Program
June 20 9:30-11:00am
Topological Data Analysis
The session explores emerging topological methods for brain image and network analysis. Topological methods in brain imaging can be traced to Keith Worsely and his research on random field theory. Session Chair: Moo K. Chung, University of Wisconsin-Madison
Armin Schwartzman, University of California - San Diego, USA
Title: Estimation of expected Euler characteristic curves of nonstationary smooth random fields, with application to fMRI analysis
Abstract: The expected Euler characteristic (EEC) of excursion sets of a smooth Gaussian-related random field over a compact manifold approximates the distribution of its supremum for high thresholds. Viewed as a function of the excursion threshold, the EEC of a Gaussian-related field is expressed by the Gaussian kinematic formula (GKF) as a finite sum of known functions multiplied by the Lipschitz–Killing curvatures (LKCs) of the generating Gaussian field. This paper proposes consistent estimators of the LKCs as linear projections of “pinned” Euler characteristic (EC) curves obtained from realizations of zero-mean, unit variance Gaussian processes. As observed, data seldom is Gaussian and the exact mean and variance is unknown, yet the statistic of interest often satisfies a CLT with a Gaussian limit process; we adapt our LKC estimators to this scenario using a Gaussian multiplier bootstrap approach. The proposed methods are illustrated on voxelwise testing in 3D fMRI volumes.
Alexander Strang, University of California - Berkeley, USA
Title: Towards Topological Data Analysis of Directed Data
Abstract: Topological Data Analysis (TDA) promises tools for approximating the topology of a point cloud via persistent homology. These tools can summarize the "shape" of high-dimensional data sets. Traditional TDA uses undirected graphs with edges weighted by a distance metric which evaluates the dissimilarity of data at the endpoints. Simplicial complexes are then constructed across a range of threshold distances, and structures which persist over a wide range of thresholds are extracted as significant topological features. Methods for extending these tools to fundamentally directed data (e.g. causal information flow on networks) will be proposed, discussed, and criticized.
Anqi Qiu, Hong Kong Polytechnic University, China
Title: Simplicial Attention Network for Complex Data
Abstract: Graph neural networks have been effective in learning the representation of relationships among nodes of a graph. However, the complexity of graph-structured data often goes beyond those defined simply on the nodes. Our study breaks new ground by treating a graph as a simplicial complex, thus allowing graph-structured data to be defined on any k-dimensional simplex. We propose a distinct graph attention network, the Simplicial Attention Network (SAN), employing Hodge-Laplacian (HL) operators and attention mechanisms. This network adeptly learns varied signal representations across k-dimensional simplices. The cornerstone of SAN lies in two innovative components: convolutional filters and attention pooling operators, implemented on n-dimensional simplices. The convolutional filters function on k-dimensional simplices, utilizing their unique topology encoded by the HL operator. These filters operate within the spectral domain of the k-th HL operator, demanding the calculation of k-th HL eigenfunctions. However, for larger graphs, this process can be computationally onerous. We mitigate this by introducing a generic polynomial approximation for HL-filters, showcasing intriguing spatial location properties relative to the polynomial order. Furthermore, we propose a pooling operator, designed by coarsening k-dimensional simplices and amalgamating features defined on these simplices through attention mechanisms. These mechanisms include self-attention via a transformer, and cross-attention through a projection operator that learns the topological interconnection between k-1 and k-dimensional simplices. HL-HGAT, tested across diverse graph applications, including the Travel Salesman Problem, prediction of molecular properties and peptide functions, image classification, and cognitive and age prediction, outperforms standard GNN techniques.
June 20 11:00-Noon
Plenary Talk
Paul M. Thompson, Director at ENIGMA, Associate Director, USC Mark and Mary Stevens Neuroimaging and Informatics Institute, University of Southern California
June 20 Noon-2:00pm
Lunch break & Poster Session
June 20 2:00-3:30am
Functional Brain Imaging
The session explores the current methodological issues in functional brain imaging (fMRI). The session will be chaired by Jean-Baptise Poline, Montreal Neurological Institute (MNI).
