Modern Statistical Methods for Image Analysis

Simon Vandekar Chair
Vanderbilt University
 
Sunday, Aug 4: 2:00 PM - 3:50 PM
5013 
Contributed Papers 
Oregon Convention Center 
Room: CC-D140 

Main Sponsor

Section on Statistics in Imaging

Presentations

A Chi-Square Type Test for Time-Invariant Fiber Pathways of the Brain

A longitudinal diffusion tensor imaging (DTI) study on a single
brain can be remarkably useful to probe white matter fiber connectivity that
may or may not be stable over time. We consider a novel testing problem
where the null hypothesis states that the trajectories of a coherently oriented
fiber population remain the same over a fixed period of time. Compared to
other applications that use changes in DTI scalar metrics over time, our test is
focused on the partial derivative of the continuous ensemble of fiber trajectories
with respect to time. The test statistic is shown to have the limiting chi-square
distribution under the null hypothesis. The power of the test is demonstrated
using Monte Carlo simulations based on both the theoretical and empirical
critical values. The proposed method is applied to a longitudinal DTI study
of a normal brain. 

Keywords

Functional central limit theorem

Nadaraya-Watson type kernel estimator


White matter fiber tractography 

View Abstract 1963

Co-Author(s)

Juna Goo, Boise State University
David C Zhu, Michigan State University

First Author

Lyudmila Sakhanenko, Michigan State University

Presenting Author

Lyudmila Sakhanenko, Michigan State University

A High-Dimensional Latent Regression Item Response Model for Psychometric-Neuroimaging Association Studies

Neuroimaging data are increasingly used in evaluating cognitive patterns underlying psychometric measurements. Challenges in such analysis include: 1. Latent cognitive traits are not directly observable. 2. Neuroimaging data are high dimensional. To address these issues, we propose a novel high-dimensional latent regression item response model. Unlike conventional item response models that model the latent trait solely based on the observed psychometric measurements, our model associates latent traits to high-dimensional neuroimaging features and filters the important features predictive of the latent traits and outcome measurements. We implemented a high-dimensional EM algorithm, it employs a Metropolis-Hasting resampling mechanism in E-steps to evaluate the latent traits and a regularization regression in M-steps to select predictive features. Through extensive simulations, our approach showed great performance in feature selection and outcome prediction. The method is applied to the National Alzheimer's Coordinating Center psychometric and neuroimaging datasets to learn the association patterns between cognitive abilities and brain regions. 

Keywords

High-dimensional data

Item response model

Metropolis-Hasting algorithm

Neuroimaging

Psychometrics measurements

Variable selection 

Abstracts


Co-Author

Xiaohang Mei

Presenting Author

Xiaohang Mei

Analysis of Independent and Correlated Imaging Features using Scalar-on-matrix Regression

Image features that characterize objects from kidney biopsies may offer insight into disease prognosis as novel biomarkers. For each subject, we construct a matrix of image features that are measured for each object from that subject's biopsy. We proposed the CLUstering Structured laSSO (CLUSSO), a novel scalar-on-matrix regression method that allows for unbalanced numbers of objects across subjects, to predict scalar outcomes from matrices of independent image features. CLUSSO averages images feature values within subgroups of objects as determined by cluster analysis. We showed through simulations that CLUSSO has fewer false positives (FPs) and more true positives for identifying truly predictive features relative to a naive method that averages feature values across all objects. To handle correlated image features, we developed the Random CLUSSO, an extension of CLUSSO that averages estimated feature coefficients across bootstrapped samples with subsampled features. These methods are applied to tubular image features of kidney biopsies from glomerular disease patients to predict kidney function. 

