Statistical Methods in Imaging (SMI) 2026

In the analysis of brain functional MRI (fMRI) data using regression models, Bayesian methods are highly valued for their flexibility and ability to quantify uncertainty. However, these methods face computational challenges in high-dimensional settings typical of brain imaging, and the often pre-specified correlation structures may not accurately capture the true spatial relationships within the brain. To address these issues, we develop a general prior specifically designed for regression models with large-scale imaging data. We introduce the Soft-Thresholded Conditional AutoRegressive (ST-CAR) prior, which reduces instability to pre-fixed correlation structures and provides inclusion probabilities to account for the uncertainty in choosing active voxels in the brain. We apply the ST-CAR prior to scalar-on-image (SonI) and image-on-scalar (IonS) regression models—both critical in brain imaging studies—and develop efficient computational algorithms using variational inference (VI) and stochastic subsampling techniques. Simulation studies demonstrate that the ST-CAR prior outperforms existing methods in identifying active brain regions with complex correlation patterns, while our VI algorithms offer superior computational performance. We further validate our approach by applying the ST-CAR to working memory fMRI data from the Adolescent Brain Cognitive Development (ABCD) study, highlighting its effectiveness in practical brain imaging applications.

To understand Alzheimer's disease progression, we aim to model abnormal accumulation of Tau protein via connected brain networks. We propose a novel Bayesian latent space model that formulates interactions between brain network edges and regional Tau protein in a co-evolution framework. We propose multiplicative coevolution model to neuroimaging and reproduce observed network–pathology dependencies and quantifies posterior uncertainty. We apply this model to the Alzheimer's Disease Neuroimaging Initiative (ADNI). We found that the baseline Tau volume of female is higher than male in subcortical regions, suggesting a different tau accumulation pattern with sex.

Some evidence suggests that people with autism spectrum disorder exhibit patterns of brain functional dysconnectivity relative to their typically developing peers, but specific findings are difficult to replicate. To facilitate this replication goal with data from the Autism Brain Imaging Data Exchange (ABIDE), we propose a flexible and interpretable model for participant-specific voxel-level brain functional connectivity. Our approach efficiently handles massive participant-specific whole brain voxel-level connectivity data that exceed one trillion data points. The key component of the model is to leverage the block structure induced by defined regions of interest to introduce parsimony in the high-dimensional connectivity matrix through a block covariance structure. Associations between brain functional connectivity and participant characteristics—including eye status during the resting scan, sex, age, and their interactions—are estimated within a Bayesian framework. A spike-and-slab prior facilitates hypothesis testing to identify voxels associated with autism diagnosis. Simulation studies are conducted to evaluate the empirical performance of the proposed model and estimation framework. In ABIDE, the method replicates key findings from the literature and suggests new associations for investigation.

Historically, neuroimaging results and figures only include regions that surpass a given statistical threshold. However, this common step strongly biases interpretations and meta-analyses, particularly towards non-reproducibility. We advocate instead for "transparent thresholding," which highlights statistically significant regions but also includes subthreshold locations for key experimental context. This balances distilling modeled results with enabling informed interpretations. We highlight transparent thresholding's many benefits with several examples and real-world datasets. We also bring attention to the many software packages that have now implemented it. Compared to standard results reporting, transparent thresholding removes ambiguity, decreases hypersensitivity to non-physiological features, catches potential artifacts, improves cross-study comparisons, reduces non-reproducibility biases, and clarifies interpretations.

Abstract to be announced

Abstract to be announced

Abstract to be announced

Abstract to be announced

Multi-center medical image analysis is limited by privacy, heterogeneous data, and high communication costs. We propose a communication-efficient decentralized learning framework where each medical site performs local updates and only exchanges model parameters. The method maintains privacy, greatly reduces communication, and achieves accuracy comparable to centralized training on multi-institution classification tasks.

With variety of imaging and EHR data available in clinical settings, integrative approaches towards management and assessment of surgical procedures for complex complications such as stroke have become a necessity in clinical settings. This talk will discuss a hierarchical AI approach to create such as an integrative solution for stroke management.

Improving outcomes in cancer surgery increasingly depends on intelligent tools that can guide real-time decision-making. While many computer-assisted interventions rely on conventional imaging technologies, mass spectrometry—a technique that captures detailed molecular signatures of tissues—is emerging as a powerful alternative modality. When integrated with AI-driven tissue classification models, mass spectrometry can help surgeons accurately distinguish between healthy and cancerous tissues. However, to enable practical intraoperative use, predictions must be localized and communicated effectively to surgeons to support decision-making. In this study, we evaluate the performance of navigated rapid evaporative ionization mass spectrometry (REIMS) as a tool for intraoperative margin assessment in breast cancer surgery.

Accurate intraoperative distinction between breast tumour and healthy tissue could improve surgical precision and reduce re-operations for patients requiring breast conserving surgery. While mass spectrometry data can be used to classify breast cancer tissue, its lack of explainability limits clinical implementation. Linking predictions to biological pathways allows clinicians to rapidly understand the classification logic. By mapping mass spectrometry data to biological pathways using publicly available databases, we created concept-based models that attribute specific pathways to the prediction of breast cancer or healthy tissue. With further refinement and validation, this could act as a decision-support tool, combining clinician expertise with rapid, explainable tissue analysis to optimize patient treatment.

Abstract to be announced

Abstract to be announced

Abstract to be announced

Abstract to be announced

Brain networks are typically represented by adjacency matrices, where each node corresponds to a brain region. In traditional brain network analysis, nodes are assumed to be matched across individuals, but the methods used for node matching often overlook the underlying connectivity information. This oversight can result in inaccurate node alignment, leading to inflated edge variability and reduced statistical power in downstream connectivity analyses. To overcome this challenge, we propose a novel framework for registering high resolution continuous connectivity (ConCon), defined as a continuous function on a product manifold space—specifically, the cortical surface—capturing structural connectivity between all pairs of cortical points. Leveraging ConCon, we formulate an optimal diffeomorphism problem to align both connectivity profiles and cortical surfaces simultaneously. We introduce an efficient algorithm to solve this problem and validate our approach using data from the Human Connectome Project (HCP). Results demonstrate that our method substantially improves the accuracy and robustness of connectome-based analyses compared to existing techniques.

Recent studies have shown that Alzheimer's disease imaging biomarkers remained highly variable even after state-of-the-art harmonization methods application. Furthermore, data harmonization and standardization across different clinical domains including Positron Emission Tomography (PET) imaging and blood biomarkers continue to hinder our ability to accurately estimate differences across different clinical groups. In this study we present different harmonization methods for PET imaging and blood biomarkers outcomes in Alzheimer's Disease and aim to explain the variability across several harmonization methods.

