2025 Statistical Methods in Imaging Conference

May 19-21, 2025

KEYNOTE SPEAKERS

Monday, May 19, 2025 (3:00 pm – 4:15 pm)
Imaging via Signal Processing-Inspired Implicit Neural Representations
Richard Baraniuk, Professor, Rice University

Deep leaning-based Implicit Neural Representations (INRs) have emerged as a powerful new tool for modeling diverse signals and solving associated inverse problems. In this talk, we will overview two INR innovations inspired by classical signal processing concepts. First, we will show how INRs can be combined with Laplacian pyramids to obtain a multiscale INR framework for efficient processing of even gigapixel images. Second, we will show that wavelet functions provide an attractive family of activation nonlinearities for INRs solving a wide range of inverse problems, including image denoising, super resolution, tomography, and novel view synthesis. Throughout, harmonic analysis and geometry provide a firm theoretical framework to study these new deep networks. This is joint work with Vishwanath Saragadam, UC Riverside.

Tuesday, May 20, 2025 (10:15 am - 11:30 am)
Lifespan Changes in Human Brain Networks
Tingting Zhang, Professor, University of Pittsburgh

Understanding the evolution of human brain networks throughout the lifespan is a key scientific challenge. However, previous studies have reported inconsistent findings, largely due to the inherent complexity of brain network lifespan changes and the limited sample sizes available. To address these issues, we developed a new clustering-enabled regression approach that identifies clusters of brain regions undergoing synchronized age- related changes, thereby revealing diverse patterns of connectivity alterations across different regions.

To ensure reproducibility, we applied our method to a comprehensive dataset comprising functional magnetic resonance imaging (fMRI) and diffusion MRI (dMRI) data from three Human Connectome Project studies. Our analysis indicates that only a small subset of network connections exhibits practically significant age-related changes in either functional connectivity (FC) or structural connectivity (SC). Specifically, FC between clusters within the same functional network tends to decline with age, while FC between clusters from different networks displays a variety of change patterns. Additionally, our findings highlight sex-specific trends in FC alterations. Notably, our investigation into dMRI-derived structural networks reveals that lifespan changes in FC do not consistently mirror those observed in SC. These results underscore the complexity and heterogeneity of brain network evolution over the lifespan and provide new insights into the underlying mechanisms of brain aging.

Wednesday, May 21, 2025 (10:15 am - 11:30 am)
Multimodal, Generative, and Agentic AI for Pathology
Faisal Mahmood, Associate Professor of Pathology, Harvard School of Public Health

Advances in digital pathology and artificial intelligence have presented the potential to build models for objective diagnosis, prognosis and therapeutic-response and resistance prediction. In this talk we will discuss our work on: (1) Data-efficient methods for weakly-supervised whole slide classification with examples in cancer diagnosis and subtyping (Nature BME, 2021), identifying origins for cancers of unknown primary (Nature, 2021) and allograft rejection (Nature Medicine, 2022) (2) Discovering integrative histology-genomic prognostic markers via interpretable multimodal deep learning (Cancer Cell, 2022; IEEE TMI, 2020; ICCV, 2021; CVPR, 2024; ICML, 2024). (3) Building unimodal and multimodal foundation models for pathology, contrasting with language and genomics (Nature Medicine, 2024a, Nature Medicine 2024b, CVPR 2024). (4) Developing a universal multimodal generative co-pilot and chatbot for pathology (Nature, 2024). (5) 3D Computational Pathology (Cell, 2024) (6) Bias and fairness in computational pathology algorithms (Nature Medicine, 2024; Nature BME 2023) (7) Agentic AI workflows for diagnostic pathology and biomedical research.

TUTORIAL I

Instructor: Hsin-Hsiung Bill Huang (U Central Florida)
Title: Low-Rank Regression and Classification for Neuroimaging: Statistical Methods in R
Programming language: R
Description:  This tutorial introduces low-rank statistical models for neuroimaging data, focusing on MRI images and EEG brain signals for disease classification and prediction. Participants will learn matrix and tensor decomposition, and sparse low-rank models with hands-on implementation in R and Python. Applications include MRI and EEG analysis, and functional connectivity modeling. The tutorial covers data preprocessing, optimization techniques, and scalable computation. Participants will gain practical experience in low-rank neuroimaging analysis with provided datasets and code.

TUTORIAL II

Instructor: Zhengjia Wang (UPenn)
Title: RAVE: Reproducible Analysis and Visualization for Human Intracranial Electroencephalography (iEEG)
Programming language:  The tutorial will be given in R.
Description: Human intracranial electroencephalography (iEEG) is an emerging technique that opens a unique window into human brain function with unprecedented temporal resolution and spatial precision. This tutorial will showcase how iEEG data can benefit from rigorous statistical analyses and how the RAVE platform facilitates every step of the workflow from raw data to interactive visualizations. By the end of the tutorial, the attendees will be able to 

• Understand the benefit of iEEG for studying the brain and the data challenges
• Use the RAVE platform to process and analyze iEEG data through a case study
• Apply their own algorithms to iEEG data using RAVE packages

Monday, May 19, 2025  (1:15 pm – 2:45 pm)

Recent Advances in Statistical Methods for Spatially Resolved Omics Data

Organizer: Lulu Shang, MD Anderson Cancer Center
Speakers: Tao Wang, Ziyi Li, Xiyu Peng, Lulu Shang


Speaker 1: Tao Wang, UT Southwestern Medical Center
Presentation Title: Mapping cellular interactions from spatially resolved transcriptomics data
Abstract: Cell–cell communication (CCC) is essential to how life forms and functions. However, accurate, high-throughput mapping of how expression of all genes in one cell affects expression of all genes in another cell is made possible only recently through the introduction of spatially resolved transcriptomics (SRT) technologies, especially those that achieve single-cell resolution. Nevertheless, substantial challenges remain to analyze such highly complex data properly. Here, we introduce a multiple-instance learning framework, Spacia, to detect CCCs from data generated by SRTs, by uniquely exploiting their spatial modality. We highlight Spacia’s power to overcome fundamental limitations of popular analytical tools for inference of CCCs, including losing single-cell resolution, limited to ligand–receptor relationships and prior interaction databases, high false positive rates and, most importantly, the lack of consideration of the multiple-sender-to-one-receiver paradigm. We evaluated the fitness of Spacia for three commercialized single-cell resolution SRT technologies: MERSCOPE/Vizgen, CosMx/NanoString and Xenium/10x. Overall, Spacia represents a notable step in advancing quantitative theories of cellular communications. 

Speaker 2: Ziyi Li, MD Anderson Cancer Center
Presentation Title:  Accurate Imputation of Pathway-specific Gene Expression in Spatial Transcriptomics with PASTA
Abstract:  Mapping the entire transcriptome at single-cell resolution under its natural spatial context is essential for investigating the oncogenesis and progression of diseases. The recently emerged targeted in-situ technologies retain the spatial organization of cells at high resolution, although they remain limited in the number of genes that can be simultaneously measured. To overcome this obstacle, numerous computational methods have been developed to predict unmeasured gene expression in spatial transcriptomics data by leveraging scRNA-seq data. Most of these methods focus on the expression of individual genes and usually generate highly variable predictions. In this study, we introduce PASTA (PAthway-oriented Spatial gene impuTAtion), a novel spatial pathway expression imputation method that leverages cell type and spatial proximity to enhance prediction accuracy. PASTA integrates pathway information into the imputation process, which improves prediction robustness and enhances biological relevance in spatial transcriptomics data. Additionally, PASTA assumes that nearby cells and cells of the same type exhibit similar expression patterns. We demonstrate PASTA's superior performance across both simulated and real-world datasets, highlighting its ability to impute pathway gene expression with improved stability and biological significance.

Speaker 3:  Xiyu Peng, Texas A&M University
Presentation Title:  Decoding Spatial Tissue Architecture: A Scalable Bayesian Topic Model for Multiplexed Imaging Analysis
Abstract:  Recent progress in multiplexed tissue imaging is advancing the study of tumor microenvironments to enhance our understanding of treatment response and disease progression. Despite its popularity, there are significant challenges in data analysis, including high computational demands that limit feasibility for large-scale applications and the lack of a principled strategy for integrative analysis across images. To overcome these challenges, we introduce a spatial topic model designed to decode high-level spatial architecture across multiplexed tissue images. Our method integrates both cell type and spatial information within a topic modelling framework, originally developed for natural language processing and adapted for computer vision. We benchmarked its performance through various case studies using different single-cell spatial transcriptomic and proteomic imaging platforms across different tissue types. We show that our method runs significant faster on large-scale image datasets, along with high precision and interpretability. We also demonstrate it consistently identifies biologically and clinically significant spatial “topics”, such as tertiary lymphoid structures.