Keywords

Computational pathology

Image features

Scalar-on-matrix regression

Unsupervised clustering

Variable selection 

View Abstract 2014

Co-Author(s)

Fan Fan, Department of Biomedical Engineering, Emory University and Georgia Institute of Technology
Laura Barisoni, Division of AI and Computational Pathology, Department of Pathology, Duke University
Andrew Janowczyk, Department of Biomedical Engineering, Emory University
Jarcy Zee, University of Pennsylvania

First Author

Jeremy Rubin, University of Pennsylvania

Presenting Author

Jeremy Rubin, University of Pennsylvania

Generalized Linear Mixed Model with Matrix Response of Brain Imaging Data

In recent years, there has been a rapid emergence of multiple-subject network longitudinal data, characterized by individual connectivity matrices for each subject, mapped across a consistent set of nodes, and accompanied by information on subject-specific covariates. We introduce a novel generalized linear mixed model, designed to treat these networks as matrix-valued responses and leverage subject covariates as predictors. Our model captures the population-level connectivity patterns via a low-rank intercept matrix and articulates the impact of subject covariates using a sparse slope tensor. We have developed an efficient MCEM algorithm embedding alternating gradient descent method for parameter estimation and edge selection. The effectiveness of our approach is validated through simulations through various data settings and applied in two brain connectivity studies, showcasing its practical utility in contemporary network analysis. Extensive simulation studies demonstrate that our proposed model overperforms the element-wise penalized generalized linear mixed models with LASSO or SCAD penalty. 

Keywords

Generalized Linear Mixed Model

Monte Carlo EM Algorithm

Longitudinal data

Matrix Response

Low Rank Structure

Tensor Slope 

View Abstract 3279

Co-Author

Quefeng Li, University of North Carolina Chapel Hill

First Author

Zhentao Yu

Presenting Author

Zhentao Yu

Microstructural Quantile Profile for White Matter Tracts

In vivo fiber tractography is a 3D reconstruction technique to assess neural tracts using data collected by dMRI. The fiber tract obtained from the technique can be used for studying the brain's anatomy and its associations to function of interest covariates. Recent machine learning methods can efficiently identify subject-level white matter tracts. However, analyzing the scalar clinical/psychological factors (e.g., cognitive score) and fiber tracts is difficult. The current methods use high-level summary statistics of fiber tract; thus, the relationship investigation is based on traditional regression models. In this paper, we find the FA quantiles over the points of a fiber tract (Microstructural Quantile Profile) can be used to differentiate the effect of function of interest covariates. We adopted and further developed the quantile regression methodology with clustered data to infer the relationship between Microstructural Quantile Profile and scalar clinical/psychological factors. Insightful spatial findings were provided via our new approach. The method is more robust in identifying the relationship between fiber tract and scalar clinical/psychological factors. 

Keywords

fiber tractography

diffusion MRI

quantile regression

microstructural quantile profile 

View Abstract 2796

Co-Author

Lauren O'Donnell, Brigham and Women's Hospital, Harvard Medical School

First Author

Zhou Lan, Brigham and Women's Hospital, Harvard Medical School

Presenting Author

Zhou Lan, Brigham and Women's Hospital, Harvard Medical School

Semiparametric confidence sets for arbitrary effect sizes in longitudinal neuroimaging

Longitudinal data are increasingly prevalent in psychiatric neuroimaging as investigators aim to explore the relationships between biological factors and symptom variations on an individual level. This study addresses the complexity of longitudinal neuroimaging data to construct spatial confidence sets using a flexible semiparametric bootstrap joint (sPBJ) spatial extent inference (SEI) method. Our method involves robust estimation of the spatial covariance function based on the generalized estimating equation. We obtain more efficient effect size estimates by concurrently estimating the exchangeable working covariance and using the sPBJ bootstrap to determine the joint distribution of effect size across voxels. The bootstrap procedure is used to construct confidence sets for the effect size parameter. These confidence sets can identify the target and null regions of the image where the effect size is above or below given thresholds, respectively, with high probability. We evaluate the coverage of the proposed procedures using realistic simulations. This comprehensive approach, integrated into the pbj R package, offers a robust tool for analyzing repeated neuroimaging measurements. 

Keywords

Effect size

Confidence sets

Longitudinal neuroimaging data 

View Abstract 3472

Co-Author(s)

Kenneth Liao, Vanderbilt University Medical Center
Maureen McHugo, University of Colorado Medicine
Anna Huang, Vanderbilt University Medical Center
Kristan Armstrong, Vanderbilt University Medical Center
Suzanne Avery, Vanderbilt University Medical Center
Stephan Heckers, Vanderbilt University Medical Center
Simon Vandekar, Vanderbilt University

First Author

Xinyu Zhang

Presenting Author

Xinyu Zhang