Reference intervals, defined as intervals containing a new observation with a specified probability relative to reference data, would be clinically useful in assessing brain magnetic resonance imaging (MRI). Brain charts, which are estimates of MRI phenotypes across covariates such as age and sex, can be used to construct reference intervals. However, the reference data used to fit intervals often differs from a new sample in terms of study design, MRI acquisition, and image preprocessing. Application of MRI reference intervals to new samples remains a challenging problem. Here, we propose a new method called Reference interval calibration via conFormal prediction (ReForm) that adjusts reference intervals for a new sample. Our method builds on recent work in conformal prediction, which yields intervals with guaranteed coverage for new observations. Through resampling experiments in Lifespan Brain Chart Consortium cortical thickness data, we compare ReFormed reference intervals to refitting intervals, statistical harmonization methods, and model-based adjustment of intervals. Notably for patient privacy concerns, ReForm does not require sharing of reference data. Yet, our empirical results demonstrate that ReForm controls FPR similarly or better than alternative methods which require sharing reference data. Finally, we provide recommendations for practical applications of ReForm and an R package for calibrating reference intervals using ReForm.

The Robust Effect Size Index (RESI) is a recently proposed standardized index to quantify effect magnitude across models, with significant applications in high-dimensional data analysis, including neuroimaging studies. However, its confidence interval (CI) construction has relied on computationally intensive bootstrap procedures. We establish a general theorem for the asymptotic distribution of the RESI using a Taylor expansion that accommodates a broad class of models. Simulations under various linear and logistic settings show that RESI and its CI have smaller bias and more reliable coverage than commonly used effect sizes such as Cohen's d and f, and that combining with robust covariance estimation yields valid inference under model misspecification—a crucial aspect in neuroimaging, where model assumptions often need to account for complex spatial dependencies and heterogeneous variances across regions. Additionally, our approach shows that the asymptotic method significantly reduces computational time, achieving up to a 50-fold speedup over the bootstrap procedure. This is particularly valuable in neuroimaging applications where large-scale datasets and computational efficiency are key challenges.

To understand Alzheimer's disease progression, we aim to model abnormal accumulation of Tau protein via connected brain networks. We propose a novel Bayesian latent space model that formulates interactions between brain network edges and regional Tau protein in a co-evolution framework. We propose multiplicative coevolution model to neuroimaging and reproduce observed network–pathology dependencies and quantifies posterior uncertainty. We apply this model to the Alzheimer's Disease Neuroimaging Initiative (ADNI). We found that the baseline Tau volume of female is higher than male in subcortical regions, suggesting a different tau accumulation pattern with sex.

Understanding functional connectivity patterns in Autism and ADHD remains an unsettled scientific problem with existing literature pointing to evidence of both over- and under-connectivity. Resting-state fMRI is a popular modality with rich spatiotemporal information but poses significant modeling challenges. This work explores new methods for identifying neuroimaging biomarkers in Autism and ADHD using high-dimensional resting-state fMRI data. The proposed statistical models are flexible in that it can accommodate multiple features in fMRI time series data such as stationary, nonstationary, high-dimensional, Gaussian and non-Gaussian. Statistical inference is carried out by constructing an approximate Whittle likelihood that stems from a frequency domain factor model. The modeling approach also enables estimation of frequency-specific functional connectivity matrices. Two group comparisons are carried out by finding differences in the distributions and means of the number of edges in the functional connectivity matrices. These two group comparisons via testing are achieved using multiple candidate discrete and zero-inflated discrete distributions. Using our proposed approach led to interesting results that indicate altered functional connectivity in Autism and ADHD with varying patterns witnessed across different age groups, resting-state networks and frequencies.

The study of functional brain networks increasingly emphasizes how connectivity patterns evolve over time. While hidden Markov models (HMMs) have become a popular tool for modeling dynamic functional connectivity (FC), their assumption of geometrically distributed sojourn times (consecutive time spent in a state) forces unrealistically frequent state switching. Hidden semi-Markov models (HSMMs) address this limitation by explicitly modeling the sojourn distribution, enabling more accurate characterization of the duration spent in each network state. I develop an HSMM framework for inferring dynamic brain network states from functional MRI data and estimating subject-specific state sequences, dwell times, and associated connectivity patterns. Building on this foundation, I introduce an extension in which the sojourn distribution is modeled as a function of covariates. This covariate-dependent formulation allows state dwell times to vary systematically with demographic, clinical, or behavioral factors, providing a direct mechanism for comparing temporal dynamics across patient populations. I illustrate the utility of this approach using fMRI data from older adults with obesity, demonstrating how covariate effects on sojourn behavior can reveal meaningful differences in network stability and flexibility. This framework opens new avenues for studying how individual characteristics shape the temporal organization of functional brain networks.

Craniosynostosis, the premature fusion of cranial sutures, alters skull morphology and can restrict localized brain development despite normal overall intracranial volume. These localized deformations are linked to long-term neurodevelopmental deficits, but there is a current lack of quantitative models that capture the heterogeneity in growth patterns driven by different suture fusion types. Leveraging a large dataset of CT and 3D photogrammetry scans from craniosynostosis patients and healthy controls, a novel hierarchical generalized additive model that uses soap-film smooths is proposed, and basis functions are shared to model continuous volume maps across space, age, and fusion type. This approach enables fine-scale localization of volume abnormalities and prediction of future development trajectories. The model accounts for covariates such as age and sex and is estimated using a Bayesian framework.

Abstract to be announced

Abstract to be announced

Abstract to be announced

Abstract to be announced

In biomedical research, repeated measurements within each subject are often processed to remove artifacts and unwanted sources of variation. The resulting data are used to construct derived outcomes that act as proxies for outcomes that are not directly observable. Although intra-subject processing is widely used, its impact on inter-subject statistical inference has not been systematically studied, and a principled framework for causal analysis in this setting is lacking. In this article, we propose a semiparametric framework for causal inference with derived outcomes obtained after intra-subject processing. This framework applies to settings with a modular structure, where intra-subject analyses are conducted independently across subjects and are followed by inter-subject analyses based on parameters from the intra-subject stage. We develop multiply robust estimators of causal parameters under rate conditions on both intra-subject and inter-subject models, which allows the use of flexible machine learning. We specialize the framework to a mediation setting and focus on the natural direct effect. For high dimensional inference, we employ a step-down procedure that controls the exceedance rate of the false discovery proportion. Simulation studies demonstrate the superior performance of the proposed approach. We apply our method to estimate the impact of stimulant medication on brain connectivity in children with autism spectrum disorder.