B.T. Thomas Yeo, National University of Singapore, Singapore
Title: Optimizing Scan Time & Sample Size to Improve Individual-Level Prediction
Abstract: A pervasive dilemma in neuroimaging is whether to prioritize sample size or scan duration given fixed resources. Here, we systematically investigate this trade-off in the context of brain-wide association studies (BWAS) using resting-state functional magnetic resonance imaging (fMRI). We find that total scan duration (sample size × scan duration per participant) robustly explains individual-level phenotypic prediction accuracy via a logarithmic model, suggesting that sample size and scan duration are broadly interchangeable. The returns of scan duration eventually diminish relative to sample size, which we explain with principled theoretical derivations. When accounting for fixed costs associated with each participant (e.g., recruitment, non-imaging measures), we find that prediction accuracy in small-scale BWAS might benefit from much longer scan durations (>50 min) than typically assumed. Most existing large-scale studies might also have benefited from smaller sample sizes with longer scan durations. These results replicate across phenotypic domains (e.g., cognition and mental health) from two large-scale datasets with different algorithms and metrics. Overall, our study emphasizes the importance of scan time, which is ignored in standard power calculations. Standard power calculations inevitably maximize sample size at the expense of scan duration. The resulting prediction accuracies are likely lower than would be produced with alternate designs, thus impeding scientific discovery.
Hernando Ombao, King Abdullah University of Science and Technology (KAUST), Saudi Arabia
Title: Causality Inference via Spectral Entropy
Abstract: The goal is to develop statistical inference on spectral-specific measures of dependence between nodes in a brain network. Brain connectivity reflects how different regions of the brain interact during resting state or the execution of some task. Here, connectivity will be explored via an information-theoretic causal measure called transfer entropy (TE). To improve utility of TE in brain signal analysis, we propose a novel methodology to capture cross-channel information transfer in the frequency domain by examining dependence between oscillations. We introduce a new measure, the spectral transfer entropy (STE), to quantify the magnitude and direction of information flow from a certain frequency-band oscillation of a channel to an oscillation of another channel. The main advantage of our proposed approach is that it allows adjustments for multiple comparisons to control family-wise error rate. Another novel contribution is a simple yet efficient estimation method based on vine copula theory that enables estimates to capture zero (boundary point) without the need for bias adjustments. With the vine copula representation, a null copula model, which exhibits zero STE, is defined, making significance testing for STE straightforward through a standard resampling approach. Lastly, we illustrate the advantage of our proposed measure through some numerical experiments and provide interesting and novel findings on the analysis of EEG recordings linked to a visual task. This is joint work with Paolo Redondo (Ph.D. student) and Raphael Huser.
Woo-Young Ahn, Seoul National University
Title: Understanding real-world addictive behaviors using naturalistic paradigms, neuroimaging, and AI methods
Abstract: Computational psychiatric research has played a pivotal role in revealing the neurocognitive mechanisms underlying psychiatric disorders including addictive behaviors. However, many laboratory tasks and traditional analysis techniques often fall short in capturing real-world behaviors and their corresponding neural correlates due to their oversimplified nature. In this talk, I will discuss recent advances in computational psychiatry, with a particular focus on addiction. I will highlight recent studies employing naturalistic paradigms, fMRI, and AI techniques for better capturing the dynamic aspects of neurocognitive processes.
June 20 3:30-3:00pm Break
June 20 4:00-5:00pm
Brain Connectivity
The session explores methodical development in brain connectivity
Jaroslaw Harezlak, Indiana University - Bloomington, USA
Title: Incorporation of brain spatial and connectivity-based information in statistical regularization
Abstract: Prior information used in a principled manner can improve the quality of the regression coefficient estimation. We give an overview of the statistical regularization methods incorporating structural connectivity derived from the Diffusion Weighted Brain Imaging (DWI), functional connectivity derived from the functional Magnetic Resonance Imaging (fMRI) and cortical spatial distance derived from the structural MRI (sMRI) in the penalized approach. Extensions of the previously developed methods informing the estimation of the regression coefficients incorporate such information via a Graph Laplacian and a Hodge Laplacian matrices. The penalty terms are constructed via a weighted sum of the Laplacian matrices. Simulation studies show greatly improved regression coefficient estimation accuracy. We apply our approach to the data collected in the Human Connectome Project showing associations between the structural and functional MRI measures and neurocognitive outcomes.
