitive decline stages, mainly AD and MCI against HCs. Moreover, using publicly available datasets (ROSE and OCT-500), various DL models were trained to better segment vasculature and foveal avascular zone (FAZ) from OCTA en-face scans, followed by arguably new parameters extraction not previously provided by commercial solutions. Additionally, another complex dataset consisting of OCT images was acquired from three distinct OCT systems with manual delineation of retinal layers performed by expert annotators. This OCT dataset was used to train another DL model to automatically segment retinal layers specifically centred around the fovea. Results: The systematic review paper, successfully published, indicated the association between various neurodegenerative disorders and specific retinal biomarkers extracted by OCT/OCTA imaging modality. On the other hand, the statistical analysis performed on the extracted machine parameters demonstrated a significant vascular density reduction in different sections around the fovea for MCI and Dem compared against HCs. Conversely, some of the studied machine parameters provided dissimilar results to the literature, and hence, these finding were written in a form of another manuscript draft for the purpose of publishing another paper. On the other hand, the performance of developed DL models for retinal layers segmentation around fovea (from OCT