Functional magnetic resonance imaging (fMRI) has transformed our understanding of brain organization; however, conventional analysis pipelines often rely on anatomically fixed regions of interest that fail to adequately capture the dynamic and individualized nature of brain function. Although data-driven approaches estimate subject-specific intrinsic connectivity networks and improve individual-level representation, they typically continue to assume static spatial patterns. In this lecture, I present recent advances in Dynamic Precision Functional Mapping, a framework that moves beyond conventional approaches by capturing individualized, time-resolved functional sources through mathematically grounded, data-driven modeling. Furthermore, I will highlight recent efforts to incorporate nonlinear representations and interactions to enhance the fidelity of individualized mapping. Finally, I will demonstrate the utility of these methods in studying psychosis.

Alzheimer's disease (AD) is marked by cognitive decline along with the widespread of tau aggregates across the brain cortex. Due to the challenges of imaging pathology spreading flows in-vivo, however, quantitative analysis on the cortical pathways of tau propagation and its interaction with the cascade of amyloid-beta (Aβ) plaques lags behind the experimental insights of underlying pathophysiological mechanisms. To address this challenge, we present a physics-informed neural network, empowered by mean-field theory, to uncover the biologically meaningful spreading pathways of tau aggregates between two longitudinal snapshots. Following the notion of `prion-like' mechanism in AD, we first formulate the dynamics of tau propagation as a mean-field game (MFG), where the spread of tau aggregate at each location (aka. agent) depends on the collective behavior of the surrounding agents as well as the potential field formed by amyloid burden. Given the governing equation of propagation dynamics, MFG reaches an equilibrium that allows us to model the evolution of tau aggregates as an optimal transport with the lowest cost in Wasserstein space. By leveraging the variational primal-dual structure in MFG, we propose a Wasserstein-1 Lagrangian generative adversarial network (GAN), in which a Lipschitz critic seeks the appropriate transport cost at the population level and a generator parameterizes the flow fields of optimal transport across individuals. Additionally, a symbolic regression module derives an explicit Aβ-tau crosstalk formula. Experimental results on public neuroimaging datasets demonstrate that our explainable deep model yields accurate tau predictions for unseen subjects and offers new insights into AD pathology spread.

Functional magnetic resonance imaging (fMRI) provides an indirect measurement of neuronal activity via hemodynamic responses that vary across brain regions and individuals. Ignoring this hemodynamic variability can bias downstream connectivity estimates. Furthermore, the hemodynamic parameters themselves may serve as important imaging biomarkers. Estimating spatially varying hemodynamics from resting-state fMRI (rsfMRI) is therefore an important but challenging blind inverse problem, since both the latent neural activity and the hemodynamic coupling are unknown. In this work, we propose a methodology for inferring hemodynamic coupling on the cortical surface from rsfMRI. Our approach avoids the highly unstable joint recovery of neural activity and hemodynamics by marginalizing out the latent neural signal and basing inference on the resulting marginal likelihood. To enable scalable, high-resolution estimation, we employ a deep neural network combined with conditional normalizing flows to accurately approximate this intractable marginal likelihood, while enforcing spatial coherence through priors defined on the cortical surface that admit sparse representations. Uncertainty in the hemodynamic estimates is quantified via a double-bootstrap procedure. The proposed approach is extensively validated using synthetic data and real fMRI datasets, demonstrating clear improvements over current methods for hemodynamic estimation and downstream connectivity analysis.

Causal mediation analysis is critical for understanding how changes in the brain mediate the effects of environmental and genetic factors on neurological outcomes in neuroimaging studies. However, traditional mediation methods often face challenges when mediators are high-dimensional structured objects, such as complex brain imaging data, due to the curse of dimensionality and reduced statistical power. This study introduces a methodology that leverages envelope-based structured dimension reduction to enhance pathway identification and statistical power in detecting indirect effects. The proposed approach is applied to Alzheimer's Disease Neuroimaging Initiative (ADNI) data to examine how structural changes in brain regions mediate the impact of genetic factors on cognitive decline. Simulation studies validate the asymptotic properties of the estimators and demonstrate that the method outperforms existing techniques in improved power and reduced variance in estimation. This approach advances mediation analysis for neuroimaging and extends to other high-dimensional multivariate contexts, offering a robust framework for disease detection and intervention strategies.

In neuroscience, high dimensionality and complex dependence structures often lead to substantial variability across variable selection methods, even when predictive performance is comparable. We propose a knockoff framework with a novel generation strategy that extends the conventional sequential conditional independent pairs (SCIP) algorithm through conditional diffusion modeling and feature partitioning, enabling computationally efficient construction while preserving complex dependence patterns and spatial contiguity in high-dimensional data. Our approach is general and flexible, accommodating diverse data types and supporting both unimodal and multimodal settings, while rigorously controlling the false discovery rate (FDR) and producing robust inference. We apply the method to investigate behavioral outcomes using vectorized structural MRI features, matrix-valued functional connectivity from fMRI, and their joint representations. Across feature importance measures derived from multiple scalar-on-image models, our framework yields stable and consistent variable selection results, demonstrating enhanced robustness and interpretability for imaging-based inference.

Multi-site neuroimaging studies enhance statistical power and generalizability but face challenges from data privacy constraints where direct data sharing is prohibited. In addition, neuroimaging data across multiple institutions often exhibit systematic heterogeneity driven by scanner differences, which could bias downstream analyses if not addressed. We propose distributed covariance graph-guided ComBat (dG-ComBat), a multivariate harmonization framework that operates in a fully privacy-preserving distributed setting. dG-ComBat identifies and leverages the latent covariance graph structure among neuroimaging features to guide mitigation of both mean and covariance batch effects. With only two rounds of summary-statistics exchange, dG-ComBat achieves identifical harmonization performance (i.e., lossless) to centralized setting in which data from all participating institutions are shared. Extensive simulations and applications to amyloid-β and tau PET data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) showed that dG-ComBat effectively reduces scanner-related variability while preserving biologically relevant diagnostic signals, and improving downstream analyses. The method provides a scalable, summary-statistics-based approach for harmonizing multi-institutional neuroimaging data when individual-level data sharing is infeasible.

