2794
Educational Course - Half Day (4 hours)
Most functional MRI analysis pipelines begin by analyzing the functional time-series data for each voxel separately. This includes modeling the hemodynamic response for each subject across experimental conditions, but also the (multi-level) modeling of brain activation across (groups of) participants. Second, the brain maps resulting from the first step are analyzed to localize significant regions of brain activity or contrast, either at the individual or group level. That is, the statistics from each separately-analyzed voxel of the first step are combined to answer the question most researchers are interested in: where in the brain is my signal?
It is in this second step that spatial information (potentially) enters the equation. Often this information enters the analysis implicitly, for example as an assumption of the multiple comparisons correction (e.g., cluster-extent analysis) and is not used in any optimal fashion. Given the nature of the FMRI signal, and the hypotheses researchers have, there is much to be gained when spatial information is incorporated into the analysis explicitly, either at the first level, the second level, or both. In this course we will teach participants the state-of-the-art methods for incorporating spatial information into their analyses. We will cover extensions of the traditional analyses that take into account the spatial nature of data (True Discovery Proportions, All-Resolutions Inference (ARI), and threshold-free cluster enhancement (TFCE)). We will also cover explicit modeling approaches that incorporate spatial information (spatial Bayesian GLMs) and methods to explicitly test spatial hypotheses across studies (Confidence Sets).
In this half-day course participants will learn state-of-the-art approaches to make explicit use of spatial information, and will gain understanding about how these approaches improve the quality of analyses as well as how to implement them in their own analysis pipelines.
At the end of the course, participants will have knowledge on (1) the state-of-the-art methods for incorporating spatial information, (2) will be able to perform these analyses using software implementations, and (3) will be equipped to interpret the results from these analyses in a meaningful way. All course materials will be made freely available to all participants.
Neuroimaging researchers using (functional) MRI. The course is explicitly aimed at all researchers, from any level, doing (functional) MRI analysis.
Presentations
Massive univariate FMRI analysis, i.e. the general linear model (GLM), emerged in the 1990s, a time when computational considerations made more sophisticated models impractical or even intractable. Today, however, models that explicitly account for spatial dependence across voxels/vertices are easy to use, thanks to advances in computing power, Bayesian methods, and spatial statistics. In a spatial Bayesian GLM, a multivariate prior distribution on activation maps encodes expected similarities between neighboring locations. Spatial GLMs result in higher accuracy and power, both for individual and group-level analysis. Power is often high enough (even in individual-level analysis) that minimum scientifically relevant effect sizes (e.g. 1% signal change) can be considered when identifying statistically significant activations. Surface-based and subcortical parcel-constrained spatial GLMs are implemented in the BayesfMRI R package.
Presenter
Amanda Mejia, Indiana University Bllomington, IN
United States
Standard cluster-extent analysis of functional MRI data suffers from the spatial specificity paradox. This paradox states that the larger the cluster that is detected, the less we know about the location of activity within it. This is because a significant cluster means that “there is at least one voxel in this cluster active”, and not “all voxels in this cluster are active”. In essence, once we have detected a cluster, we don’t know exactly how many voxels within that cluster are significant, making localization problematic.
Recently developed methods based on closed testing provide a quantification of the number of truly active voxels within a cluster, called the True Discovery Proportion (TDP). The All-Resolutions Inference framework allows users to calculate the TDP for any cluster in the data without losing family-wise error (FWER) control. In practice, this means that a researcher can freely adjust the size of clusters, intersect clusters with anatomical masks, rerun analysis with a different cluster forming threshold, all with full FWER control. This flexibility requires an interactive implementation that goes beyond what is available in standard analysis packages. In this session we will teach how to use the ARIBrain Explorer software package (Python). For example, we will show users how to find clusters with a minimum TDP and how to interactively increase/decrease the size of each cluster to reach an optimal TDP. In addition, we’ll cover how to properly interpret and convey results.
Presenter
Lucas Peek, Leiden University Leiden, ZH
Netherlands
In a standard functional MRI data analysis pipeline the final step is often a correction of multiple comparisons. This step determines which regions of the brain show significant activation. The SPM software package provides users with extensive information on brain activity across different levels of analysis: set (whole brain), cluster (random-field based clusters), and voxelwise statistics. Interpretation of results differs across these levels, as for example cluster level statistics only tell researchers that there is activity in a cluster, but not where in that cluster activity is located. In other words, cluster-level statistics are about the extent of a cluster, but provide no information of the activity within a cluster. We recently developed a cluster-level compatible measure for the amount of signal within a cluster: clusterTDP. This measure estimates the amount of significantly activated voxels within a cluster, thus providing the researcher with valuable information on activity within the cluster. Here we introduce the clusterTDP SPM extension: a SPM compatible extension that provides True Discovery Proportions (TDP) for cluster-level analyses. The extension seamlessly integrates in the SPM pipeline making running the analysis and interpreting results easy. In this session we will show how to install and use the SPM extension, as well as how to interpret results.
Presenter
Xu Chen, University of Essex
School of Mathematics, Statistics and Actuarial Science
Colchester, Essex
United Kingdom
Threshold-free Cluster Enhancement (TFCE) is one of the most popular methods for multiple comparisons correction. The strength of TFCE lies in the use of spatial information of neighbouring voxels to improve estimates of brain activity. We recently developed a closed testing implementation of the TFCE statistic: Localized Cluster Enhancement (LCE). LCE leverages the strength of TFCE by using information from neighbouring voxels but provides improved error control and better localization. In this session we will show how to use the LCE software and how to interpret results.
Traditionally, uncertainty estimation in fMRI inference has focused on how signal magnitude at each specific voxel varies under repeated sampling. However, very little attention has been given to the variability in signal location. In this session, we shall provide a practical introduction to spatial confidence regions; regions which act as probabilistic bounds for the locale of observed clusters and excursion sets, allowing the researcher to assess how reliably a spatial region has been estimated.
The session shall cover the generation of confidence regions for excursion sets derived from %BOLD maps, standardized (Cohen’s D) effect size images, and conjunctions (overlaps) for both, and will use Jupyter notebooks to demonstrate a Python toolbox for confidence regions on the Human Connectome Project (HCP) dataset. By the end of this workshop, participants will have an understanding of why spatial confidence regions quantify uncertainty in fMRI inference and how to apply the methods in practice.