Speaker 4: Lulu Shang, MD Anderson Cancer Center
Presentation Title: Statistical identification of cell type-specific spatially variable genes in spatial transcriptomics
Abstract: An essential task in spatial transcriptomics is identifying spatially variable genes (SVGs). Here, we present Celina, a statistical method for systematically detecting cell type-specific SVGs (ct-SVGs)—a subset of SVGs exhibiting distinct spatial expression patterns within specific cell types. Celina utilizes a spatially varying coefficient model to accurately capture each gene’s spatial expression pattern in relation to the distribution of cell types across tissue locations, ensuring effective type I error control and high power. Celina proves powerful compared to existing methods in single-cell resolution spatial transcriptomics and stands as the only effective solution for spot-resolution spatial transcriptomics. Applied to five real datasets, Celina uncovers ct-SVGs associated with tumor progression and patient survival in lung cancer, identifies metagenes with unique spatial patterns linked to cell proliferation and immune response in kidney cancer, and detects genes preferentially expressed near amyloid-β plaques in an Alzheimer’s model.

Monday, May 19, 2025  (1:15 pm – 2:45 pm)

Advanced Methods in Neuroimaging and Brain Connectivity

Organizer:  Ruiwen Zhou, Washington University in St. Louis
Speakers:  Aris Sotiras, Yezhi Pan, Eardi Lila, Andrew An Chen

Speaker 1: Aris Sotiras, Washington University in St. Louis
Presentation Title:  Modeling Neurodegenerative Disorders as Anomalies Using Deep Learning
Abstract:  Neurodegenerative disorders, such as Alzheimer’s disease, pose significant challenges due to their heterogeneous presentation, complex progression, and the difficulty of early detection. Individual disease trajectories vary widely, influenced by genetic, environmental, and comorbid factors, making it difficult to develop generalized diagnostic models. Traditional approaches rely on established biomarkers and clinical assessments, yet these methods often fail to detect subtle brain changes before significant symptoms emerge. In this talk, I will explore how deep learning can model neurodegenerative disorders as anomalies, providing a data-driven approach to personalized diagnosis and prognosis. Specifically, I will discuss how unsupervised deep learning can learn normative brain aging patterns, which can then be used to identify deviations indicative of disease. This framework enables subject-specific abnormality mapping, offering interpretable insights into pathology. By framing disease-related changes as outliers, anomaly detection models can capture subtle yet informative patterns that distinguish pathological from healthy aging. Through experimental results, I will demonstrate how this approach enhances disease identification, progression prediction, and the characterization of disease heterogeneity. This talk will be relevant to researchers in machine learning, neuroimaging, and computational medicine. Attendees will gain insights into both the potential and limitations of deep learning for anomaly-based neurodegenerative disorder modeling, and how these methods can contribute to early diagnosis and personalized interventions.

Speaker 2:  Yezhi Pan, University of Maryland
Presentation Title:  Multilayer Network Model for Brain Structural-functional Connectivity Coupling
Abstract:  The network model for brain structural-functional connectivity coupling is challenging because both structural and functional brain connectivity (SC and FC) exhibit organized and complex network structures. We represent SC and FC as graphs, where nodes denote brain regions and edges represent connections. The node sets for SC and FC networks may differ. Our focus is on assessing which sets of SC edges are associated with sets of FC edges and understanding the underlying network topological structure. We propose a triple-layer network model to capture the complex associations and identify dense SC-FC subgraph pairs. In a dense SC-FC subgraph pair, we maximize the proportion of SC-FC related edges in both SC and FC subgraphs while ensuring that most related SC-FC edges are covered by SC-FC subgraph pairs. We conduct extensive simulations and apply the proposed method to data from the Human Connectome Project and UK Biobank.

Speaker 3: Eardi Lila, University of Washington 
Title: A physics-informed geometric deep learning approach to electrophysiological source reconstruction
Abstract: Electrophysiological brain signals are acquired through indirect and noisy measurements that represent transformed versions of the underlying neural activity. Source reconstruction --- the process of inferring the underlying signals from these measurements --- is essential for accurate brain function mapping but remains challenging. Deep learning methods have shown promise, but they often discard or only implicitly capture the physics governing neural signal propagation, leading to inefficient learning. To address this, we introduce a novel physics-informed geometric deep learning framework that embeds potentially ill-posed physics constraints into a deep learning model. This approach reduces sample size requirements and enhances generalizability across sensor locations and configurations without retraining. We apply the proposed method to magnetoencephalography source reconstruction.

Speaker 4:  Andrew An Chen, Medical University of South Carolina
Presentation Title:  Kernel-based association tests for brain cortical gradients
Abstract:  Recent methodological advances describe the topological organization of the brain cortex as a continuous map called brain cortical gradients. These gradients are consistent with seminal research on brain functional organization, neurodevelopmental trajectories, and key multimodal brain metrics. However, statistical methods for the analysis of brain cortical gradients are limited and current approaches either ignore population variability or key properties of gradient data. Brain cortical gradients and methods for deriving gradients are first introduced. Then, the unique properties of this novel data type and the limitations of existing approaches are discussed. Finally, nonparametric hypothesis testing methods appropriate for gradient data are proposed. Application to the Philadelphia Neurodevelopmental Cohort reveals that the proposed methods can improve power for detecting neurodevelopmental changes and sex differences. Potential extensions and statistical frameworks are explored for further methodological developments of our method. Additionally, the generalizability of our approach is validated with an external neuroimaging study, demonstrating its applicability across different clinical studies. This highlights the potential of our framework as a versatile tool for predicting longitudinal brain volumes.

Tuesday, May 20, 2025  (8:30 am – 10:00 am)

Advancing Statistical Methods for Imaging: Innovations in Modeling, Mediation, and Connectivity Analysis

Organizer:  Jian Kang, University of Michigan
Speakers:  Timothy D. Johnson, Shuo Chen, Yi Zhao, Yuliang Xu

Speaker 1:  Timothy D Johnson, University of Michigan
Presentation Title:  Statistical Modeling of fMRI Data for Pre-Surgical Planning
Abstract:  Spatial smoothing is an essential step in the analysis of functional magnetic resonance imaging (fMRI) data. The standard method is to convolve the image data (at each point in the time-series) with a three-dimensional Gaussian kernel that applies a fixed amount of smoothing to the entire image. In pre-surgical planning, using fMRI to determine functionally eloquent, peritumoral regions of the brain image, spatial accuracy is paramount. Thus methods that rely on global smoothing may not be reasonable as global smoothing can blur the boundaries between activated and non-activated regions of the brain. Moreover, in a standard fMRI analysis strict false positive control is desired. For pre-surgical planning false negatives may be of greater concern. In this talk, we present two Bayesian models that allow spatially adaptive smoothing that circumvent the problem with global smoothing. For both models, with start with the unsmoothed Z-statistic image obtained from standard software, such as SPM or FSL, and segment the image into three classes: deactivated, activated, and null classes. The first model is based on a Pott’s prior model for image segmentation that does not smooth over sharp boundaries in the Z-statistic image and a Dirichlet process prior. Our non parametric Pott’s model favors image configurations where neighboring voxels belong to the same class. We adopt a Bayesian decision theoretic approach to determine which voxels belong to each of the three classes. In this approach, false negatives and false positives are penalized asymmetrically in the loss function allowing false negatives to be penalized more heavily than false positives. Our second model is a spatially adaptive, conditionally autoregressive model. Similar to the Potts model, this model reduces smoothing at boundaries between regions of no activation and activation in the Z-statistic image. After we fit the model to the data, we again take a Bayesian decision theoretic approach that allows false negatives and false positives to be penalized differently. We apply both models to pre-surgical fMRI data from a patient with an Oligodendroglioma. During surgery, electrical stimulation mapping (EMS) was performed to determine functionally eloquent, peritumoral, regions of the brain responsible for speech. Post-surgery assessment of our modeling results with EMS show promise for these models and fMRI in pre-surgical planning.

Speaker 2:  Shuo Chen, University of Maryland
Presentation Title:  A mediation model for assessing the impact of aging on functional connectome networks via brain structural changes
Abstract:  Brain aging is characterized by consistent changes in brain structure and function, reflected in the decline of cognitive abilities such as learning, memory, and attention. However, the impact of age-related structural changes on brain function remains unclear. To address this gap, we develop a novel network mediation model to investigate how structural brain changes influence functional connectome (FC) networks with aging. Our approach systematically evaluates structure-FC coupling and identifies a minimal set of mediators shaping age-related FC networks. We apply this method to 40,126 participants from the UK Biobank for discovery and 512 participants from the Human Connectome Project-Aging cohort for validation. The findings reveal robust structural changes that influence FC networks associated with fine motor function and attention.