Structural connectivity (SC) derived from diffusion MRI and functional connectivity (FC) derived from functional MRI provide complementary summaries of brain organization. Because coherent brain function is supported by anatomically grounded, polysynaptic communication pathways, we model SC and FC as arising from a shared, biologically meaningful scaffold and propose a bidirectional graph autoencoder to estimate a latent backbone common to both modalities. The method fits coupled forward and reverse mappings (SC→FC and FC→SC) that share a single latent adjacency matrix, which serves as the propagation operator for message passing in both modalities and yields an explicit backbone estimate. To account for heterogeneity across subjects and modalities, we augment the shared backbone with modality specific deviations that adapt subject-level topology while remaining anchored to the common foundation. We characterize the estimator via a walk-algebra representation and show that the regularized bidirectional objective recovers a minimal shared backbone under identifiability conditions. In simulations, the proposed approach improves bidirectional reconstruction and more accurately recovers the ground-truth backbone than competing unidirectional and higher-order graph neural baselines. An application to Human Connectome Project (HCP) and Adolescent Brain Cognitive Development (ABCD) data yields reproducible backbone networks that align with established large-scale functional systems. Moreover, backbone estimates show moderate cross-cohort consistency between HCP and ABCD, illustrating the utility of the proposed framework for interpretable structure–function modeling.

Variability in repeated measurement longitudinal processes, such as fMRI, is subject to intrinsic (domain space) latent phase states stochasticity and extrinsic (range space) chronological time dispersion typically modeled as additive noise. I'll discuss complex-time (kime) representation of time-varying processes as parametric manifolds and kime-phase tomography (KPT), a mathematical statistics framework that extends the representation of longitudinal observations from classical time to kime. The kime phase tracks the intrinsic kime-phase domain stochasticity, due to cross-sectional variation in the repeated sampling. The phase is complementary to the classical extrinsic additive noise in the signal range space. The probabilistic formulation of the kime manifold induces different metrics and a family of non-commutative operators that satisfy canonical commutation relations analogous to quantum mechanics. This operator-theoretic foundation enables tomographic reconstruction of latent phase distributions from mixed-moment statistics via spectral deconvolution on the circle. We will show some examples related to time-varying medical images.

The rapid growth of high-resolution imaging across multiple subjects and imaging modalities presents remarkable opportunities to advance our understanding of complex biological and neurological processes. However, jointly analyzing these massive, multi-modal imaging datasets on a large number of subjects poses major challenges due to their extreme dimensionality, intricate spatial relationships, and heterogeneous sources. In this talk, I will introduce DeepSpFactor, a novel, interpretable deep learning framework for large-scale, multi-subject, multi-modal imaging integration. DeepSpFactor leverages spatial factor models, powered by deep neural networks (DNNs), to flexibly capture nonlinear and multi-level spatial patterns in imaging data. Our approach dissects image variation into meaningful latent components, models how these patterns relate to subject-specific traits, and robustly accounts for complex spatial dependencies. By combining Bayesian inference with scalable DNN-based computational architecture, DeepSpFactor provides reliable uncertainty quantification, supports the imputation of missing modalities, and delivers scientifically interpretable factors that distinguish shared from modality-specific structures, and offers rapid computation. This work opens new possibilities for statistically rigorous and computationally efficient analysis of large imaging studies, and I will discuss both methodological foundations and applications to real-world neuroscience data.

The majority of neuroimaging inference focuses on hypothesis testing rather than effect estimation. With concerns about replicability, there is growing interest in reporting standardized effect sizes from neuroimaging group-level analyses. Confidence sets for effect sizes were recently developed for neuroimaging, but are restricted to simple univariate contrasts (e.g., one-sample or two-sample test Cohen's d) and cross-sectional data. Thus, existing methods exclude increasingly common longitudinal associations of biological brain measurements with potentially nonlinear, inter- and intra-individual variations in diagnosis, development, or symptoms. We use modern methods for confidence sets combined with a recently proposed robust effect size index to provide a very general approach and software for effect size confidence set inference in neuroimaging. These confidence sets identify regions of the image where the lower or upper simultaneous confidence interval is above or below a given threshold with high probability. We evaluate the coverage and interval width of the proposed procedures using realistic simulations and perform longitudinal analyses of aging and diagnostic differences of cortical thickness in Alzheimer's disease. This comprehensive approach, along with the visualization functions integrated into the pbj R package, offers a robust tool for analyzing repeated neuroimaging measurements.

This talk presents a flexible, predictor-dependent joint modeling framework designed for analyzing multimodal neuroimaging data in aging. The framework accommodates network data from multiple subjects over a shared set of spatial nodes and incorporates spatially correlated attributes. It permits simultaneous inference on nodes associated with predictors, spatial relationships of nodal attributes, and predictor-attribute regression relationships. Simulations demonstrate its superior performance by accounting for both network and spatial data correlations. The method is used to analyze data from the Lifespan Cognitive and Motor Neuroimaging Laboratory, integrating structural and functional MRI data to analyze brain connectivity and region-specific attributes. It includes aging-related features and ROI spatial locations to enhance understanding of the interactions between brain structure, function, aging, and other predictors. As a Bayesian approach, it provides uncertainty quantification, delivering robust results even with small samples.

Abstract to be announced

Abstract to be announced

Tensor data, often characterized as multidimensional arrays, have become increasingly prevalent in biomedical studies, particularly in neuroimaging applications. Analyzing these complex datasets can be challenging due to the high dimensionality and inherent structures within tensors. In this work, we propose the Sparse Partial Generalized Tensor Regression (SPGTR) method for modeling diverse types of responses, including continuous, binary, and count data, in terms of both tensor and vector/scalar predictors. Our novel mode-wise penalized manifold optimization techniques enable us to achieve dimension reduction and biologically meaningful sparsity in tensor regression coefficient estimation. We establish the asymptotic properties of the proposed estimation facilitated by the envelope concept. We demonstrate the effectiveness of the SPGTR through extensive simulation studies. An application of SPGTR to a functional magnetic resonance imaging (fMRI) study yields new insights into the association between posttraumatic stress disorder (PTSD) and brain connectivity measures, which are not uncovered by existing methods.

Details to be announced

Abstract to be announced

We propose BPSiR-GSS: Bayesian Partition-based Scalar-on-Image Regression with Group Spike-and-Slab priors, a novel framework for regression models with image predictors designed to adaptively identify spatially contiguous regions associated with clinical outcomes. Unlike existing approaches that rely on an external fixed partition or a chosen atlas, our approach recursively partitions the spatial domain itself, producing an interpretable representation of region-specific effects. We employ a discrete group spike-and-slab prior on the partition nodes, which enables multi-level sparsity and region-level voxel selection. Theoretical considerations align the model with emerging posterior contraction results for discrete spike-and-slab priors in spatial settings. For posterior inference, we develop a trans-dimensional MCMC algorithm that employs a marginalization over the regression coefficients, resulting in an efficient updating scheme over partitions and sparsity patterns. Extensive simulation studies demonstrate that BPSiR-GSS achieves superior prediction accuracy and variable selection performance compared to state-of-the-art spatial Gaussian processes and low-rank tensor regression models. We apply the method to structural MRI data to predict brain age, successfully identifying localized patterns of neural degeneration that provide clean, interpretable maps of aging-related regions of interest. This is a joint work with Marina Vannucci at Rice University and Suprateek Kundu at M.D. Anderson Cancer Center.

