Introduction
Our statistical methodology research and collaborations focus on various biomedical imaging areas. Primarily, we work with brain imaging data, including both functional and structural neuroimaging modalities. We have also worked in cardiac imaging and in cancer applications, including brain tumor, breast and prostate cancer imaging.
Generally, functional neuroimaging is a method for mapping measures of localized brain activity in vivo. We apply our methods to functional neuroimaging applications that seek to characterize changes in distributed neural processing associated with psychiatric disorders, drug cravings, behaviors, emotions, and decision-making. CBIS also conducts research on brain networks and functional connectivity, which provides insights into the relationships between different brain regions when performing a particular task or during resting state. In recent years, the collection of multimodal neuroimaging (fMRI, sMRI and DTI, etc.) has become common practice to provide different views of brain function or structure. CBIS has been working on developing effective analytical tools for fusing multimodal imaging to obtain more accurate and informative results on brain function and connectivity.
Our research seeks state-of-the-art biostatistical methods that are applicable (1) to describe functional associations between brain regions, (2) to determine functional connectivity and hierarchical networks in the brain, (3) to make inferences concerning task-related changes in brain activity that ultimately produce maps revealing distributed patterns of task/activity associations, (4) to address various prediction objectives and (5) to conduct integrative analysis of multimodal imaging data. Specifically, we have applied our statistical methods to help better understand the neural correlates underlying.
brain disorders such as depression, PTSD, Alzheimer’s disease, and schizophrenia
task-related and resting-state functional connectivity differences in Zen meditators
social anxiety disorder and its response to pharmacotherapy
the sensitivity to ethical issues related to justice and care
cue-induced nicotine craving
cocaine dependence
We briefly describe our statistical methodology research below.
Research
Independent Component Analysis
Independent component analysis (ICA) is the most commonly used computational tool for identifying and characterizing underlying brain functional networks. One of the challenging research topics in ICA is how to perform group ICA for multi-subject imaging studies. Our research on group ICA methodology focused on development of probabilistic group ICA framework for estimating brain functional networks based on multi-subject fMRI data. We have proposed general temporal-concatenation group ICA (TC-GICA) models that can accommodate different types of between-subject variability in temporal responses and model wide variety of neural signals under different experimental tasks (Guo and Pagnoni, 2008; Guo, Biometrics, 2011). In addition to TC-GICA, we have developed a novel hierarchical group ICA method to formally model subject-specific effects in both temporal and spatial domains in fMRI data (Guo and Tang, 2013). Furthermore, we have recently developed a hierarchical covariate-adjusted ICA (hc-ICA) model that provides a formal statistical framework for estimating covariate effects and testing differences between brain functional networks. Our method provides a more reliable and powerful statistical tool for evaluating group differences in brain functional networks while appropriately controlling for potential confounding factors. We plan to release a matlab toolbox for hc-ICA in 2016.
Fig1. p-values for testing group differences between PTSD+ and PTSD- subjects in DMN
Reference
Shi, R and Guo, Y. (2016+). Investigating Differences in Brain Functional Networks Using a Hierarchical Covariate-adjusted ICA Model. Annals of Applied Statistics. Accepted.
Guo, Y and Tang, Li. (2013). A hierarchical probabilistic model for group independent component analysis in fMRI studies, Biometrics, doi: 10.1111/biom.12068.
Guo Y (2011). A general probabilistic model for group independent component analysis and its estimation methods. Biometrics. 67(4): 1532-1542.
Guo Y and Pagnoni G (2008). A unified framework for group independent component analysis for multi-subject fMRI data. NeuroImage 42: 1078-1093. Listed in ScienceDirectðs Top 25 hottest articles of NeuroImage between July-Sep. 2008.
Guo Y (2008). Group Independent Component Analysis of Multi-subject fMRI data: Connections and Distinctions between Two Methods. IEEE Proceedings of the 2008 International Conference on BioMedical Engineering and Informatics v2: 748-752.
C
Meta-analysis
In recent years, the number of imaging studies on the structure and function of human brains has grown significantly. Most of these studies recruit a number of subjects for brain scans and apply various statistical methods to assess the association patterns between brain locations and human behavior as well as neurological disease or dysfunction. However, the limited sample sizes in many studies may lead to low statistical power. At the same time, results from different literatures on similar topics could sometimes be inconsistent.
To address these issues, researchers adopt meta-analysis methods to integrate scientific findings across different studies. Meta-analysis is typically performed on the spatial coordinates of peaks of activation (peak coordinates), reported in the standard MNI or Talairach coordinate systems, and combined across studies. This information is typically provided in most neuroimaging papers and simple transformations between the two standard spaces exist. Current statistical methods for imaging meta-analysis mainly focus on identifying activated brain locations for one type of study and do not incorporate brain-wide correlations between different studies. We are now developing systematic Bayesian frameworks to: 1) jointly model the brain activation patterns for multiple studies; 2) use hierarchical models to borrow information across studies; 3) capture location specific spatial dynamics of between study correlations. A primary scientific question that our method can address is how activations related to different emotions are correlated at different parts of brain.