Chee Ming Ting, Monash University, Malaysia
Title: Deep Generative Modelling of Brain Connectome
Abstract: The study of human brain connectome including structural connectivity (SC) and functional connectivity (FC) provide insights into healthy and disordered brain structures and functions. Recent advances in deep learning (DL) techniques have shown promises for prediction and characterization of brain connectivity. However, DL can be hindered by typically limited amount of neuroimaging data available for model building. For example, for constructing SC, tractography analysis requires diffusion MRI data with a high signal-to-noise ratio and spatial-angular resolution which is typically expensive to acquire. In this talk, I will introduce our recent effort on synthesizing realistic brain connectivity using deep learning-based generative models, with applications to data augmentation to improve DL performance. We first present a graph-regularized manifold-aware conditional Wasserstein generative adversarial networks (GR-SPD-GAN) for FC data generation that can preserve the symmetric positive definite (SPD) structure and encode inter-subject similarity in the generated FC matrices. Conditional models are used to facilitate generation according to different experimental groups (diseased, control and different ages). We also consider extensions to synthesizing dynamic FC using the TimeGAN. We show that augmentation with the generated static and dynamic FC data substantially improves classification of major depression from resting-state fMRI. Next, we present a deep autoencoder capable of generating bundle-specific streamlines directly from anatomical T1-weighted MRI. The method can generate streamlines without resorting to voxel-to-voxel tracing, hence sidestepping challenges involved in tracing across complex configurations such as crossings, bending, and bottlenecks where multiple bundles converge before re-emerging. We show that tract streamlines can be estimated directly from anatomical MRI, allowing tractography in the absence of diffusion MRI and enabling anatomy tractography to guide diffusion tractography.
June 20 5:00-6:00pm
Imaging Genetics
The session explores the methodological issues in imaging genetics
Li Shen, University of Pennsylvania, USA
Title: Enhancing Alzheimer’s Research with AI and Informatics: Strategies for Mining Brain Imaging Genomics Data
Abstract: Alzheimer’s disease (AD) is a major public health crisis, affecting millions worldwide, with a substantial social and economic burden. Effective strategies are urgently needed to discover new AD genes for disease modeling and drug development. Studying AD genetics using multimodal imaging and multi-omics data is becoming a rapidly growing field with distinct advantages in power over categorical diagnosis under imaging and omics traits as well as in capturing new insights into disease mechanism and heterogeneity from genetic determinants to omics-level molecular signatures, to brain imaging biomarkers, and to AD outcomes. In this talk, we will discuss AI and informatics strategies for discovering AD risk and protective genes through analyzing multidimensional genetics, omics, imaging and outcome data from landmark and local AD biobanks. We show that the wide availability of these rich biobank data, coupled with advances in trustworthy AI and informatics, provides enormous opportunities to contribute significantly to gene and biomarker discovery in AD and to impact the development of new diagnostic, therapeutic and preventative approaches.
Yu-Ping Wang, Tulane University, USA
Title: Interpretable multimodal deep learning for the integration of brain imaging and genomics data with application to brain development study
Abstract: Multi-modal deep network-based models have been developed to integrate complementary information from multi-modal datasets while capture their complex relationships. However, deep learning models are often difficult to interpret, bringing about challenges for uncovering biological mechanisms using these models. In this work, we develop an interpretable multimodal deep learning-based integration model to perform cognitive group classification and result interpretation simultaneously. We incorporate gradient weighted class activation mapping (Grad-CAM) into convolutional collaborative learning to obtain feature maps in a multi-modal convolution network. Recently, we further improve the Grad-CAM with Score-CAM for better feature interpretation. The proposed model can generate interpretable activation maps to quantify pixel-level contributions of the input imaging features. Moreover, the estimated activation maps are class-specific, which can therefore facilitate the identification of biomarkers such as brain regions and functional connectivity networks underlying different populations such as age, gender and cognitive groups. Finally, we apply and validate the model in the study of brain development with integrative analysis of multi-modal brain imaging and genomics data. We demonstrate its successful application to both the classification of cognitive function groups and the discovery of underlying genetic mechanisms.