Details to be announced

Details to be announced

Mediation analysis is widely used to elucidate the mechanisms through which an exposure influences an outcome of interest. However, standard causal mediation methods may yield inconsistent conclusions, potentially due to unmodeled heterogeneity in mediation effects across individuals. To address this challenge, we propose a new framework that incorporates covariate-treatment and mediator-treatment interactions within a linear structural equation modeling system. Causal assumptions are discussed and heterogeneous natural direct and indirect effects are parameterized as functions of patient characteristics. A modified covariate approach is introduced to relax hierarchical constraints in models with interactions, and generalized lasso regularization is employed to achieve parsimony in high-dimensional settings. We establish asymptotic properties of the proposed estimators and demonstrate their finite-sample performance through extensive simulation studies. Application to data from the Alzheimer's Disease Neuroimaging Initiative reveals substantial heterogeneity in the mediation effect of cerebrospinal fluid volume change on the association between APOE-ε4 carrier status and cognitive decline, identifying a subgroup of individuals who may be particularly susceptible to cognitive impairment through this disease pathway.

Converging evidence indicates a complex interaction between Tau-protein propagation and alterations in brain connectivity. Aimed at characterizing Tau protein spread along functional networks in the early course of Alzheimer's disease, we develop a Bayesian graphical model to jointly model tau propagation, functional connectivity structure, and subgroup heterogeneity using cross-sectional data. By integrating graph-constrained infection dynamics with connectivity patterns, the model infers plausible propagation pathways and subgroup-specific infection sequences. We show interesting results on heterogeneous Tau accumulation patterns by applying the method to the A4 study, with the findings further extended to the ADNI study.

The availability of large-scale neuroimaging datasets and powerful computational resources has created an unprecedented opportunity to explore the structure, function, and dynamics of the brain as what it truly is—a complex system. However, conventional statistical models have inherent limitations that restrict the ability of clinicians and scientists to fully leverage the wealth of information contained in the neuroimaging data. To address these challenges, we have developed novel multivariate modeling frameworks and accompanying software tools, built on mixed-effects regression backbones, that enable relating phenotypic characteristics (e.g., age, disease phenotype, etc.) to whole-brain networks and specific subnetworks such as the default mode network (DMN) and drawing principled statistical inference. These modeling tools overcome critical limitations of existing methods and have successfully been applied across a range of studies examining the effects of various phenotypes including aging, weight loss, alcohol drinking, pesticide exposure, HIV and Marijuana use, and cerebral microbleeds on brain networks.

The integration of statistical network analysis with neuroimaging has catalyzed a paradigm shift in our understanding of brain function. Inferring these directed interactions, however, is particularly challenging when functional connections are not static, but nonstationary. This talk bridges dynamic network analysis for neuroimaging with temporal segmentation. We introduce a flexible and adaptive random featurization framework to discern abrupt and gradual structural changes in the connectivity network, avoiding the traditional computational burden of deep learning methods. Post-segmentation, network inference is simplified, allowing for recovery of directed information flow within each quasi-stable regime.

Brain connectivity—the complex network of neural pathways enabling communication among brain regions—provides a powerful framework for understanding individual differences in cognition, emotion, and behavior. This project introduces ConnectMVR, a penalized multivariate matrix-variate regression framework designed for supervised brain connectivity analysis and multimodal neuroimaging integration. ConnectMVR identifies statistically reliable brain-behavior associations across multiple connectivity matrices at both the regional and connection levels, while accounting for the high dimensionality and structured dependencies inherent in connectome data. Using data from the Human Connectome Project in Development (HCP-D), we will apply ConnectMVR to integrate intrinsic functional connectivity from resting-state fMRI and structural connectivity from diffusion MRI, linking these measures to behavioral assessments. This framework offers a rigorous, interpretable, and generalizable approach for uncovering multiscale connectivity patterns underlying behavioral variability in youth.

Aging clocks aim to estimate an individual's underlying physiological and molecular age, distinct from chronological age. Because biological age cannot be measured directly, aging clocks are computed indices derived from high-throughput multi-omics and imaging data using techniques such as machine learning (ML) and artificial intelligence (AI). Aging clocks derived from standard ML/AI regression models are known to exhibit systematic prediction bias that shrinks predicted values toward the mean, leading to underestimation in older participants and overestimation in younger participants. Here, we elaborate on the potential causes of this bias inherited from ML/AI regression models and emphasize that it can compromise downstream analyses, for example, by inducing biased association estimates between accelerated aging clocks and risk factors. We further present ML regression strategies that automatically remove systematic bias during the model training stage and ensure unbiased inference in downstream analyses.

Modeling individual variability in brain functions is essential for understanding neurological disorders such as post-stroke aphasia. We propose a unified topological framework to characterize this heterogeneity and its impact on both behavioral deficits and treatment response. Our approach begins with a novel method for lesion-symptom mapping (LSM) that leverages persistent homology to capture higher-order topological features in functional brain networks. By moving beyond region-wise or edge-level comparisons, this method identifies mesoscale structures—such as cycles—that reflect distributed lesion effects. Statistical inference is performed via heat kernel representation of persistence diagrams and a permutation-based testing procedure. We also extend this framework to investigate treatment response by introducing a topological clustering method based on heat-kernel representation of persistence diagrams, enabling integration of Euclidean covariates and automated selection of the number of clusters. Together, these methods provide a cohesive topological perspective on brain-behavior relationships, offering multiscale, data-driven tools for analyzing functional brain networks in neuroimaging studies.

Alzheimer's disease (AD) is a progressive neurodegenerative disorder characterized by gradual structural brain changes that precede clinical symptoms. The Alzheimer's Disease Neuroimaging Initiative (ADNI) provides longitudinal structural MRI (sMRI) data—cortical thickness, surface area, and volumetric measures across brain regions—that serve as important biomarkers for tracking early AD progression. However, predicting time-to-AD conversion from these data is challenging due to inter-subject variability, complex spatial dependencies among brain regions, and temporal heterogeneity across visits. Traditional and deep learning classifiers often overlook censoring and fail to jointly model spatiotemporal patterns of neurodegeneration. We propose a temporal graph neural network (GNN) framework for time-to-Alzheimer's prediction that integrates anatomical structure with longitudinal trajectories. Our approach combines landmark analysis to handle censoring with a GNN that encodes structural relationships among brain regions, along with LSTM/Transformer-based temporal modules that capture both local dynamics and long-range progression. These hybrid architectures jointly model spatial organization and evolving sMRI patterns, producing individualized and clinically meaningful time-to-event risk predictions. The proposed framework offers a flexible and powerful tool for early detection of AD and may improve the identification of high-risk individuals in longitudinal neuroimaging studies.

