Reading minds with Deep Learning: predicting behavioural states from functional imaging data

博士项目的目的：目标是：开发用于大脑功能时空深度学习的工具用于预测脆弱早产儿的神经发育结果，以及对ADHD和自闭症风险敏感的生物标志物的开发项目描述/背景：复杂认知障碍（如自闭症和ADHD）的精确诊断极具挑战性，因为此类疾病的特征是认知和行为特征的高度异质性。这些特征反映了大脑活动的时空动态的细微特征，因此很难被描述。目前，用于分析静息态功能成像数据的最流行的技术是空间独立成分分析（spatial-ICA，Independent Component Analysis）。这将大脑建模为一个宏观网络，由一组功能专门化的区域形成，每个区域与一个时间过程相关联。然后通过使用相关性测量来估计时间过程之间的相似性来推断网络连接性[1]。尽管伊卡等矩阵因子分解方法显着提高了我们对大脑功能与行为之间关系的理解，但它们消除了受试者间差异的重要来源。具体来说，伊卡分析着眼于大脑状态随时间的平均特性，而已经表明，许多行为测量可以通过动态测量更好地预测[2]。此外，研究假设皮层组织的单一全局平均模型;然而，越来越多的证据表明情况并非如此[3，4]。所需要的是可以从数据中学习时间和空间特征的工具，而无需对数据进行事先建模或空间归一化。这个问题适用于深度学习;因此，我们寻求从最近的自然图像处理[5，6]，心脏成像[7]和功能磁共振成像的时空卷积深度学习工作中获得灵感（fMRI [8]），为了对病理性大脑状态进行分类，并支持神经发育障碍的精确诊断。鉴于认知研究需要对大脑表面进行精确分析，（或皮层，[9，10]），一个关键目标将是将模型扩展到几何深度学习[11，12]，它在表面流形上训练，而不是2D或3D网格。还将重点放在可解释模型的开发上[6]。这将支持临床解释。该项目最合适的候选人将具有Python编程专业知识和深度学习经验。具有时空数据集或几何深度学习的工作经验将是一个重要的优势。

Aim of the PhD Project:The goal is to:Develop tools for spatio-temporal Deep Learning of brain functionFor prediction of neuro-developmental outcome in vulnerable preterm babies,And development of biomarkers sensitive to risk of ADHD and AutismProject Description / Background:Precision diagnosis of complex cognitive disorders, such as Autism and ADHD, is extremely challenging since such disorders are characterised by a highly heterogeneous range of cognitive and behavioural traits. Such traits are extremely difficult to characterise as they reflect subtle features of the spatio-temporal dynamics of brain activity.Currently, the most popular technique for analysing resting-state functional imaging data is to perform spatial-ICA (independent component analysis [1]). This models the brain as a macroscale network, formed from a set of functionally specialised regions, each associated with a time course. Network connectivity is then inferred by estimating similarities between time courses using correlation measures [1].Although matrix factorisation approaches such as ICA, have significantly improved our understanding of how brain function relates to behaviour, they smooth out vital sources of inter-subject variation. Specifically, ICA analyses look at the average properties of brain states over time, whereas, it has been shown that many behavioural measures are better predicted by dynamic measures [2]. Further, studies assume a single global average model of cortical organisation; however, there is growing evidence that this is not the case [3,4].What is required are tools that can learn temporal and spatial features from the data without requirement for prior modelling or spatial normalisation of the data. This problem lends itself to deep learning; we therefore seek to take inspiration from recent works on spatio-temporal convolutional deep learning for natural image processing [5,6], cardiac imaging [7] and functional Magnetic Resonance Imaging (fMRI [8]), in order to classify pathological brain states, and support precision diagnosis of neuro-developmental disorders.Given that studies of cognition require precision analysis of the brain's surface (or cortex, [9,10]), a key objective will be to extend models to geometric deep learning [11,12], which trains on surface manifolds, rather than 2D or 3D grids. Significant emphasis will also be placed on the development of interpretable models [6]. This will support clinical interpretation.The most suitable candidate for this project will have programming expertise in Python, and experience in Deep Learning. Experience in working with spatio-temporal data sets or geometric deep learning would be a significant plus.