June 21 9:30-10:45am
Special Lecture
Jean-Baptise Poline, Montreal Neurological Institute, McGill University, Canada
Title: An epistemological perspective on the field of neuroimaging statistics and on its future
With the emergence of neuroimaging as a subfield of cognitive and clinical neuroscience 25 years ago came a community of statistics or applied mathematics aiming to i) solve for the challenges brought by these high dimensional and complex data ii) propose a number of "new questions" or research directions that are only possible through advanced analytical techniques. Today machine learning and new deep learning models are potential solutions to some of the challenges traditionally tackled by classical statistics.
In this short talk I will propose a retrospective view of this field, highlighting both some of the notable successes that the statistical community encountered within neuroimaging but also the areas where, after 25 years, progress has been slow or even non existent. I will relate this view to the Popperian and Khunian epistemological views and lay out some yet unfaced challenges and areas of future research.
Multimodal Methods
Neda Jahanshad, University of Southern California, USA
Title: Working with continuously larger-scale neuroimaging analyses
Abstract: Large-scale neuroimaging collaborations have been established to help tackle the reproducibility crisis in brain imaging research. These collaborations often involved dozens of research groups pooling together neuroimaging, clinical and computational resources together to amass study sample sizes on the order of tens of thousands of participants. However, data sharing and data heterogeneity remain necessary and important hurdles on the path towards efficient data analysis. Here, we will discuss common clinical and technical challenges faced when working with distributed and diverse data as well as the collaborative and neuroinformatic backbone of these collaborations. The discussion will emphasize the growing role that data science methods including machine learning and artificial intelligence are playing, and will play, in multiple aspects of collaborative neurosciences. In this talk, the methods and applications discussed will be in the context of the ENIGMA consortium, but these may be applicable across many multi-site collaborative studies.
Vince D. Calhoun, Tri-institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS), USA
Title: Functionally Adaptive Structural Basis Sets of the Brain: A Dynamic Fusion Approach
Abstract: The precise relationship between brain structure and dynamic neural function has long been an open area of research, and recently, a topic of much debate. Whether investigated through the lens of interregional white matter connectivity or cortical surface morphology, the common thread that links many studies in this subfield of research is the focus on identifying a singular structural basis set, upon which functional activation signals are reconstructed to define the linkage between structure and function. Such approaches are limited in two respects; first, these basis sets are defined solely upon structural data and ignore the influence of functional coupling entirely, and second, these approaches operate on the somewhat narrow assumption that a single structure-function coupling governs the activity of the whole brain at all times. The first limitation can be addressed with the use of multimodal data fusion, which identifies hidden linkages between structural and functional brain imaging data; however, many multimodal fusion approaches still necessitate functional data to be heavily summarized over the time dimension, resulting in temporally rigid structure-function linkages. Regarding the second limitation, given that functional brain activity and connectivity vary over multiple timescales, it is natural to consider this also might be true of structure-function couplings. Here, we introduce dynamic fusion, implemented as an ICA-based symmetric fusion approach, which enables flexible, time-resolved linkages between structure and function utilizing dynamic functional connectivity (dFNC) states. We show evidence that challenges current claims regarding structural basis sets and suggests that temporally evolving structural basis sets can better reflect dynamic functional manifolds and better capture diagnostically relevant structure-functional coupling than traditionally computed structural bases.
June 21 10:45-11:00am Break
June 21 11:00-12:00am
Keith Worsley Lecture
Named after Professor Keith J. Worsley at McGill University and University of Chicago.
