Realising the potential of retinal image analysis via AI methods for early detection of brain disease in the community

点击翻译按钮获取中文摘要

Changes in the brain leading to brain disease are thought to start decades before cognitive symptoms emerge. If biomarkers for these early stages could be identified, it would contribute to a more accurate estimation of an individual's risk of developing disease and enable the monitoring of high-risk (presymptomatic) people. This would also provide the means for assessing the efficacy of new interventions. The retina is an extension of the brain sharing embryological origins as well as a blood supply and nerve tissue. It, therefore, has huge potential as a site for biomarker investigation through easy, non-invasive imaging. Computational image analysis can then be used to reveal valuable information about microvascular health, deposition, and neurodegenerative damage. [1] The Scottish Collaborative Optometry-Ophthalmology Network e-research (SCONe, https://www.ed.ac.uk/ophthalmology/scone ) is a pioneering project established in 2020 which aims to build a world-leading nationwide retinal image resource for innovation in eye research and healthcare. In Scotland, community-based optometrists routinely collect millions of retinal images every year and many have been doing so for more than a decade. This represents a unique opportunity to create a large-scale longitudinal image resource that is representative of the primary care population. Within its initial two-year pilot phase, SCONe demonstrated the feasibility of bringing community-acquired retinal images for people aged 60+ together with other routinely collected healthcare data within the NHS National Safe Haven and has successfully acquired and linked over 200,000 images. Community-based early identification of individuals at risk of cognitive decline and dementia is currently a major unmet clinical need. Our solution is to exploit the potential embedded within retinal images to predict neurovascular health in the aging brain using routinely captured retinal images, cross-linked hospital datasets with information on symptoms/diagnosis, and the latest advances in Artificial Intelligence [2]. Based on this, we will devise a retinal neurovascular biomarker toolkit for cognitive decline, stroke, and dementia. Predictive modelling based on routinely collected retinal images poses a broad range of novel image analysis problems related to image quality control and enhancement, longitudinal image registration, and feature engineering [3]. AimsThe main aim of the study is to develop Artificial Intelligence approaches to the early identification of individuals at risk of cognitive decline and dementia, which is a major unmet clinical need. These solutions can then be deployed in the community (GP practices, optometrists, consumer devices). Progress towards this aim will be delivered based on the following objectives: Taking advantage of the data linkage capabilities of Public Health Scotland, link the SCONe repository of retinal image with hospital inpatient/outpatient electronic health records (e.g. SMR datasets) to identify a population of individuals over 60 years old with retinal images spanning over 10 years or more and incident diagnosis of cognitive decline, stroke or dementia. Develop novel approaches to quality scoring and retinal image enhancement using generative adversarial networks (GANs) to make this community acquired dataset suitable for latest advances in deep learning for image classification. Develop novel approaches to longitudinal retinal image registration and identification of key retinal phenotypes changing over time. Extend existing epidemiological models predicting the risk of brain health deterioration and incident disease with the SCONe linked retinal images, either in raw format at multiple points in time or after having characterised their temporal evolution in point 3.