Reference
Kang, J., Nichols, T.E., Wager, T.D. Johnson, T.D. (2014) A Bayesian hierarchical spatial point process model for multi-type neuroimaging meta-analysis, Annals of Applied Statistics, 8 (3), 1800.
Xue, W., Kang, J., Bowman F. D., Guo, J. Wager, T. (2014) Identifying Functional Co-activation Patterns in Neuroimaging Studies via Poisson Graphical Models, Biometrics, doi: 10.1111/biom.12216.
Kang, J., Johnson, T.D. (2014) A slice sampler for the hierarchical Poisson/gamma random field model, Proceedings for Perspectives on Big Data Analysis, Contemporary Mathematics, American Mathematical Society, 622
Taylor, S. F., Kang, J., Brege, I. S., Tso, I. F., Hosanagar, A., Johnson, T. D. (2012) Meta-analysis of functional neuroimaging studies of emotion perception and experience schizophrenia, Biological Psychiatry, 71(2): 136-45.
Kang, J., Johnson, T. D., Nichols, T. E., Wager, T. D. (2011). Meta analysis of functional neuroimaging data via Bayesian spatial point processes. Journal of the American Statistical Association, 106(493), 124-134
Multimodality Methods
We have developed several methods that combine modalities of neuroimaging data, namely fMRI and DTI data, to study the relationship between brain structure and function and to investigate the connectivity disruption pathways that characterize certain brain diseases. Resting-state and task-related brain activity, measured by fMRI, reflects the functional connectivity (FC) or associations between different brain regions. Diffusion tensor imaging (DTI), which enables the reconstruction and probabilistic quantification of major fiber tracts in the brain, provides structural connectivity (SC) information that may improve our understanding of FC. We have developed a method of anatomically-weighted FC (awFC), that implements a hierarchical clustering algorithm to identify FC networks, weighted by the evidence of underlying SC (Bowman et al., 2012). We have also proposed a novel statistical framework for measuring and testing the strength of SC (sSC) underlying FC networks estimated by data-driven methods like ICA. We found that the estimated FC networks with higher sSC also tend to be more reliably identified in ICA. We have also developed a unified Bayesian framework for analyzing FC utilizing the knowledge of associated structural connections, which extends an approach by Patel et al. (2006) that considers only functional data (Xue et al, 2015).
Reference
Bowman, F. D., Zhang, L., Derado, G., & Chen, S. (2012). Determining functional connectivity using fMRI data with diffusion-based anatomical weighting. NeuroImage, 62(3), 1769-1779
Xue, W., Bowman, F. D., Pileggi, A. V., & Mayer, A. R. (2015). A multimodal approach for determining brain networks by jointly modeling functional and structural connectivity. Frontiers in computational neuroscience, 9.
Patel, R. S., Bowman, F. D., & Rilling, J. K. (2006). Determining hierarchical functional networks from auditory stimuli fMRI. Human brain mapping, 27(5), 462-470.
Copyright © 2015 Center for Biomedical Imaging Statistics. All rights reserved.
CBI
Brain Networks and Connectivity
Recently, network analyses for fMRI data have emerged that characterize the functional relationships between brain regions. A typical neuroimaging network analysis involves defining brain regions (nodes), quantifying a measure of association between all pairs of brain regions (edges) to produce a connectivity matrix, thresholding these associations to obtain a more sparse connectivity matrix, and then calculating summary statistics that reflect properties of the network. Network summaries include metrics that reflect local or global communication ability (e.g., clustering coefficient, path length, and efficiency), centrality metrics (e.g., betweenness, closeness, and eigenvector centrality), and community structure (e.g., small-worldness).
Fig1. Wang et al., (2016). Highly Consistent Positive (a) and Negative Edges (b) between Partial Correlation and Full Correlation.
Functional connectivity (FC) is commonly measured by correlation, but this reflects the marginal association between two nodes, which could be mediated by connections to a third-party node. Partial correlation, however, measures direct FC by calculating correlation after regressing out the effects from all other nodes. We have developed a Dens-based method that allows fast and reliable computation of the partial correlation matrix, at a specified network sparsity level (Wang et al., 2016). We are also currently developing methods for network estimation via graphical models, as well as dynamic FC methods that measure the change in functional networks over time.
Fig2. Figure from Wang et al., (2016). Parial correlation matrices estimated at two sparsity levels. The sparse regularization has more shrinkage effects on negative functional connections and between-module connections.
Functional connectivity (FC) is commonly measured by correlation, but this reflects the marginal association between two nodes, which could be mediated by connections to a third-party node. Partial correlation, however, measures direct FC by calculating correlation after regressing out the effects from all other nodes. We have developed a Dens-based method that allows fast and reliable computation of the partial correlation matrix, at a specified network sparsity level (Wang et al., 2016). We are also currently developing methods for network estimation via graphical models, as well as dynamic FC methods that measure the change in functional networks over time.
Fig1. Wang et al., (2016). Highly Consistent Positive (a) and Negative Edges (b) between Partial Correlation and Full Correlation.
Reference
Wang, Y., Kang, J., Kemmer, P.B., & Guo, Y. (2016) An efficient and reliable statistical method for estimating functional connectivity in large scale brain networks using partial correlation. Frontiers in Neuroscience. In press.