Michael I. Miller, Bessie Darling Massey Professor and Director of Biomedical Engineering, co-director of the Kavli Neuroscience Discovery Institute at Johns Hopkins University
Title: Molecular CA: From Euler to Dirac
Abstract: I will discuss progress over the years in Molecular Computational Anatomy (CA), representing gross anatomy as diffeomorphic transformations of templates.Recent results will be presented extending CA from the tissue scales via continuum mechanics for MRI to the molecular scales via generalized functions for spatial transcriptomics. Applications to understanding the spatio-temporal organization of the progression of Alzheimer's disease in digital pathology, from high field MRI to misfolded Tau proteins, will be presented.
References:
https://spj.science.org/doi/10.34133/2022/9868673
https://www.sciencedirect.com/science/article/abs/pii/S1361841523003286
June 21 Noon-2:00pm
Lunch Break & Poster Session
June 21 2:00-3:30am
Structural Brain Imaging
The session explores current issues in structural brain imaging. The session will be chaired by Anqi Qiu.
Ilwoo Lyu, Pohang University of Science and Technology (POSTECH), Korea
Title: Advancing Cortical Surface Annotation for Brain Shape Analysis
Abtract: Accurate labeling of cortical surfaces plays a pivotal role in unraveling their intricate structure and understanding their functions. Despite recent advancements in machine learning, the limited availability of annotated data limits their full potential. In this talk, we present our recent efforts aimed at achieving accurate cortical surface annotation even in the face of data scarcity. We begin by introducing our spherical data augmentation technique, which significantly enhances the learning process. Subsequently, we delve into the intricacies of spherical harmonics-based neural networks, meticulously engineered to adeptly capture the unique geometry and intricate patterns inherent in the cortical surface. Building upon this foundation, we unveil enhanced iterations of these networks, distinguished by their proficiency in efficient learning through the simplification of cortical features. We also show practical applications of these methods, illustrating their efficacy in addressing difficult challenges such as cortical parcellation and sulcal region labeling. Through these applications, we emphasize the effectiveness of our approaches in mitigating data scarcity.
Michael Breakspear, University of Newcastle, Australia
Title: Generation of surrogate brain maps preserving spatial autocorrelation through random rotation of geometric eigenmodes
Abstract: The brain expresses activity in complex spatiotemporal patterns, reflecting spatially distributed cytoarchitectural, biochemical, and genetic influences. The correspondence between these multimodal “brain maps” may reflect underlying causal pathways and is hence a topic of substantial interest. However, these maps possess intrinsic smoothness (spatial autocorrelation, SA) which can inflate spurious cross-correlations, leading to false positive associations. Identifying non-trivial associations requires knowledge about the distribution of correlations that arise by chance in the presence of SA. This null distribution can be generated from an ensemble of surrogate brain maps that preserve internal SA but break correlations between maps. Here we introduce “eigenstrapping”, using a spectral decomposition to derive geometric eigenmodes from cortical and subcortical surfaces, then randomly rotating these eigenmodes to produce SA-preserving surrogate brain maps. We show that these surrogates appropriately represent the null distribution of chance pairwise correlations, with similar or superior false positive control than current state-of-the-art. Eigenstrapping is fast, eschews the need for parametric assumptions about the nature of the SA, and works with maps defined on smooth surfaces with or without a boundary. It generalizes to broader classes of null models than existing techniques, offering a unified approach for inference on cortical and subcortical maps, spatiotemporal processes, and complex patterns possessing higher-order correlations.
Moo K. Chung, University of Wisconsin-Madison, USA
Title: Differential Structural and Functional Connectivity in Sulci and Gryi
Abstract: The cerebral cortex of the human brain, characterized by its complex folding, comprises both concave regions known as sulci and convex regions called gyri, which resembles the Turing pattern often observed in a reaction-diffusion process. Distinct patterns of functional and structural connectivity are observed between these regions. In a preliminary study involving T1-MRI, diffusion-MRI, and resting-state fMRI from approximately 400 subjects, we identified 70% higher white matter fiber density in gyri compared to sulci. This finding correlates with markedly stronger functional connectivity within gyral regions (correlation 0.981) than within sulcal regions (correlation 0.969) as well (p <0.000001). The negligible connectivity between gyri and sulci (correlation 0.0027) indicates a functional segregation between these cortical areas. We will try to elucidate how the dynamic patterns of human brain networks at rest are influenced by the gyral and sulcal architecture. Additionally, we investigate the implications of these connectivity disparities on the overall dynamics of brain networks. Part of talk is based on draft arXiv:2201.00087.