Speaker 3:  Yi Zhao, Indiana University 
Presentation Title:  Longitudinal regression of covariance matrix outcomes
Abstract: In this study, a longitudinal regression model for covariance matrix outcomes is introduced. The proposal considers a multilevel generalized linear model for regressing covariance matrices on (time-varying) predictors. This model simultaneously identifies covariate-associated components from covariance matrices, estimates regression coefficients, and captures the within-subject variation in the covariance matrices. Optimal estimators are proposed for both low-dimensional and high-dimensional cases by maximizing the (approximated) hierarchical-likelihood function. These estimators are proved to be asymptotically consistent, where the proposed covariance matrix estimator is the most efficient under the low-dimensional case and achieves the uniformly minimum quadratic loss among all linear combinations of the identity matrix and the sample covariance matrix under the high-dimensional case. Through extensive simulation studies, the proposed approach achieves good performance in identifying the covariate-related components and estimating the model parameters. Applying to a longitudinal resting-state functional magnetic resonance imaging data set from the Alzheimer's Disease (AD) Neuroimaging Initiative, the proposed approach identifies brain networks that demonstrate the difference between males and females at different disease stages. The findings are in line with existing knowledge of AD and the method improves the statistical power over the analysis of cross-sectional data.

Speaker 4:  Yuliang Xu, Duke University
Presentation Title:  Bayesian Structured Mediation Analysis with Unobserved Confounders
Abstract:  We explore methods to reduce the impact of unobserved confounders on the causal mediation analysis of high-dimensional mediators with spatially smooth structures, such as brain imaging data. The key approach is to incorporate the spatial latent individual effects, which influence the structured mediators, as unobserved confounders in the outcome model, thereby potentially debiasing the mediation effects. We develop Bayesian Structured Mediation analysis with Unobserved confounders (BASMU) framework, and establish its model identifiability conditions. Theoretical analysis is conducted on the asymptotic bias of the Natural Indirect Effect (NIE) and the Natural Direct Effect (NDE) when the unobserved confounders are omitted in mediation analysis. For BASMU, we propose a two-stage estimation algorithm to mitigate the impact of these unobserved confounders on estimating the mediation effect. Extensive simulations demonstrate that BASMU substantially reduces the bias in various scenarios. We apply BASMU to the analysis of fMRI data in the Adolescent Brain Cognitive Development study, focusing on four brain regions previously reported to exhibit meaningful mediation effects. Compared with the existing image mediation analysis method, BASMU identifies two to four times more voxels that have significant mediation effects, with the NIE increased by 41\%, and the NDE decreased by 26%.

Tuesday, May 20, 2025 (8:30 am – 10:00 am)

Complex Scientific Questions and Neurophysiological Data: Methodological Perspectives

Organizer: Donatello Telesca, University of California at Los Angeles
Speakers:  Damla Senturk, Donatello Telesca, Aaron Scheffler, Nicholas Marco

Speaker 1: Damla Senturk, UCLA Biostatistics
Presentation Title: Modeling intra-individual inter-trial EEG response variability in autism
Abstract: Autism spectrum disorder (autism) is a prevalent neurodevelopmental condition characterized by early emerging impairments in social behavior and communication. EEG represents a powerful and non-invasive tool for examining functional brain differences in autism. Recent EEG evidence suggests that greater intra-individual trial-total variability across EEG responses in stimulus-related tasks may characterize brain differences in autism. Traditional analysis of EEG data largely focuses on mean trends of the trial-averaged data, where trial-level analysis is rarely performed due to low neural signal to noise ratio. We propose to use nonlinear (shape-invariant) mixed effects (NLME) models to study intra-individual inter-trial EEG response variability using trial-level EEG data. By providing more precise metrics of response variability, this approach could enrich our understanding of neural disparities in autism and potentially aid the identification of objective markers. The proposed multilevel NLME models quantify variability in the signal’s interpretable and widely recognized features (e.g., latency and amplitude) while also regularizing estimation based on noisy trial-level data. Even though NLME models have been studied for more than three decades, existing methods cannot scale up to large data sets. We propose computationally feasible estimation and inference methods via the use of a novel minorization-maximization (MM) algorithm. Extensive simulations are conducted to show the efficacy of the proposed procedures. Applications to data from a large national consortium find that children with autism have larger intra-individual inter-trial variability in P1 latency in a visual evoked potential (VEP) task, compared to their neurotypical peers.

Speaker 2: Donatello Telesca, UCLA Biostatistics
Presentation Title: Modeling mixed membership and phase variability in EEG 
Abstract: A common concern in the field of functional data analysis is the challenge of temporal misalignment, which is typically addressed using curve registration methods. Currently, most of these methods assume the data is governed by a single common shape or a finite mixture of population level shapes. We introduce more flexibility using mixed membership models. Individual observations are assumed to partially belong to different pure mixtures, allowing for variation across multiple functional features. We propose a Bayesian hierarchical model to estimate the underlying shapes, as well as the individual time-transformation functions and levels of membership. Motivating this work is data from EEG signals in children with autism spectrum disorder (ASD). Our method agrees with the neuroimaging literature, recovering the 1/f pink noise feature distinctly from the peak in the alpha band. Furthermore, the introduction of a regression component in the estimation of time-transformation functions quantifies the effect of age and clinical designation on the location of the peak alpha frequency (PAF). 

Speaker 3: Aaron Scheffler, University of California San Francisco
Presentation Title: Constrained covariate-dependent smoothing and curve registration with applications to disease progression modeling
Abstract: Disease progression can be tracked via a cascade of changes in biomarkers and clinical measurements over the disease time course. For example, in progressive neurodegenerative diseases (ND), such as Alzheimer's Disease, changes in biomarkers (neuroanatomical images, cerebrospinal fluid) may precede clinical measurements (cognitive batteries) by months or years. Viewing repeated measurements of biomarkers and clinical measurements as a multivariate time series composed of continuous and discrete values, successful modeling of disease progression balances capturing stereotypic patterns in disease progression across subjects with subject-level variability in timing, acceleration, and shape of disease progression trajectories. A Bayesian model combining parametric and constrained semi-parametric disease progression models is proposed for curve alignment and covariate-dependent smoothing of exponential family outcomes across the disease time course, allowing for the characterization of typical disease progression as well as heterogeneity in the timing, speed, ordering, and shape of disease progression at the population-level and at the subject-level via random effects structure that partitions phase and amplitude variance. The framework will accommodate continuous and count outcomes, allowing for the incorporation of measurements ranging from neuroimaging features to sensitive sub-scales of cognitive batteries. A working example is provided from patients experiencing progressive ND.

Speaker 4: Nicholas Marco, Duke University
Presentation Title: Modeling Neural Switching via Drift-Diffusion Models
Abstract: A neural encoding theory known as multiplexing posits that when multiple stimuli are present, individual neurons can switch over time between encoding each member of the stimulus ensemble, causing a fluctuating pattern of firing rates. Here, we introduce a new statistical framework to test and analyze rate fluctuations of multiplexing neurons by proposing a state-space model for point process data. Our state-space model that differs from most of the existing statistical literature in that the state changes are (1) continuous-time, (2) non-Markovian, and (3) endogenous. Statistical inference for the proposed state-space model is challenging due to these differences, causing typical MCMC methods for state-space models to have poor sampling performance. To efficiently conduct posterior inference, we propose a novel MCMC algorithm for inference on state-space models with similar dependence structures. To statistically assess the necessity of our multiplexing-specific model to explain the observed data, we pitted it against a simpler and more general point process model that encapsulates alternative encoding theories with some level of abstraction. Using this statistical framework, we provide compelling evidence of multiplexing in the inferior colliculus of two macaque monkeys, with novel insight into the timescale of the switching process.

Tuesday, May 20, 2025  (1:00 pm – 2:30 pm)

How Can We Model Multivariate Time Series Data?

Organizer: Mark Fiecas, University of Minnesota
Speakers:  Zeda Li, Ellery Island, Heather Shappell, Scott Bruce

Speaker 1: Zeda Li, Baruch College
Presentation Title: CPSCA: Conditional Principal Spectral Component Analysis for Covariate-dependent Multivariate Time Series
Abstract: In this talk, we introduce a novel frequency-domain dimension reduction method for covariate-dependent high-dimensional time series, named conditional principal spectral component analysis (CPSCA).  This method decomposes the covariate-dependent multivariate time series into two parts: one with spectral densities dependent on the covariates and the other with spectral densities independent of the covariates.  To uncover the latent frequency-domain dependent structures, a new metric called the spectral martingale difference divergence matrix (specMDDM) is proposed.  The proposed method can serve as an initial step in the analysis of covariate-dependent high-dimensional time series, transforming a potentially high-dimensional problem into a lower-dimensional one.  Consistency results for the methods are established under both fixed dimensions and diverging dimensions.

