Accurate, reliable, and interpretable machine learning for assessment of neonatal and pediatric brain micro-structure

项目摘要 这个项目的目标是提高扩散加权磁共振成像的能力 （dMRI）用于新生儿和儿科受试者。目前，dMRI是唯一可行的非侵入性探测方法 大脑微观结构在过去的二十年里， 基于dMRI信号的脑微结构模型。不幸的是，准确和可靠的估计，这些 模型需要大量的高质量测量，这可能是困难的或不可能获得的， 新生儿和儿科受试者。因此，迫切需要能够准确、 鲁棒地估计从减少和低质量的测量微观结构生物标志物。为了解决这个 需要，这项研究将开发和验证数据驱动和机器学习（ML）技术的方法， 估计新生儿和儿科受试者的dMRI生物标志物。这些方法的潜力极大地 随着大型高质量dMRI数据集（如人类连接组项目）的可用性， (HCP)数据最近的工作，包括我们自己的研究，已经证明ML技术有很大的优势。 克服现有分析工具的局限性并实现上级估计精度的潜力。 这项研究将大大扩展我们的前期工作，并产生重要的新能力， 目前不存在。具体来说，我们将开发和验证新的方法来估计重要的微观， 结构模型和生物标志物，从扩散张量到高级多室模型， 在这方面，我们的工作的两个主要的新方面将包括1）使用空间， 时间地图集，以提高主题级分析的准确性，2）开发新的深度神经网络 基于自我关注的网络架构。此外，我们将开发新技术， 用于dMRI分析的ML方法的可靠性、鲁棒性和可解释性。这将包括以下技术： 计算校准良好的不确定性估计，技术，可以检测腐败，嘈杂，和出- 分布测量和技术，使解释和解释的预测， 这些ML方法。我们将使用重测和自举方法以及通过 大脑解剖学和微观结构专家的评估。在这项研究中开发的方法将 能够定量评估新生儿和儿科大脑的微观结构， 在大脑发育的关键阶段， 细节和可重复性是目前无法达到的。

Project Summary The goal of this project is to enhance the capabilities of diffusion-weighted magnetic resonance imaging (dMRI)for neonatal and pediatric subjects. Currently, dMRI is the only viable non-invasive method for probing brain microstructure. The past two decades have witnessed development of more powerful and more complex modelsof brain microstructure based on dMRI signal. Unfortunately, accurate and reliable estimation of these models require large numbers of high-quality measurements, which may be difficult or impossible to obtain in neonatal and pediatric subjects. Therefore, there is an urgent need for methods that can accurately and robustly estimatethe micro-structural biomarkers from reduced and low-quality measurements. To address this need, this researchwill develop and validate data-driven and machine learning (ML) techniques methods for estimating dMRI biomarkers for neonatal and pediatric subjects. The potential of these methods has greatly increased by the availability of large high-quality dMRI datasets such as the Human Connectome Project (HCP) data. Recent works, including our own studies, have demonstrated that ML techniques have a great potential to overcome limitations of the existing analysis tools and to achieve superior estimation accuracy. This research will substantially extend our preliminary work and generate important new capabilities that currently do not exist. Specifically, we will develop and validate novel methods for estimating important micro- structural models and biomarkers, ranging from diffusion tensor to advanced multi-compartment models, with far fewer measurements.In this regard, the two main novel aspects of our work will include 1) the use of spatio- temporal atlases to improvethe accuracy of subject-level analysis and 2) development of new deep neural network architectures based on self-attention. Furthermore, we will develop new techniques for enhancing the reliability, robustness, and explainability of ML methods for dMRI analysis. This will include techniques for computing well-calibrated uncertainty estimations, techniques that can detect corrupt, noisy, and out-of- distribution measurements, and techniques that enable interpretation and explanation of the predictions of these ML methods. We will evaluate the new methods using test-retest and bootstrapping methods and via assessment by experts in brain anatomyand micro-structure. The methods developed in this research will enable quantitative assessment of neonatal and pediatric brain micro-structure and the impact of developmental factors and neurological disorders at thesecritical stages in brain development with accuracy, detail, and reproducibility that is currently beyond reach.