June 21 3:30-4:00 Break
June 21 4:00-5:30pm
Statistical Methods
The session explores the methodological issues in modeling brain images
Thomas Nichols, Big Data Instittue, University of Oxford, UK
Title: Understanding the Changing Brain: Longitudinal Models for Neuroimaging Data
Abstract: Longitudinal neuroimaging studies offer a unique window into understanding how the human brain evolves over time, shedding light on developmental processes, the impact of aging, and the influence of various factors on brain structure and function. In this talk I will provide an overview of longitudinal modelling techniques for neuroimaging data, elaborating on what the standard tools (SPM & FSL) can and cannot do, and discuss a growing collection of tools that can flexibly model longitudinal brain imaging data. Specifically I will review standard linear mixed effect (LME) models, and the tools that can fit such models voxel-wise/element-wise. However, LMEs require iterative optimisation which can be problematic for large datasets, so I will also discuss fast, non-iterative methods that provide inferences that are valid and are almost as efficient as a proper LME model. I will illustrate these models on data from the UK Biobank and Adolescent Brain Cognitive Development project.
Bertrand Thirion, Head of Science, Inria Saclay, France
Title: Variable importance for population studies in brain imaging.
Abstract: Variable importance assessment has become a crucial step in machine-learning applications when using complex learners, such as deep neural networks, on large-scale data. Removal-based importance assessment is currently the reference approach, particularly when statistical guarantees are sought to justify variable inclusion. It is often implemented with variable permutation schemes. On the flip side, these approaches risk misidentifying unimportant variables as important in the presence of correlations among covariates. Here we develop a systematic approach for studying Conditional Permutation Importance (CPI) that is model agnostic and computationally lean, as well as reusable benchmarks of state-of-the-art variable importance estimators. We show theoretically and empirically that CPI overcomes the limitations of standard permutation importance by providing accurate type-I error control. When used with a deep neural network, CPI consistently showed top accuracy across benchmarks. An experiment on real-world data analysis in a large-scale medical dataset showed that CPI provides a more parsimonious selection of statistically significant variables. Our results suggest that CPI can be readily used as drop-in replacement for permutation-based methods.We show that CPI provides meaningful assessments of variable importance on several neuroimaging datasets.
Amanda Mejia, Indiana University - Bloomington, USA
Title: The hidden cost of stringent motion scrubbing
Abstract: Motion scrubbing is a common practice to mitigate the influence of subject head motion on functional connectivity (FC) analyses. Stringent motion scrubbing is often endorsed for more thorough noise removal, but at what cost? A basic statistical tenant is that estimation error is determined by population variance and sample size, both of which are decreased through scrubbing. Depending on these two competing forces, scrubbing may ultimately improve or worsen estimation error. Here, we use data from the Human Connectome Project (HCP) retest dataset to establish long-run ground truth FC for 42 subjects. We examine the effect of motion scrubbing on estimation error of FC and quantify the increase in scan time required to maintain accuracy due to over-scrubbing. Since sessions are often excluded after scrubbing if insufficient scan time remains, we also examine the effect of motion scrubbing on brain-wide association studies (BWAS), considering the reduced sample size due to scrubbing. Compared with more lenient scrubbing, we find that stringent motion scrubbing ultimately results in less accurate FC and necessitates approximately 8% longer scans to maintain accuracy. This is driven by its high censoring rates: nearly 18% in healthy adults, a low-motion population, versus less than 5% with more lenient methods. We also find that stringent motion scrubbing induces greater downward bias of brain-behavior correlations unless low-duration sessions are excluded, which comes at the cost of higher variance. We conclude that lenient motion scrubbing strikes a near-optimal balance of noise reduction and data retention, ultimately facilitating smaller data collection budgets and/or larger sample sizes for BWAS.