Speaker 2: Ellery Island, University of Minnesota
Presentation Title:  Children with ADHD Have More Dynamic Brain Networks: A Novel Changepoint Detection Approach in fMRI Data
Abstract: Elucidating the differences in brain function between children with and without mental health or neurodevelopmental disorders is necessary for growing our understanding of these disorders and of the brain more fully. Measuring dynamic functional connectivity (dFC) provides a glimpse of brain function, but a relative lack of statistical methods are available for characterizing dFC. In this study, we pioneer a new changepoint detection method, PELT-LMEC, and apply it to data from the Adolescent Brain Cognitive Development (ABCD) Study (n = 4,424). We examine the associations between the number of changepoints in four networks and dimensional measures of psychopathology and neurodevelopmental disorder in children ages 9-11. PELT-LMEC uses a generalized linear model framework for covariance matrices and an efficient, accurate changepoint detection algorithm to locate changes in the structure of large covariance matrices. We find that the number of changepoints is a replicable statistic. Furthermore, our results suggest that different networks exhibit notably distinct distributions of the number of changepoints, and that ADHD is associated with greater dynamicity in FC.

Speaker 3: Heather Shappell, Wake Forest University
Presentation Title:  A Hidden Semi-Markov Model Approach to State-Based Dynamic Brain Network Analyses: Recent Developments and Future Directions
Abstract: The study of functional brain networks has grown tremendously over the past decade. Most functional connectivity (FC) analyses assume that FC networks are stationary across time.
However, there is interest in studying changes in FC over time. Hidden Markov models (HMMs) are a useful modeling approach for FC. However, a severe limitation is that HMMs assume the sojourn time (number of consecutive time points in a state) is geometrically distributed. This encourages state switches too often. I propose a hidden semi-Markov model (HSMM) approach for inferring functional brain networks from functional magnetic resonance imaging (fMRI) data, which explicitly models the sojourn distribution. Specifically, I propose using HSMMs to find each subject's most probable series of network states, the cumulative time in each state, and the networks associated with each state. This approach is demonstrated on fMRI data from a study on older adults with obesity. Lastly, I propose an extension to the HSMM, where the sojourn distribution may depend on a number of covariates. This extension allows for a direct comparison of sojourn times across patient populations.

Speaker 4: Scott Bruce, Texas A&M University
Presentation Title:  Frequency Band Analysis of Nonstationary Multivariate Time Series
Abstract: Information from frequency bands in biomedical time series provides useful summaries of the observed signal. Many existing methods consider summaries of the time series obtained over a few well-known, pre-defined frequency bands of interest. However, there is a dearth of data-driven methods for identifying frequency bands that optimally summarize frequency-domain information in the time series. A new method to identify partition points in the frequency space of a multivariate locally stationary time series is proposed. These partition points signify changes across frequencies in the time-varying behavior of the signal and provide frequency band summary measures that best preserve nonstationary dynamics of the observed series. An L_2-norm based discrepancy measure that finds differences in the time-varying spectral density matrix is constructed, and its asymptotic properties are derived. New nonparametric bootstrap tests are also provided to identify significant frequency partition points and to identify components and cross-components of the spectral matrix exhibiting changes over frequencies. Finite-sample performance of the proposed method is illustrated via simulations. The proposed method is used to develop optimal frequency band summary measures for characterizing time-varying behavior in resting-state electroencephalography (EEG) time series, as well as identifying components and cross-components associated with each frequency partition point.

Tuesday, May 20, 2025 (1:00 pm – 2:30 pm)

Advanced Statistical Methods in Neuroimaging and Functional Data Analysis

Organizer:  Xinyu Zhang, Vanderbilt University
Speakers: Raphiel J. Murden, Zihang Wang, Kaidi Kang, Tsung-Hung Yao

Speaker 1:  Raphiel J. Murden, Emory University
Presentation Title:  Functional Connectivity, PTSD, and Ambulatory Blood Pressure In Early-Middle Aged Black Women
Abstract:  Stressful experiences may contribute to an increased risk for cardiovascular and neurovascular disease via increased blood pressure and blood pressure variability, which are associated with deleterious changes in cardiovascular organs and the brain. For Black women, early middle age (30-50 years) has been identified as an age-range when BP and BPV increase much more sharply than for other race-sex groups. In this presentation, we use data integration via the joint and individual variance explained (JIVE) framework to examine common (i.e., joint) sources of variance shared among measures of functional connectivity (FC) and behavioral and emotional factors such as PTSD symptoms. Subject level summaries (i.e., scores) of variance sources are then examined for associations with ambulatory blood pressure (ABP) indices. Preliminary results from n = 77 participants revealed significant associations between unique variance sources unique to FC, especially within the executive control network, and nighttime dipping in systolic BP. Neither joint variance sources nor those unique to behavioral factors were associated with BP indices.

Speaker 2:  Zihang Wang, Emory University 
Presentation Title:  Robust Hierarchical Causal Inference for Treatment Effects on Brain Connectivity in Autism
Abstract:  Autism spectrum disorder (ASD) is associated with atypical functional connectivity, commonly measured using resting-state functional magnetic resonance imaging (rs-fMRI) data. However, head movements during scanning introduce artifacts that can bias functional connectivity estimates, particularly in studies involving children with ASD. We present a hierarchical statistical framework that integrates flexible machine learning techniques to estimate treatment effects on functional connectivity while accounting for motion artifacts. At the first level, we model motion effects on fMRI time series non-parametrically and estimate functional connectivity using residuals. At the second level, we develop augmented inverse probability weighted (AIPW) estimators for both the average treatment effect (ATE) and the natural direct effect (NDE) not mediated by motion on the functional connectivity estimates from the first level. Our framework ensures multiple robustness and asymptotic normality of the AIPW estimators, even when all nuisance models are estimated adaptively. Additionally, we employ a multiple testing procedure that controls the false discovery proportion using multiplier bootstrap techniques with influence functions. Simulation studies demonstrate the superior performance of our approach in estimation accuracy, robustness, and false discovery control. Applying our method to a study of school-age autism children, stimulants increase connectivity between posterior and anterior parts of the default mode network.

Speaker 3:  Kaidi Kang, Vanderbilt University
Presentation Title:  A Unified Effect Size Reporting Tool for Cross-sectional and Longitudinal Neuroimaging Studies
Abstract:  Reporting of effect sizes (ES), such as Cohen's d and odds ratios (OR), alongside their confidence intervals, has gained attention for its ability to convey both the strength and precision of scientific findings simultaneously. However, existing ES indices are model-specific, presenting a challenge for researchers attempting to compare effect sizes across studies addressing similar questions but using different statistical models. In this work, we introduce a robust ES index (RESI) that is not conditional on statistical models to facilitate ES reporting. Moreover, longitudinal and cross-sectional study designs have systematic differences in ESs, thereby complicating comparisons between the two. To resolve this, we propose a new version of RESI tailored for longitudinal studies, which estimates ES as if the study were conducted cross-sectionally, thereby improving comparability across different study designs. We applied our proposed framework to the Lifespan Brain Chart Consortium (LBCC) to demonstrate that our RESI unifies the ES reporting across studies with different designs (i.e., cross-sectional or longitudinal), bridging a critical gap in ES communication.

Speaker 4:  Tsung-Hung Yao, University of Texas MD Anderson Cancer Center
Presentation Title:  Flexible Bayesian Nonparametric Product Mixtures for Global-local Functional Clustering
Abstract:  There is a rich literature on clustering functional data with applications to time-series modeling, trajectory data, and even spatio-temporal applications. Existing methods are classified into two categories of (i) global clustering that assigns an identical membership over the whole domain, and (ii) local clustering that allows multiple memberships based on the sub-domains. However, global clustering struggles with degenerated clustering for high-dimensional functions. While there is some limited literature on local clustering approaches to deal with the above problems, these methods are not scalable to high-dimensional functions and ignore the spatial correlation, with their theoretical properties remaining underexplored. Focusing on basis expansions for high-dimensional functions, we propose a novel ``global-local'' framework that partitions coefficients into multiple subsets based on different resolution levels and imposes independent Dirichlet process (DP) priors on each subset. The proposed framework generalizes existing global and local clustering by introducing a non-trivial partition of coefficients and enhances computational efficiency by reducing the number of priors through resolution-wise partitioning, while preserving clustering flexibility and performance. Moreover, the proposed method incorporates spatially correlated errors to provide better model fitting and shows posterior consistency properties that asymptotically recover the true density of random functions. An efficient Markov chain Monte Carlo (MCMC) algorithm is developed for implementation. Extensive simulations illustrate the improved computation, clustering, and function estimation under the proposed method compared to classical approaches. We apply the proposed approach to a spatial transcriptomics application where the goal is to infer clusters of genes with distinct spatial patterns of expressions. Our method makes an important contribution by expanding the limited literature on existing clustering methods for high-dimensional functions with theoretical guarantees.