The Gaia satellite has provided first-ever positions and motions for billions of stars over the last decade. However, observing constraints mean that Gaia's measurements get worse for faint stars and truncate past a certain brightness. These faint stars are particularly interesting for many open astrophysical questions, such as the nature of dark matter and the precise formation history of the Milky Way. Using my BP3M tool (Bayesian Positions, Parallaxes, and Proper Motions), I am able to combine Hubble images with Gaia data to measure order-of-magnitude improved stellar motions, especially for the faintest stars. In this talk, I will explain BP3M's robust statistical techniques, highlight ongoing applications with different astronomical questions, and present work expanding to James Webb data. I will also present in-progress applications of modern computer vision techniques, such as Neural Radiance Fields (NeRFs), to improve calibration of the complex distortions present in telescope images. Finally, I will discuss the future of BP3M-like methods and ML-enhanced calibration for the next generation of telescopes and surveys (e.g., Euclid, Roman, Rubin/LSST).

Details to be announced

Images in astronomy can be represented as a combination of several components: point sources (e.g., stars), extended compact sources (e.g., distant galaxies), and diffuse emission. Separating images into these individual components is required for precision measurements of the colors of stars/galaxies and to unveil the structure of gas/dust in the interstellar medium. In this talk, I will describe recent progress in Bayesian component separation techniques for astronomical imaging, using data-driven priors on each component. I will use as a case study a recent imaging program from JWST that reveals the turbulent substructure of dust in the interstellar medium using thermal light echoes. From the "clean" diffuse component of the images, we can learn about the hydrodynamic properties of the interstellar medium. By leveraging a time series for tomography, we use the diffuse components and Gaussian process methods to achieve the highest Cartesian resolution 3D dust map to date. I will conclude by highlighting the scalability of this class of component separation methods and their application to some of the largest surveys in astronomy, including SPHEREx.

Details to be announced

Longitudinal biomedical studies monitor individuals over time to capture dynamics in brain development, disease progression, and treatment effects. However, estimating trajectories of brain biomarkers is challenging due to biological variability, inconsistencies in measurement protocols (e.g., differences in MRI scanners) as well as scarcity and irregularity in longitudinal measurements. Herein, we introduce a novel personalized deep kernel regression framework for forecasting brain biomarkers, with application to regional volumetric measurements. Our approach integrates two key components: a population model that captures brain trajectories from a large and diverse cohort, and a subject-specific model that captures individual trajectories. To optimally combine these, we propose Adaptive Shrinkage Estimation, which effectively balances population and subject-specific models. We assess our model's performance through predictive accuracy metrics, uncertainty quantification, and validation against external clinical studies. Benchmarking against state-of-the-art statistical and machine learning models—including linear mixed effects models, generalized additive models, and deep learning methods—demonstrates the superior predictive performance of our approach. Additionally, we apply our method to predict trajectories of composite neuroimaging biomarkers, which highlights the versatility of our approach in modeling the progression of longitudinal neuroimaging biomarkers. Furthermore, validation on three external neuroimaging studies confirms the robustness of our method across different clinical contexts.

Diffusion models have recently emerged as powerful generative priors for solving inverse problems. However, training diffusion models is data-intensive and computationally demanding, which restricts their applicability for high-dimensional and high-resolution data such as medical imaging. In this talk, I will introduce our recent works on how to improve the efficiency (data, time, and memory efficiency) of diffusion-based generative models for solving general inverse problems through constrained posterior sampling. Particularly, I will introduce two promising solutions we propose to enable learning diffusion priors for solving high-dimensional imaging problems through latent diffusion and patch diffusion models. The results are demonstrated in solving various inverse problems for both natural and medical images including 3D medical image reconstruction, showing the effectiveness of our proposed methods in both model efficiency and model performance. If time permits, I will also introduce our recent works to enhance the generalization and controllability of diffusion sampling. Our recent findings in the initial noise space reveal new opportunities for uncertainty quantification, strengthening weak prior, and enabling fine-grained control in image restoration and reconstruction tasks. This research opens the door to leverage diffusion-based generative models in tackling complex real-world data for addressing crucial inverse problems in many biomedical and scientific applications.

Function recovery from irregular Fourier samples is a ubiquitous problem in a variety of fields. When measurements in k-space are not given on a lattice, however, a variety of artifacts and numerical issues can become problematic sources of signal pollution. In this work, I will discuss a recent advance in the design of weighting schemes to significantly mitigate these problematic artifacts that borrows from the theory of numerical quadrature, which can be adapted to a wide variety of sampling settings and geometries while retaining attractive algorithmic runtime costs. Particularly in the case of densely sampled and highly non-gridded data, this approach provides significant improvements over current popular weighting methods such as density compensation factors. After a discussion of the method and a few theoretical observations, demonstrations in exploratory analysis of spatio-temporal data and statistical inverse problems will be discussed.

In functional magnetic resonance imaging (fMRI), it is important to observe the functioning brain as fast as possible and at as high of a spatial resolution as possible. Increased spatial and temporal speed results in voxels with increased noise relative to signal and contrast. There is much evidence to suggest that there is important biological information contained within the phase component of the fMRI signal which is directly linked to local changes in magnetization. When the signal-to-noise ratio within a voxel is low, as when there is ultra-high resolution, the marginal statistical distribution of the phase is non-standard and difficult to work with. This non-standard marginal phase distribution at high signal-to-noise ratios is Normally distributed, but at low signal-to-noise ratios needs to be utilized for accurate modeling. In this work, phase-only activation will be computed directly from Lathi's mathematically correct non-Normal distribution, yielding additional physiological information beyond what is typically observed.

Predicting missing segments in partially observed functions is challenging due to infinite-dimensionality, complex dependence within and across observations, and irregular noise. These challenges are further exacerbated by the existence of two distinct sources of variation in functional data, termed amplitude (variation along the y-axis) and phase (variation along the x-axis). While registration can disentangle them from complete functional data, the process is more difficult for partial observations. Thus, existing methods for functional data prediction often treat phase variation as negligible. Furthermore, they typically require precise model specifications and/or rely on computationally intensive tools, such as bootstrapping, to construct prediction intervals. We propose a unified registration and prediction approach for partially observed functions using conformal prediction. Our method integrates registration and prediction in one algorithm while ensuring exchangeability through carefully constructed predictor-response pairs. Using a neighborhood smoothing algorithm, the framework produces pointwise prediction bands with finite-sample marginal coverage guarantees under weak assumptions. The method is easy to implement, computationally efficient, and permits simple parallelization. Numerical studies and real-world data examples demonstrate the effectiveness and practical utility of our method.

