CAREER: New Learning-based Algorithms for the Analysis of Very-Large-Scale Neuroimaging Data

在机器学习最新进展的推动下，人工智能有望改变医疗保健和生物医学研究。机器学习算法允许研究人员分析大型数据集中的复杂模式，以促进我们对生物机制的理解和开发临床工具。该项目考虑了非常大规模的脑成像研究，包括，例如，数以万计的个人提供头部MRI扫描和其他生物医学数据，如全基因组序列或临床记录。这些数据使研究人员能够绘制出基因、环境和其他因素对大脑结构和功能的影响，从而提高我们对阿尔茨海默氏症等疾病的认识。今天，开发大规模脑成像数据集的主要障碍是计算性的，因为现有的软件工具不能很好地扩展，并且缺乏质量保证能力。这个项目将产生一个基于机器学习的计算管道来填补这一空白。在同类最大的研究中，我们将展示开发的软件工具来绘制大脑结构形状的遗传性。此外，该项目还将实施一系列多样化的教育推广举措，例如为代表性不足的少数族裔高中学生提供定制化的研究体验。神经影像学正在进入一个前所未有的规模和复杂性的新时代。很快，我们将拥有超过10万人的数据集。分析和利用这些数据的基本挑战是计算。今天，广泛使用的神经图像分析工具在计算上要求很高，产生的结果对混杂因素很敏感，并且在质量控制能力方面受到限制，这使得它们在规模上不可行。该项目将扩展机器学习的最新进展，开发一种创新的计算管道，解决现有方法的缺点。首先，将开发一个计算效率高且灵活的脑MRI分割框架，该框架将集成丰富的神经解剖学先验模型。通过使用对抗性学习策略，分割工具将对混淆效应（如受试者运动）具有鲁棒性。将进一步研究基于学习的方法，以避免耗时的人工分割质量控制。最后，一种创新的度量学习方法将用于研究遗传对英国生物银行大脑形态的影响。该项目还将实施一项以跨学科、实践和终身学习为重点的综合教育计划。研究人员将投入大量精力开发核心教育材料，以适应不同背景和阶段的观众。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Artificial intelligence, fueled by recent advances in machine learning, is poised to transform healthcare and biomedical research. Machine learning algorithms allow researchers to analyze complex patterns in large datasets, in the service of advancing our understanding of biological mechanisms and developing clinical tools. This project considers very large scale brain imaging studies, including, for example, tens of thousands of individuals contributing head MRI scans and other biomedical data such as whole-genome sequences or clinical records. Such data allow researchers to map the effects of genetic, environmental, and other factors on the structure and function of the brain, which in turn advances our knowledge of disorders like Alzheimer's. Today, the primary obstacle in exploiting very large scale brain imaging datasets is computational, because existing software tools don't scale well and lack in quality assurance capabilities. This project will produce a machine-learning based computational pipeline that will fill this gap. In the largest study of its kind, we will showcase the developed software tools to chart the heritability of shapes of brain structures. In addition, the project will implement a diverse set of educational outreach initiatives, such as a customized research experience for under-represented minority high-school students. Neuroimaging is entering a new era of unprecedented scale and complexity. Soon, we will have datasets including more than 100,000 individuals. The fundamental challenge in analyzing and exploiting these data is computational. Today, widely-used neuroimage analysis tools are computationally demanding, produce results that are sensitive to confounds, and are limited in quality control capabilities, making them infeasible at scale. This project will extend recent advances in machine learning to develop an innovative computational pipeline that addresses the drawbacks of existing methods. First, a computationally efficient and flexible brain MRI segmentation framework will be developed that integrates rich neuroanatomical prior models. The segmentation tool will be made robust to confounding effects such as subject motion via the use of an adversarial learning strategy. Learning-based methods will be further investigated to obviate the time-consuming manual quality control of segmentations. Finally, an innovative metric learning approach will be used to study genetic influences on brain morphology in the UK Biobank. The project will also implement an integrated educational plan that is focused on interdisciplinary, hands-on and lifelong learning. The researchers will devote significant effort to developing core educational material that will be adapted and utilized for audiences of various backgrounds and stages.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.