Tuesday, May 20, 2025 (2:45 pm – 4:15 pm)

Advanced statistical modeling of brain imaging data

Organizer:  Yi Zhao, Indiana University School of Medicine
Speakers:  Alex Petersen, Chao Huang, Amanda Mejia, Jian Kang

Speaker 1:  Alex Petersen, Brigham Young University
Presentation Title:  Modeling and Regularized Estimation of the Covariance Operator of Multivariate Functional Data
Abstract:  Functional MRI scans result in a set of regional BOLD signals for each subject, which can be modeled as multivariate functional data. The covariance operator of multivariate functional data is a complex object that can be difficult to estimate, especially if the multivariate dimension is large, making extensions of statistical methods for standard multivariate data to the functional data setting challenging. Compared with multivariate data, a key difficulty is that the covariance operator is compact and thus does not have a bounded inverse. This talk will address covariance modelling for multivariate functional data utilizing nested structures of separability, as well as regularized estimation of the functional precision operators.

Speaker 2:  Chao Huang, University of Georgia
Presentation Title:  Patch-wise Intensity Mapping for Individualized Brain Abnormality Detection in Alzheimer’s Disease
Abstract:  Detecting and localizing abnormal brain regions is crucial for early Alzheimer’s disease (AD) prevention and progression monitoring. While traditional statistical and machine learning approaches have advanced brain imaging analysis, they often struggle to capture individualized abnormalities due to spatial correlations and pathological heterogeneity in brain imaging data, i.e., (i) voxel-wise intensity abnormalities are influenced by their local neighborhoods and (ii) abnormal patterns vary across subjects in location, shape, size, and number. To address this challenge, we propose a patch-wise statistical learning framework that leverages local intensity distributional properties for whole-brain abnormality detection in AD. Our method constructs a scalable and robust brain-wide normal aging model using patch-wise Fréchet regression and Principal Component Analysis (PCA). With this brain chart, we perform patch-wise residual analysis to detect voxel-wise risk patterns linked to cognitive impairment. We validate our approach through simulation studies and real-world experiments on the Human Connectome Project (HCP) and Alzheimer's Disease Neuroimaging Initiative (ADNI) diffusion tensor image (DTI) datasets, demonstrating superior performance over existing statistical and machine learning methods. Our framework enhances the precision of whole-brain abnormality detection in AD and identifies signs of early mild cognitive impairment (EMCI), highlighting its potential for early diagnosis and neurodegenerative disease monitoring.

Speaker 3:  Amanda Mejia, Indiana University Bloomington
Presentation Title:  The hidden cost of stringent motion censoring in fMRI
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.

Speaker 4:  Jian Kang, University of Michigan
Presentation Title:  Network Latent Source Separation
Abstract:  In recent years, the study of brain connectomics, particularly multimodal connectomics, has attracted significant interest in the field of neuroscience research. Analyzing brain connectome and the interplay between multimodal networks is challenging due to various factors, such as noisy imaging data and different types of connectivity measures across modalities. In this paper, we propose a new method called Network Latent Source Separation (NLSS) that uses a Bayesian framework to decompose discrete representations of brain networks and enable joint analysis of multimodal connectome. NLSS decomposition naturally preserves the network topology as the latent sources identified by the model have the same structures as the observed brain networks, and their loadings represent the prominence of the latent sources in each individual. In this way, NLSS provides more interpretable results compared to existing methods. We develop a Gibbs sampling scheme based on subspace approximation to achieve highly efficient posterior computation. We apply NLSS to decompose functional and functional-structural joint network data obtained from the Philadelphia Neurodevelopmental Cohort (PNC) study. Our analysis of functional networks identifies reliable subnetwork systems within the estimated sources, and the analysis of joint networks provides additional insights into the interplay between functional connectivity (FC) and structural connectivity (SC), for example, strong positive FC exists between regions with or without SC while negative FC usually occurs due to global functional organization in the brain instead of direct structural pathways. Our simulation studies demonstrate the superior performance of NLSS for decomposing networks generated from various mechanisms.

Tuesday, May 20, 2025 (2:45 pm – 4:15 pm)

Recent Advancements in Spatial Methods for Geospatial Imaging in Environmental Applications

Organizer:  Mikyoung Jun (University of Houston)
Speakers:  Dorit Hammerling, Emily Hector, Samuel WK Wong, Pulong Ma

Speaker 1:  Dorit Hammerling (Colorado School of Mines)
Presentation Title:  Nonstationary Spatial Modeling of Massive Global Satellite Data
Abstract: Earth-observing satellite instruments obtain a massive number of observations every day. For example, tens of millions of sea surface temperature (SST) observations on a global scale are collected daily by the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument. Despite their size, such datasets are incomplete and noisy, necessitating spatial statistical inference to obtain complete, high-resolution fields with quantified uncertainties. Such inference is challenging due to the high computational cost, the nonstationary behavior of environmental processes on a global scale, and land barriers affecting the dependence of SST. We develop a multi-resolution approximation (M-RA) of a Gaussian process (GP) whose nonstationary, global covariance function is obtained using local fits. The M-RA model requires domain partitioning, which can be set up application-specifically. In the SST case, we partition the domain purposefully to account for and weaken dependence across land barriers. Our M-RA implementation is tailored to distributed-memory computation in high-performance-computing environments. We analyze a MODIS SST dataset consisting of more than 43 million observations, to our knowledge the largest dataset ever analyzed using a probabilistic GP model. We show that our nonstationary model based on local fits provides substantially improved predictive performance relative to a stationary approach.

Speaker 2:  Emily Hector (NCSU)
Presentation Title:  Distributed model building and recursive integration for big spatial data modeling
Abstract: Motivated by the need for computationally tractable spatial methods in neuroimaging studies, we develop a distributed and integrated framework for estimation and inference of Gaussian process model parameters with ultra-high-dimensional likelihoods. We propose a shift in viewpoint from whole to local data perspectives that is rooted in distributed model building and integrated estimation and inference. The framework’s backbone is a computationally and statistically efficient integration procedure that simultaneously incorporates dependence within and between spatial resolutions in a recursively partitioned spatial domain. Statistical and computational properties of our distributed approach are investigated theoretically and in simulations. The proposed approach is used to extract new insights into autism spectrum disorder from the autism brain imaging data exchange.

Speaker 3:  Samuel WK Wong (U Waterloo)
Presentation Title:  Spatio-temporal data fusion of satellite imagery and in situ sampling with application to harmful algae blooms
Abstract: A harmful algae bloom (HAB) created by toxic, microscopic algae can be detrimental to freshwater and marine ecosystems. Chlorophyll-a levels, which serve as a proxy for algal biomass, are measured by in situ monitoring programs or can be inferred from satellite imagery. While satellites provide large-scale data for studying the frequency and intensity of HABs, images are limited by cloud cover, pixel resolution, and potential bias. Thus, it is of interest to integrate the in situ and satellite data for more comprehensive analyses. We present a spatio-temporal data fusion model to tackle this problem and perform inference in a scalable way by leveraging the INLA-SPDE approach. As a case study, we consider HABs in the western basin of Lake Erie.

Speaker 4:   Pulong Ma (Iowa State Univ)
Presentation Title:  Uncertainty Quantification for OCO-2/3 Missions
Abstract: Satellite measurements provide wealthy information for understanding Earth systems. NASA’s Orbiting Carbon Observatory-2/3 (OCO-2/3) missions produce various data products through a complex data processing chain. A key procedure known as retrieval algorithm produces estimates of geophysical variables such as atmospheric CO2 and many other components of the atmosphere by solving an inverse problem.  One key task in the OCO-2/3 missions is to perform probabilistic assessment of retrievals through large-scale observing system uncertainty experiments (OSUE). The OSUE need to run a complex forward model that describes the mathematical relationship between radiances and atmospheric state. This talk will introduce an emulator - a fast probabilistic approximation to the forward model - to address the computational challenges in performing OSUE through an emulator. The proposed emulator can make fast prediction for functional output of radiance across three spectrum bands (O2-A band, weak CO2 band, strong CO2 band) based on high-dimensional input of the atmospheric state through a statistical model that has built in dimension reduction in input space and output space. Validation experiments demonstrate that this emulator outperforms other competing statistical methods and a reduced order model that approximates the full-physics forward model. Downstream applications such as data fusion involving estimates of geophysical variables from multiple satellite instruments will also be illustrated if time permitting.