Machine learning is used in neuroscience to examine brain-phenotype associations and facilitate individual prediction from high-dimensional brain imaging. For continuous phenotypes, Pearson's correlation between the observed and predicted phenotype is used to quantify model accuracy in testing data. However, recent research suggests millions of samples may be needed to reliably estimate the maximum achievable predictive accuracy (MAPA). We formally define the MAPA and show that Pearson's estimator is biased for this quantity and its confidence intervals fail to capture the target. We develop a semiparametric (double machine learning) one-step estimator that more accurately estimates the MAPA and yields valid confidence intervals across flexible machine learning settings. Analyzing data from the Reproducible Brain Charts dataset, we show that this estimator has smaller bias when estimating brain-phenotype associations of neuroimaging data with age and psychopathology phenotypes. We further extend by improving the estimation of area under the curve (AUC) for MAPA for binary phenotypes.

Mass-univariate analysis remains the workhorse of brain-wide association studies, where tens of thousands of imaging-derived phenotypes are modeled independently. Yet neuroimaging outcomes demonstrate strong, structured correlation patterns driven by latent brain network organization, and ignoring this dependence structure reduces statistical efficiency and limits replicability. We present a network-guided multivariate regression framework that integrates covariance network structure into multivariate regression analysis to explicitly account for inter-correlations among neuroimaging measures. This approach provides an alternative to conventional mass-univariate modeling with dependence adjustment, analogous to the role of mixed-effects models for repeated measures in lower-dimensional settings. Through extensive simulations and an application to imaging-derived phenotypes from the ABCD study, we show that network-guided modeling substantially improves inferential accuracy and cross-study replicability.

Discussant

Functional brain networks estimated from fMRI data are increasingly used in cognitive analyses. However, most statistical methods impose sparsity at the level of individual connections. We propose a group-structured sparse precision matrix framework for estimating functional connectivity that aligns statistical regularization with scientifically meaningful network organization. The proposed approach performs model selection over groups of connections, yielding interpretable network estimates. The framework is formulated through a general penalized-likelihood approach that accommodates multilevel group structures and unifies several existing sparse precision estimators as special cases. We investigate identifiability conditions, scalable optimization algorithms for high-dimensional settings, and stability tools for assessing the reliability of selected network groups. Simulation studies demonstrate improved recovery and stability of group-level structure relative to existing competing methods. An application to the Tennessee Alzheimer's Project shows how group-structured sparsity produces brain networks that are more interpretable and more directly related to cognitive outcomes. All reproducible simulation code and user-friendly utility functions are provided in an open-source R package.

Causal mediation analysis provides critical insights into how exposures influence outcomes through intermediate variables, or mediators. In this study, we examine mediation effects of complex-structured data represented as whole-brain connectivity networks derived from resting state fMRI (rs-fMRI) data on Alzheimer's Disease (AD) subjects and healthy controls. Our modeling approach incorporates low-dimensional latent scales representations of the high-dimensional connectivity matrices, while preserving node-level characteristics and facilitating the identification of key mediating brain regions. It also accommodates flexible non-linear relationships between the mediators and the outcomes of interest. Using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI-1), we perform cross-sectional and longitudinal analyses to investigate the effects of several factors known to affect cognitive decline. Results confirm the important role of amyloid-beta as a risk factor for cognitive impairment. They also confirm the key mediation role played by the default mode network, reinforcing its central role in AD. The longitudinal analysis additionally identifies a distinct subset of brain network nodes, beyond those found in the cross-sectional analysis, that uniquely drive baseline and/or time-varying indirect causal effects, particularly in the visual network.

Abstract to be announced

Functional magnetic resonance imaging (fMRI) data are inherently complex, characterized by high dimensionality, intricate inter-regional dependencies, and substantial individual variability across experimental paradigms. Traditional linear mixed models (LMMs) often fail to adequately capture nonlinear relationships inherent in neuroimaging data. To address these limitations, we introduce nonlinear mixed model (NMM), an extension of the LMM framework that integrates neural networks to flexibly model complex fixed-effect relationships while preserving the random effects structure to account for individual differences. We applied the NMM to Philadelphia Neurodevelopmental Cohort (PNC) across emotion, n-back, and resting-state paradigms, which achieved superior model fit relative to classical LMMs. This framework offers a statistically rigorous and practically explainable approach for modeling large-scale functional connectivity from modest covariates while explicitly separating population-level effects from stable individual variability in functional brain organization.

AI agents can serve as virtual research collaborators in imaging genetics, accelerating discovery and translation by orchestrating complex, end-to-end workflows that span genomic data processing, imaging phenotype engineering, association and prediction modeling, replication, and interpretation. Realizing this potential requires robust domain-specific datasets, standardized pipelines, and scalable computational infrastructure, as well as interdisciplinary collaboration to establish community standards that ensure transparency, reproducibility, and responsible deployment. In this talk, we will present an agentic framework for imaging genetics, highlight practical use cases in complex disease and aging-related brain phenotypes, and discuss how human-in-the-loop, source-traceable evidence synthesis can improve real-world utility of data resources and AI tools.

Many modern instruments, such as wearables, imaging devices, and geospatial sensors, produce rich subject-level data streams containing thousands to millions of measurements. Reducing these data to simple summaries (e.g., means) can obscure meaningful structure. We introduce a general distributional regression framework for distribution-on-scalar or distribution-on-function regression, enabling characterization of how subject-specific distributions vary across conditions and covariates. Our modeling strategy applies a series of transformations to each subject's empirical distribution function, producing a compact, low-dimensional embedding that is near-lossless under the quantile-domain Wasserstein metric. These embeddings preserve essential distributional information, accommodate measurement discreteness arising from instrument precision, satisfy Gaussian assumptions required for regression in the embedding space, and maintain monotonicity in the quantile domain. We further describe efforts to construct generative models that produce valid subject-level distributions reflecting the salient structure of the underlying population and examine the statistical properties of inference derived from these models, and discuss inferential strategies.

Imaging data are often summarized by a single value per region of interest, a practice that can reduce statistical power and obscure important spatial heterogeneity. Voxel-wise analyses better capture the complexity of imaging data, but accurate estimation of standard errors in this setting is challenging due to spatial correlation among voxels. We propose a spatial bootstrap framework for voxel-level inference that accounts for spatial dependence while preserving statistical power. The method is demonstrated through an application examining the association between Quantitative Susceptibility Mapping and Myelin Water Fraction, illustrating its ability to provide reliable standard error estimates and robust statistical inference.

