Accelerating discovery of the human foveal microconnectome with deep learning

点击翻译按钮获取中文摘要

Project Summary The human retina is one of the most complex microcircuits of the central nervous system (CNS) and is a model of CNS neurodegenerative disease with unique advantages for microconnectomics technology advancement. The central retina or fovea mediates high acuity vision, drives activity in half of the brain, and is a critical locus for prevalent blinding disease. The fovea is small (<1 mm), accessible, and relevant to CNS disease diagnosis through advanced cellular-level clinical imaging. The full foveal microconnectome comprises both the diverse neural circuits that create parallel visual pathways as well as complex microconnectivity with two specialized cell types of neuroectodermal origin, the retinal pigment epithelium (RPE) and the Müller glia. Our group has pioneered ultra-short recovery times of eyes from organ donors, to create exquisitely preserved retinal tissue volumes suitable for the first microconnectomic analysis of an intensively investigated human CNS structure. The goal of this proposal is to accelerate the human foveal microconnectome by refining and augmenting a highly successful and professionally supported software platform, Dragonfly by Object Research Systems (ORS), an industry leader in implementation of deep learning methods for auto-segmentation of complex structure. Our collaboration with ORS will target development of deep learning (DL) models as well as annotation and proofreading tools that will have broad applicability to neuroscience microconnectomics. In preliminary studies we discovered that RPE cells give rise to extremely dense neural-like projections to photoreceptor cells and that foveal Müller glia similarly have a specialized and complex relationship to foveal microcircuits. Moreover, single foveal cone photoreceptors were presynaptic to dozens of parallel visual circuits of extreme complexity. To advance understanding of these complex microconnectomes ORS will augment fast auto-segmentation using newly developed convolutional neural networks and refine sophisticated tools for rapid annotation, proofreading, data visualization, and quantitative analysis. In Aims 1 and 2 we will develop complete deep learning models of the human RPE cell-neuronal microconnectome and the Müller cell- neuronal microconnectome respectively that will transform our understanding of the critical roles these cell types play in foveal function and disease. In Aim 3 we will develop a deep learning model of the multiple neural cell types and microconnectome of parallel visual pathways for form, color, and motion vision. The major outcome will be the transformation of a powerful, widely used, professionally supported, DL-based platform for broad application to neuroscience microconnectomics, free for academic research via a no-cost license. The ORS-Dragonfly platform will accelerate microconnectomics of complex CNS circuitry and impact systems neuroscience, human neuro-pathophysiology, and interpretation of cellular-level clinical imaging. This proposal combines expertise and innovation in neurobiology, vision science, clinical ophthalmology and connectomics, with DL software development and application.