Wednesday, May 21, 2025 (1:00 pm – 2:30 pm)

Causal, Dynamic, and Multiview Modeling in High-Dimensional Brain Data

Organizer and Chair: Michele Guindani, University of California – Los Angeles
Speakers:  Beniamino Hadj-Amar, Xi (Rossi) Luo, Jun Young Park, Abhra Sarkar

Speaker 1: Beniamino Hadj-Amar, Rice University
Presentation Title:  Discrete Autoregressive Switching Processes in Sparse Graphical Modeling of Multivariate Time Series Data.
Abstract: We propose a flexible Bayesian approach for sparse Gaussian graphical modeling of multivariate time series. We account for temporal correlation in the data by assuming that observations are characterized by an underlying and unobserved hidden discrete autoregressive process. We assume multivariate Gaussian emission distributions and capture spatial dependencies by modeling the state-specific precision matrices via graphical horseshoe priors. We characterize the mixing probabilities of the hidden process via a cumulative shrinkage prior that accommodates zero-inflated parameters for non-active components, and further incorporate a sparsity-inducing Dirichlet prior to estimate the effective number of states from the data. For posterior inference, we develop a sampling procedure that allows estimation of the number of discrete autoregressive lags and the number of states, and that cleverly avoids having to deal with the changing dimensions of the parameter space. We thoroughly investigate performance of our proposed methodology through several simulation studies. We further illustrate the use of our approach for the estimation of dynamic brain connectivity based on fMRI data collected on a subject performing a task-based experiment on latent learning.
 
 

Speaker 2: Xi (Rossi) Luo, UT Health Houston
Presentation Title: Causal Mediation Analysis for Multilevel and Functional Data
Abstract: Causal mediation analysis typically involves conditions that may not be applicable in neuroimaging studies. We introduce a multilevel causal mediation framework to overcome this limitation and more accurately quantify information flow in brain pathways. This framework is designed to tackle several challenges: unmeasured mediator-outcome confounding, multilevel time series analysis, and the estimation of functional causal effects. Our approach is grounded in multilevel structural equation modeling, complemented by relaxed likelihood estimation methods. Interestingly, certain causal estimates, typically unobtainable in simpler data structures, become identifiable in our more complex data setting. We provide proof of the asymptotic properties of our estimators and illustrate the numerical properties through empirical analysis. Additionally, we utilize real fMRI data to demonstrate the practical effectiveness of our proposed framework.

Speaker 3: Jun Young Park, University of Toronto
Presentation Title:  A robust approach to improve statistical power in high-dimensional multi-view association testing 
Abstract: Understanding the interplay between high-dimensional data from different views is essential in biomedical research, in fields like genomics and neuroimaging. Existing statistical association tests for two random vectors often do not fully capture dependencies between views due to limitations in modeling within-view dependencies, particularly in unstructured high-dimensional data without clear dependency patterns, leading to a potential loss of statistical power. In this work, we propose a novel approach, devariation, which is considered as a simple yet effective preprocessing method to address the limitations by adopting a penalized low-rank factor model to flexibly capture within-view dependencies. Theoretical asymptotic power analysis shows that devariation increases statistical power, especially when within-view correlations impact signal-to-noise ratios, while maintaining robustness in scenarios without strong internal correlations. Simulation studies highlight devariation's superior performance over existing methods in various scenarios. We further validated devariation in neuroimaging data from the UK Biobank study, examining the associations between imaging-driven phenotypes (IDPs) derived from functional, structural, and diffusion magnetic resonance imaging (MRI). 
 

Speaker 4: Abhra Sarkar, University of Texas, Austin
Presentation Title:  Bayesian Semiparametric Orthogonal Tensor Factorized Mixed Models for Longitudinal Neuroimaging Data
Abstract: We introduce a novel longitudinal mixed model for analyzing complex neuroimaging data, addressing challenges such as high dimensionality, structural complexities, and computational demands. Our approach integrates dimension-reduction techniques, including basis function representation and Tucker tensor decomposition, to model spatial and temporal variations, group differences, and individual heterogeneity while drastically reducing model dimensions. The model accommodates irregular observation times and multiplicative random effects whose marginalization yields a novel Tucker-decomposed covariance-tensor framework. To ensure scalability, we employ semi-orthogonal mode matrices implemented via a novel graph-Laplacian-based smoothness prior with low-rank approximation, leading to an efficient posterior sampling method. A cumulative shrinkage strategy promotes sparsity and enables semi-automated rank selection. We establish theoretical guarantees for posterior convergence and demonstrate the method’s effectiveness through simulations, showing significant improvements over existing techniques. Applying the method to Alzheimer’s Disease Neuroimaging Initiative (ADNI) data reveals novel insights into local brain changes associated with disease progression, highlighting the method’s practical utility for studying cognitive decline and neurodegenerative conditions.

Wednesday, May 21, 2025 (8:30 am – 10:00 am)

Deep-Learning, Trees and Tensor Modeling of Imaging Data

Organizer: Hengrui Luo, Rice University
Chair:  Abhra Sarkar, UT-Austin
Speakers:  Ashok Veeraraghavan, Hengrui Luo, Michele Guindani, Akira Horiguchi  

Speaker 1: Ashok Veeraraghavan, Rice University
Presentation Title:  Beyond Pretty Pictures: Application-driven Prediction Bounds for Sparse Reconstruction
Abstract: Modern deep-learning reconstruction algorithms generate impressively realistic reconstructions from sparse inputs but can often produce significant inaccuracies. This makes it challenging to provide statistically guaranteed claims about the true state of the object. We propose a framework for computing provably valid prediction bounds on claims derived from probabilistic black-box image reconstruction algorithms. The key insights behind our framework are to represent reconstructions with a derived metric of interest and to calibrate bounds on the ground truth metric with conformal prediction (CP) using a prior calibration dataset. These bounds convey interpretable feedback about the object's state, and can also be used to retrieve nearest-neighbor reconstructed scans for visual inspection. We demonstrate the utility of our framework on sparse-view computed tomography (CT) for fat mass quantification and radiotherapy planning tasks. Results show that our framework produces bounds with better semantical interpretation than conventional pixel-based bounding approaches, and flags dangerous outlier reconstructions that look plausible but have statistically unlikely metric values. Our work lays the foundation for more interpretable and trustworthy test-time assessments of sparse image reconstruction.

Speaker 2: Hengrui Luo, Rice University
Presentation Title:  Efficient Decision Trees for Tensor Regressions
Abstract: This talk covers recent progress in tree-based methods for tensor regressions, which is an interpretable nonparametric tool that assists in various learning tasks. In particular, we develop single and ensemble tree methods for tensor-input regressions. We begin with scalar-on-tensor (tensor input and scalar output) regression and design efficient computational strategies to handle tensor inputs, which is a more complex search space than vector input space. Then we extend our tensor tree model to tensor-on-tensor (tensor input and tensor output) regressions with ensemble approaches with theoretic guarantees. We will also identify some existing challenges in applying tree-based and nonparametric methods for ultra-high dimensional tensor data, like MRI/fMRI data with limited individuals and possible missing data. We'll wrap up with a ranking perspective and raise a couple of open questions when extending this ranking perspective to tensor-input scenarios. 

Speaker 3: Michele Guindani, UCLA
Presentation Title:  Bayesian Time-Varying Tensor Vector Autoregressive Models for Dynamic Effective Connectivity
Abstract: In contemporary neuroscience, a key area of interest is dynamic effective connectivity, which is crucial for understanding the dynamic interactions and causal relationships between different brain regions. Dynamic effective connectivity can provide insights into how brain network interactions are altered in neurological disorders such as dyslexia. Time-varying vector autoregressive (TV-VAR) models have been employed to draw inferences for this purpose. However, their significant computational requirements pose challenges, since the number of parameters to be estimated increases quadratically with the number of time series. In this talk, we propose a computationally efficient Bayesian time-varying VAR approach. For dealing with large-dimensional time series, the proposed framework employs a tensor decomposition for the VAR coefficient matrices at different lags. Dynamically varying connectivity patterns are captured by assuming that at any given time only a subset of components in the tensor decomposition is active. Latent binary time series select the active components at each time via an innovative and parsimonious Ising model in the time-domain. Furthermore, we propose sparsity-inducing priors to achieve global-local shrinkage of the VAR coefficients, determine automatically the rank of the tensor decomposition and guide the selection of the lags of the auto-regression. We show the performances of our model formulation via simulation studies and data from a real fMRI study involving a book reading experiment.  Joint work with Wei Zhang, Ivor Cribben, Sonia Petrone. 

