Deep Learning Applications in Structural and Functional MRI

结构磁共振成像(SMRI)允许对大脑结构进行非侵入性成像。功能磁共振成像(FMRI)测量人类大脑活动的某些方面。SMRI和fMRI数据(统称为MRI数据)在脑部和大脑研究中极其重要。深度学习在2D图像识别和自然语言处理等其他领域也取得了惊人的成功。它有可能以多种方式极大地放大MRI在神经科学中的用途：对现已可用的大型MRI数据集进行自动质量控制和预处理；改进对fMRI脑信号的解码；改进从个体受试者的MRI数据中提取有用信息的功能，例如年龄、性别和临床健康状况。不过，必须克服重大挑战。 深度学习方法已被应用于核磁共振数据，例如，在2D核磁共振图像中分割大脑区域，在功能核磁共振中识别大脑活动模式，以及区分临床诊断为精神分裂症的患者和健康对照组。大多数先前的研究使用了相对较小的数据集(40名或更少的受试者)。这项研究计划的目标1是用更大的数据集复制以前的一些结果，包括100个或1000个受试者。 深度学习需要具有1万到数百万个训练实例的大型训练数据集。核磁共振数据集相对较小，在目前可用的最大核磁共振数据集中只有1,000-2,000个示例。为了解决这一问题，MRI中的许多深度学习应用程序从MRI数据中提取许多小的2D块，或者从日常2D图像的大数据集上训练的深度学习模型开始(例如：ImageNet)。目标2a是研究将先前的领域知识和假设(例如：fMRI血氧水平依赖(BOLD)信号属性)合并到深度学习模型中，作为减少训练数据集大小要求的一种手段，以及这可能如何影响模型性能。目的2b旨在扩展深度学习方法，以利用目前大多数深度学习MRI研究中忽略的sMRI的3D结构和fMRI的4D结构。 深度学习有一个“黑箱”或不可解释的问题，因为人类通常不可能遵循深度学习模型的“决策”过程。目标3是研究深度学习模型的解释方法，并将其应用于MRI数据。成功将提高神经科学家从复杂的核磁共振数据集中提取知识的能力。

Structural magnetic resonance imaging (sMRI) allows for non-invasive imaging of brain structures. Functional MRI (fMRI) measures certain aspects of human brain activity. sMRI and fMRI data (collectively, MRI data) are incredibly important in mind and brain research. Deep learning has been spectacularly successful in other fields like 2D image recognition and natural language processing. It has potential to greatly amplify the utility of MRI in neuroscience in multiple ways: automated quality control and preprocessing of the large MRI datasets now becoming available; improved decoding of fMRI brain signals; improved extraction of useful information from individual subjects' MRI data such as age, gender, and clinical health status. Important challenges must be overcome though. Deep learning methods have been applied to MRI data, for example in segmenting brain regions in 2D MRI images, identifying brain activity patterns in fMRI, and differentiating patients with a clinical diagnosis (eg: schizophrenia) vs. healthy controls. Most previous studies used relatively small datasets (40 or fewer subjects). Objective 1 of this research program is to replicate some of the previous results with larger datasets including 100s or 1000s of subjects. Deep learning requires large training data sets with 10,000s to millions of training examples. MRI datasets are comparatively small with only 1,000-2,000 examples in the largest MRI datasets currently available. To get around this, many deep learning applications in MRI extract many small 2D patches from the MRI data or else start with deep learning models trained on large datasets of everyday 2D images (eg: Imagenet). Objective 2a is to investigate incorporation of prior domain knowledge and assumptions (eg: fMRI blood oxygenation level dependent (BOLD) signal properties) into deep learning models as a means of reducing training dataset size requirements, as well as how this may affect model performance. Objective 2b is to extend deep learning methods to take advantage of 3D structure in sMRI and 4D structure in fMRI, currently ignored in most deep learning MRI studies. Deep learning has a "black box" or non-interpretability problem in that it is typically not possible for a human to follow the deep learning model's "decision making" process. Objective 3 is to investigate methods for interpreting deep learning models as applied to MRI data. Success would improve neuroscientists' ability to extract knowledge from complex MRI datasets.