Xue, W., Bowman, F. D., Pileggi, A. V., & Mayer, A. R. (2015). A multimodal approach for determining brain networks by jointly modeling functional and structural connectivity. Frontiers in computational neuroscience, 9.
Kemmer, P. B., Guo, Y., Wang, Y., & Pagnoni, G. (2015). Network-based characterization of brain functional connectivity in Zen practitioners. Frontiers in psychology, 6.
Copyright © 2015 Center for Biomedical Imaging Statistics. All rights reserved
Prediction Methods
There has been a strong interest in the neuroimaging community to utilize information in imaging data to predict individual disease status and treatment response. We developed the following statistical prediction methods based on brain images: a prediction method based on a Bayesian hierarchical model for forecasting future neural activity based on a subject’s baseline brain images and other individual characteristics (Guo, Bowman and Kilts, 2008), which can potentially help select optimal treatment plans for individual patients; a weighted cluster kernel PCA model for predicting subjects’ cognitive state using brain images (Guo, 2010) ; we also propose a statistical method for predicting individual’s future functional connectivity in longitudinal studies by using information from individual's baseline fMRI scans along with relevant subject characteristics, such as disease or treatment status (Dai and Guo, 2017). The proposed prediction method improves the accuracy of individualized prediction of connectivity by combining information from both group-level connectivity patterns as well as individual-specific connectivity features. It also offers statistical inference tools such as predictive intervals that help quantify the uncertainty or variability of the predicted outcomes.
Fig1. Guo Y. et al. (2008): Individualized predicted and observed post-treatment rCBF measurements under the low load condition for four subjects in the working memory study. (a) Predicted maps. Notable differences exist between the patients’ predicted brain responses to treatment. (b) Observed maps. There is satisfactory agreement between the predicted and observed post-treatment rCBF.
Reference
Dai, T and Guo, Y (2017). Predicting Individual Brain Functional Connectivity Using a Bayesian Hierarchical Model. NeuroImage. 147(15): 772–787. An earlier version of the paper was the Second-Place winner of the 2016 Student Paper Competition, American Statistical Association (ASA) Statistics in Imaging Section.
Guo Y, Bowman FD, Kilts C (2008). Predicting the brain response to treatment using a Bayesian Hierarchical model . Human Brain Mapping, 29(9): 1092-1109.
Guo Y (2010). A weighted cluster kernel PCA prediction model for multi-subject brain imaging data. Statistics And Its Interface. 3:103-111.
Derado, G., Bowman, F. D., and Zhang, L. (2012). Predicting Brain Activity using a Bayesian Spatial Model. Statistical Methods in Medical Research (DOI: 10.1177/0962280212448972).
Copyright © 2015 Center for Biomedical Imaging Statistics. All rights reserved.
CBI
Statistical impacts and processing for advanced image acquisition techniques
Advanced magnetic resonance imaging acquisitions can decrease scan time and reduce patient burden, but they can also lead to noise amplification. We examined the effect of multiband acquisition, also called simultaneous multislice acceleration, on estimates of functional connectivity in resting-state fMRI, with an emphasis on estimating brain connections in subcortical areas that play an important role in psychiatric and neurological disorders. Our research indicates single-band acquisitions with larger voxel sizes (e.g., 3.3 mm) should be used for studies focusing on subcortical areas. For whole brain studies, we recommend moderate multiband factors (multiband 4). We found an acquisition popularized by the Human Connectome Project (multiband 8) led to decreased effect sizes in subcortical regions. Overall, multiband acceleration led to decreases in functional connectivity in subcortical regions due to spatially varying noise amplification. Our research group has also worked with magnetic resonance spectroscopy, which can be used to measure the concentrations of brain metabolites. We collaborated with MR physicists to develop a novel method for combining data across channels in multi-channel head coils in magnetic resonance spectroscopy, and the method improves the signal-to-noise ratio.
Reference
Risk BB, Murden RJ, Wu J, Nebel MB, Venkataraman A, Zhang Z, Qiu D. Which multiband factor should you choose for your resting-state fMRI study?. Neuroimage. 2021 Jul 1;234:117965. PubMed Central PMCID: PMC8159874. https://www.sciencedirect.com/science/article/pii/S1053811921002421
Sung D, Risk BB, Owusu-Ansah M, Zhong X, Mao H, Fleischer CC. Optimized truncation to integrate multi-channel MRS data using rank-R singular value decomposition. NMR Biomed. 2020 Jul;33(7):e4297. PubMed Central PMCID: PMC7317403.
https://analyticalsciencejournals.onlinelibrary.wiley.com/doi/full/10.1002/nbm.4297
Risk BB, Kociuba MC, Rowe DB. Impacts of simultaneous multislice acquisition on sensitivity and specificity in fMRI. Neuroimage. 2018 May 15;172:538-553. PubMed PMID: 29408461.
https://www.sciencedirect.com/science/article/pii/S1053811918300788
Copyright © 2015 Center for Biomedical Imaging Statistics. All rights reserved.
CBI