Speaker 4: Akira Horiguchi , UC Davis
Presentation Title:  Multidimensional Scalable Wavelet Tree Ensembles
Abstract: We consider the problem of learning data representations that preserve existing geometric structures in an image. An existing approach is to incorporate adaptivity into discrete wavelet transforms. Such an approach provides both image denoising and hierarchical image segmentation. We also propose using this flexible wavelet basis to produce an ensemble of weak learners, since each weak learner can flexibly capture a high-signal region of a residual image. We evaluate the performance of our method on synthetic and real images and find that (1) it detects edges while avoiding splits in homogeneous regions, (2) it produces smaller errors than existing less flexible counterparts while having fast computation times, and (3) boosting produces effective representations that can be used for subsequent tasks.

Wednesday, May 21, 2025 (1:00 pm – 2:30 pm)

Medical Imaging I

Organizer: Meng Li, Rice University
Chair: Qiwei Li, UT Dallas
Speakers:  Maan Malahfji, Meng Li, David Ouyang, Cheng-Han Yu

Speaker 1: Maan Malahfji, Houston Methodist
Presentation Title: Applications of AI and Machine Learning in Cardiovascular MRI and CT
Abstract: We discuss recent advancements in AI tools and ML applications that are revolutionizing the acquisition, interpretation, and derivation of actionable findings of cardiovascular MRI and cardiac CT studies. Both cardiac CT and cardiac MRI are imaging modalities that are experiencing exponential growth in their use and applications. But cost, training requirements, and availability are key limitations for their increased utilization worldwide. Artificial intelligence based tools can democratize the availability of these advanced cardiac imaging modalities, to achieve high level imaging, using standard hardware. Machine learning statistical methods are making their way into analysis of big data in cardiology. We discuss machine learning based identification of clinically sound phenotypes in aortic valvular disease datasets, and how findings derived from ML algorithms correspond to clinical observations.

Speaker 2: Meng Li, Rice Statistics
Presentation Title:  Functional Modeling and Machine Learning for Automated Analysis of Cardiovascular Imaging Data
Abstract: We present a fully automated pipeline for assessing right ventricular (RV) function from echocardiographic tissue Doppler imaging (TDI), integrating tools from functional data analysis and machine learning. The pipeline begins with a training-free algorithm that digitizes Doppler signals from still-frame images and estimates clinical indices of RV contraction—TAPSE and S’—via dynamic peak detection and curve integration. These estimates closely align with clinician measurements, providing a reproducible alternative to manual assessment. To predict preserved RV function, we compare two approaches: traditional tabular machine learning models based on summary features, and an attention-based neural network that incorporates full Doppler waveform trajectories as functional inputs. The latter substantially improves classification performance, with external validation confirming its ability to stratify clinical outcomes. This work illustrates how automated estimation, functional data representations, and modern learning algorithms can enhance reliability and predictive power in biomedical signal analysis. I will also discuss related challenges and recent progress in cardiovascular imaging analysis.

Speaker 3: David Ouyang, Cedars-Sinai Medical Center
Presentation Title:  Deployment and Development of Cardiovascular AI
Abstract: In cardiology, we use a wide range of imaging and diagnostic tests to help diagnose cardiovascular disease, however the interpretation of such testing depends on expert interpretation. There is both imprecision in how common metrics are assessed (such as left ventricular ejection fraction in assessing heart function) as well as additional information that can be derived from the images that clinicians cannot reliably diagnose. In this presentation, we describe the history of our work in developing artificial intelligence (AI) for echocardiography, to studying its impact in blinded, randomized non-inferiority clinical trial (ClinicalTrials.gov NCT05140642) to finally FDA clearance. Similarly, we describe cardiac amyloidosis as an important disease that is underdiagnosed in routine clinical care, and how AI applied to echocardiography can help clinicians in early and accurate diagnosis.

Speaker 4: Cheng-Han Yu, Marquette University
Presentation Title: Comparison of Bayesian Convolutional Neural Network and Machine Learning Based on Lung X-ray Image Classification
Abstract: Deep neural networks have become integral to medical image analysis, yet standard approaches often struggle with overfitting, poor calibration, and inadequate uncertainty quantification—critical shortcomings in clinical decision-making contexts. To address these limitations, we explore Bayesian deep learning methodologies under a convolutional neural network architecture, emphasizing their advantages for uncertainty-aware image classification. We implement and compare a suite of Bayesian inference techniques, including stochastic gradient MCMC, variational inference, and Gaussian approximations, with classical stochastic gradient descent training and traditional machine learning models. Using a dataset of human lung X-ray images, including cases of COVID-19 and other pulmonary conditions, we assess model performance in both binary and multi-class classification tasks, with special attention to out-of-distribution generalization. Our findings show that Bayesian model averaging across multiple high-performing modes significantly improves predictive accuracy and calibration. Notably, dropout-based and subnetwork-based inference outperform fully connected networks, suggesting enhanced regularization and uncertainty modeling. Moreover, in simple binary classification tasks, Gaussian processes and ensemble methods such as gradient boosting and random forests remain competitive. These results underscore the value of principled Bayesian modeling in advancing robust and interpretable medical imaging diagnostics.

Wednesday, May 21, 2025 (8:30 am – 10:00 am)

Next-Generation Statistical Advances in Neuroimaging: Bayesian Combinatorial Analysis, Object Data Regression, and Amortized Inference

Organizer:  Rajarshi Guhaniyogi, Texas A&M University
Chair: Beniamino Hadj-Amar, Rice University
Speakers:  Leo Duan, Ranjan Maitra, John Kornak, Rajarshi Guhaniyogi

Speaker 1:  Leo Duan, University of Florida
Presentation Title:  Statistical Modeling of Combinatorial Response Data --- An Inspiration from Albert and Chib (1993)
Abstract:  Albert and Chib (1993) is a foundational paper in categorical data analysis, and has inspired a vast literature on modeling binary and polychotomous responses. With the rich development, existing methods are inadequate for handling combinatorial responses, where each response is an array of integers subject to additional constraints. Such data are increasingly common in modern image applications, such as signal propagation on the human brain network in neuroimaging. Ignoring the combinatorial structure in the response data may lead to biased estimation and prediction. The fundamental challenge for modeling these integer-vector data is the lack of link function that connects a linear or functional predictor with a probability respecting the combinatorial constraints. In this paper, we propose a novel augmented likelihood, in which a combinatorial response can be viewed as a deterministic transform of a continuous latent variable. We specify the transform as the maximizer of integer linear programming, and characterize useful properties such as dual thresholding representation. When taking a Bayesian approach and considering a multivariate normal distribution for the latent variable, our method becomes a direct generalization to the celebrated probit data augmentation, and enjoys straightforward computation via Gibbs sampler. We provide theoretic justification for the proposed method at an interesting intersection between duality and probability distribution, and develop useful sufficient conditions that guarantee the applicability of our method. We demonstrate the effectiveness of our method on a task-based EEG image data analysis for human working memory study. This is a joint work with Yu Zheng and Malay Ghosh.

Speaker 2:  Ranjan Maitra, Iowa State University 
Presentation Title:  Tensor-on-Tensor Times Series Regression for Integrated One-step Analysis of fMRI Data
Abstract:  Data acquisition in a functional Magnetic Resonance Imaging (fMRI) activation detection experiment yields a massively structured array- or tensor-variate dataset that need to be analyzed with respect to a set of time-varying stimuli and possibly other covariates. The conventional approach employs a two-stage analysis: The first stage fits an univariate regression on the time series data at each individual voxel and reduces the voxel-wise data to a single statistic. The statistical parametric map formed from these voxel-wise test statistics is then fed into a second-stage analysis that potentially incorporates spatial context between the voxels and identifies activation within them. We develop holistic yet practical tensor-variate methodology that provides one-stage tensor-variate regression modeling of the entire time series array-variate dataset. Low-rank specifications on the tensor-variate regression parameters and Kronecker separable error covariance tensors make our innovation feasible.  A block relaxation algorithm provides maximum likelihood estimates of the model parameters. A R package, with C backends for computational feasibility, operationalizes our methods. Performance on different real-data-imitating simulation studies and a functional MRI study about Major Depressive Disorder demonstrate the stability of our approach and that it can reliably identify cerebral regions that are significantly activated.

Speaker 3:  John Kornak, University of California San Francisco
Presentation Title:  Bayesian Image Analysis in Fourier and Other Transformed Spaces
Abstract:  Bayesian image analysis provides a principled framework for improving image quality by integrating prior knowledge with probabilistic models of noise and uncertainty. Traditional approaches operate in image space, where modeling spatial dependencies is computationally demanding. We introduce Bayesian Image Analysis in Transformed Space (BITS), which leverages Fourier and wavelet transforms to reformulate priors and likelihoods, facilitating efficient and scalable inference. The Bayesian Image Analysis in Fourier Space (BIFS) framework transforms spatially correlated priors into independent processes across frequency components, decomposing high-dimensional estimation into parallelizable one-dimensional tasks. Analytical priors for Fourier coefficients capture known image properties such as smoothness and edge preservation. However, to model complex real-world images, we extend to a data-driven method (DD-BIFS), while a data-driven extension (DD-BIFS) empirically learns priors, enhancing adaptability for complex medical imaging applications. Bayesian Image Analysis in Wavelet Space (BIWS) extends these principles to a multiscale representation, effectively capturing non-stationary image features such as textures and anatomical structures. By leveraging transformed domains, both BIFS and BIWS achieve computational efficiency, scalability, and suitability for modern machine learning architectures. We will demonstrate the application of BITS to perfusion arterial spin labeling (ASL) imaging and functional MRI (fMRI), showcasing its potential for improving image quality and inference in neuroimaging.

Speaker 4:  Rajarshi Guhaniyogi, Texas A&M University
Presentation Title:  Explainable Artificial Intelligence with Spatial and Network Images
Abstract:  In medical imaging studies, a key goal is to identify associations among diverse imaging modalities with varying structural patterns. This work is motivated by a scientific application that aims to predict task-based brain activation maps (t-fMRI) using spatially-varying cortical metrics (s-MRI) and brain connectivity networks (rs-fMRI). To address this, we will discuss a generative model that integrates spatially-varying imaging inputs and network-valued predictors. The network effect is captured via an unknown non-linear function, while spatially-varying inputs are modeled using corresponding spatial regression coefficients, both estimated jointly via a novel explainable deep neural network (DNN) architecture. The proposed approach captures spatial smoothness of the imaging data, models complex association between network and spatial images, enables interpretability in the DNN architecture by offering well calibrated uncertainty in the inference and offers scalability with large sample and high-resolution images. 

Wednesday, May 21, 2025 (2:45 pm – 4:00 pm)

Methodological advances in diffusion imaging and shape analysis

Organizer and Chair:  Andrew Chen, Medical University of South Carolina
Speakers:  Will Consagra, Sebastian Kurtek, Hunter Moss

Speaker 1:  Will Consagra, University of South Carolina
Presentation Title:  Deep Learning Methods for Connectivity and Microstructure Inference in Diffusion MRI
Abstract:  Diffusion MRI (dMRI) is the primary imaging modality used to study brain connectivity and microstructure in vivo. However, parameter inference in common dMRI biophysical models is a challenging inverse problem, due to factors such as high dimensional parameter spaces, low signal-to-noise ratios, and non-linear forward models. These challenges have led many existing estimation methods to impose biologically implausible simplifications to stabilize estimation, leading to significant model misspecification and reduced interpretability. In this work, we introduce a novel sequential method for parameter inference for a broad class of dMRI models. Rather than estimating all parameters simultaneously, our approach decomposes the problem into a series of manageable subproblems, each solved using deep neural networks designed to exploit problem specific structure and symmetry. The resulting inference procedure is largely amortized, enabling scalable parameter estimation and uncertainty quantification across the whole brain. Simulation studies and real imaging data analysis using the Human Connectome Project (HCP) demonstrate the advantages of our method over standard alternatives.

Speaker 2:  Sebastian Kurtek, Ohio State University
Presentation Title:  Assessment of Spatial Dependence in Shapes of Planar Curves
Abstract:  We develop a marked point process framework for spatially indexed shapes of planar, parameterized curves. Shape is a property of an object that is invariant to translation, rescaling, rotation and reparameterization. To remove these sources of nuisance variation from the representation space, we build on the elastic shape analysis framework. Importantly, we use a parameterization-invariant metric to define a shape variogram that is subsequently used in the construction of a shape mark-weighted K function that treats the spatial locations at which shape data was observed as random. The K function enables (i) assessment of spatial dependence among shapes, and (ii) a hypothesis test of whether shape marks are independent in space. We evaluate the proposed framework using simulation studies, and apply it to reveal distinct spatial dependence structures among tumorous and non-tumorous cell nucleus shapes extracted from histopathology images. This is joint work with my current Ph.D. student Yejin Choi, and Karthik Bharath at the University of Nottingham.

Speaker 3:  Hunter Moss, Medical University of South Carolina
Presentation Title:  High b-value diffusion MRI and the fiber orientation density function
Abstract:  Diffusion MRI (dMRI) uniquely probes tissue microstructure using the random movement of water molecules. Many dMRI techniques, like diffusion tensor imaging (DTI) and diffusional kurtosis imaging (DKI), use weak diffusion weighting (i.e., low b-values), where the dMRI signal includes both intra- and extracellular water contributions. This complicates the interpretation of changes in rotational invariants, such as fractional anisotropy and mean kurtosis, as numerous processes are involved. In contrast, high b-value (b ³ 4000 s/mm²) dMRI isolates the intra-axonal signal in white matter (WM), making the data axon-specific. Fiber ball imaging (FBI) is a dMRI method that uses high b-value data to estimate the fiber orientation density function (fODF) in WM through the linear inverse Funk transform. The fODF reflects the angular distribution of axon orientations and allows for the estimation of axon-specific parameters like fractional anisotropy axonal (FAA), which enhances the biophysical interpretation of low b-value rotational invariants. In this talk, the concept of high b-value dMRI will be introduced, and its utility in clinical studies will be demonstrated.

Wednesday, May 21, 2025 (2:45 pm – 4:00 pm)

Medical Imaging II

Organizer: Qiwei Li, UT Dallas
Chair: Meng Li, Rice University
Speakers: Arvind Rao, Qiwei Li, Jia Wu

Speaker 1: Arvind Rao, University of Michigan
Presentation Title:  Statistical & Machine Learning Approaches To The Interpretation Of Structured Biomedical Data for Personalized Medicine
Abstract: This presentation examines advanced methodologies for analyzing structured biomedical data to enable personalized medicine. It focuses on three key biological data structures: spatial imaging from histopathology, mIF and single cell integration, as well as transcriptomic data from various analyses. The talk demonstrates how functional data analysis, graph neural networks, and tensor methods leverage inherent geometric structures in complex datasets. Case studies include GaWRDenMap for quantifying spatial heterogeneity in pancreatic disease, Cell-Graph Attention Network for disease grading, and tensor decomposition approaches for identifying spatial neighborhoods in tumor microenvironments. These methods enable deeper characterization of disease phenotypes through quantitative measurement of cellular interactions and tumor heterogeneity, bridging multi-omic data and clinical insights to advance personalized therapeutic strategies.

Speaker 2: Qiwei Li (UT Dallas)
Presentation Title:  AI-powered Bayesian Methods for Analyzing Histopathology Images
Abstract:  Statistics traditionally emphasizes human-driven analysis supported by computational tools, whereas AI primarily depends on computer algorithms with guidance from human insight. Nonetheless, each milestone in statistical development opens new frontiers for AI and offers fresh perspectives within statistics itself. This interplay fosters discoveries initiated from either domain that ultimately enrich the other. In this talk, I will illustrate how the integration of statistical spatial and shape analysis and AI enables more interpretable and predictive pathways from histopathology images to clinically meaningful insights. Recent advances in deep learning have made it possible to detect and classify tissue regions and individual cells at scale from digital histopathology images. I will introduce several novel AI-powered Bayesian models for analyzing these images. These methods offer new insights into cell-cell interactions, spatial cellular architecture, and tumor boundaries in the context of cancer progression, supported by multiple case studies.

Speaker 3:   Jia Wu (MD Anderson)
Presentation Title: Multi-modal AI Modeling in Lung Cancer
Abstract:  This talk explores the role of multi-modal AI modeling in patient selection and stratification to optimize cancer treatment strategies. It highlights a clinical-radiomics model for identifying patients receiving Stereotactic Ablative Radiotherapy (SABR) who may benefit from the addition of immunotherapy. Findings from a multi-center clinicogenomic study will also be presented, demonstrating how patient stratification can guide decisions between immune checkpoint inhibitor (IO) monotherapy and combination therapy with chemotherapy (IO+chemo). The discussion extends to emerging AI-driven tools for radiographic analysis, including deep learning models, habitat imaging for tumor characterization, and synthetic PET modeling. When integrated with blood-based biomarkers, these imaging techniques enhance recurrence prediction and support personalized treatment planning. The talk also covers advances in AI-powered analysis of digital pathology and spatial biology datasets, emphasizing how integration across data modalities can close clinical gaps and improve outcomes. Together, these approaches demonstrate the transformative potential of AI in advancing precision oncology through more individualized and effective treatment strategies.