Interpreting brain-behavior relationships through the lens of anatomical parcellations or functional networks is commonplace in human brain mapping. However, statistical approaches for testing whether brain-behavior associations are stronger (i.e., enriched) within a region of interest remain underdeveloped. Here, we propose a permutation-based approach for network enrichment testing using ordinal dominance curves (NETDOM). In simulation studies, we demonstrate that NETDOM properly controls the type I error rate—unlike other prominent methods—while exhibiting increased statistical power when brain-behavior associations are elevated in a subset of in-network locations. Using data from two large-scale neurodevelopmental cohorts, we illustrate that NETDOM effectively detects enriched associations between structural and functional brain measures and neurocognitive performance.

The fidelity of medical diagnostic images, such as those acquired via ultrasound, directly influences the validity of downstream statistical inference and clinical decision-making. Low-resolution (LR) images impose information loss that classical deterministic upscaling methods, such as bicubic interpolation, cannot recover. These traditional methods operate effectively as low-pass filters, attenuating high-frequency structural content and introducing systematic bias in downstream classification and inference tasks. Alternatively, we propose a four-stage pipeline that leverages an Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) as a generative engine for super-resolution, integrated with a rigorous statistical verification process to ensure that the resulting high-resolution (HR) images represent a measurable improvement in structural detail over the original LR data. The pipeline begins with a preprocessing stage that standardizes image intensities and corrects for acquisition-related variation across heterogeneous inputs, ensuring the model learns biologically meaningful structure rather than scanner-specific artifacts. The generative core employs a relativistic discriminator paired with a perceptual loss functional, framing super-resolution as estimation on the HR image manifold rather than pointwise mean prediction. This overcomes the well-known regression-to-the-mean phenomenon induced by mean squared error-based objectives, which produces posterior mean estimates that systematically suppress high-frequency variation. A subsequent standardization stage applies pixel-level normalization and stochastic augmentation to stabilize empirical feature distributions prior to downstream analysis. The pipeline concludes with a formal statistical verification step, auditing pixel intensity distributions and screening for non-biological artifacts, providing a principled quality-control procedure analogous to residual diagnostics in classical regression. This work frames medical image super-resolution explicitly as a statistical estimation problem, and offers the imaging community a reproducible, auditable pipeline for enhancing LR datasets in ways that demonstrably improve sensitivity and specificity of diagnostic models.

Radiomics features often show high intercorrelation due to latent factors like tumor volume, distorting feature importance in ML models. We present RadiomiXAI, a model-agnostic explainability framework using Bayesian networks to adjust importance scores by modeling feature dependencies. Using 476,424 radiomics feature vectors extracted from OpenRadiomics via PyRadiomics, we computed Pearson correlations with tumor volume and identified 74 features with correlation coefficients >0.8, suggesting strong dependency. Multiple normalization techniques were evaluated for variance inflation factor (VIF) reduction. A LightGBM classifier was trained on BraTS 2020 T1 Contrast Enhanced MRI data using a 60/20/20 train/validation/test split. Feature importances were estimated using RadiomiXAI, a perturbation-based method that replaces each feature with sampled values from its range and quantifies the mean absolute change in predicted probabilities. Uncertainty intervals for each importance score were then derived using the BN model: incoming dependencies reduce the score (lower bound), and outgoing dependencies inflate it (upper bound), proportional to mutual information values on the BN edges. Normalization techniques failed to adequately decorrelate most features from volume. RadiomiXAI identified an outgoing dependency-adjusted upper bound of 0.2059 for volume's base importance score of 0.0087, revealing its potential downstream overrepresentation. RadiomiXAI reveals how latent common-cause variables such as tumor volume can inflate feature importance scores in radiomics pipelines. Conventional normalization fails to resolve these dependencies. BN-guided uncertainty bounds provide a novel explainability framework to adjust feature interpretations, aiding robust biomarker identification.

Deep learning with ECG data in the ICU is not only about building accurate models, but about creating AI tools that can truly help doctors make better decisions. Atrial fibrillation (AFib) is one of the most common and serious heart rhythm problems in critically ill patients, and can increase the risk of complications and death. However, ICU ECG signals are noisy and full of artifacts, and constant alarms make it hard to reliably detect these brief AFib events. In this project, we follow the evolution of AI methods for AFib detection in this challenging environment. We begin with classical machine learning based on carefully designed ECG and heart-rate-variability features, then move to deep neural networks that learn directly from raw signals, first trained on large public datasets and then adapted to ICU data using transfer learning and weak labels. We next explore large ECG foundation models pretrained on huge datasets and fine-tune them for ICU use. Finally, we develop our own large model trained from scratch using physiologically guided self-supervised contrastive learning, allowing us to fully exploit the vast amount of unlabeled ICU ECG data. Together, this work shows how ICU ECG analysis can progress from traditional modeling to powerful, robust, and clinically actionable deep learning systems.

Spatial high-dimensional data is now routine in computer vision. Modern vision foundation models produce dense embeddings that capture rich semantic structure across an image. Similar spatial high-dimensional measurements also arise in mass spectrometry imaging, where each location contains a molecular signature. These representations are valuable for discovery and downstream decision-making. However, they are difficult to interpret directly. A standard approach is global dimensionality reduction to form a low-dimensional view. In practice, global projections can wash out subtle variation and local heterogeneity. We propose an interactive spatial prompting method, with supporting software, that uses user-provided spatial cues to guide dimensionality reduction. This produces higher-contrast views that better support exploratory analysis.

In mediation analysis, the exposure often influences the mediating effect, i.e., there is an interaction between exposure and mediator on the dependent variable. When the mediator is high-dimensional, it is necessary to identify non-zero mediators (M) and exposure-by-mediator (X-by-M) interactions, research is scarce in preserving the underlying hierarchical structure between the main effects and the interactions. To fill the knowledge gap, we develop the XMInt procedure to select M and X-by-M interactions in the high-dimensional mediators setting while preserving the hierarchical structure. Our proposed method employs a sequential regularization-based forward-selection approach to identify the mediators and their hierarchically preserved interaction with exposure. Our numerical experiments showed promising selection results. Further, we applied our method to ADNI morphological data and examined the role of cortical thickness and subcortical volumes on the effect of amyloid-beta accumulation on cognitive performance, which could be helpful in understanding the brain compensation mechanism.

Mediation analysis decomposes the effect of an explanatory variable on an outcome into direct and indirect components through mediators. Conventional mediation models assume constant effects, which can be overly restrictive when causal relationships vary with covariates. We propose a dynamic high-dimensional mediation framework based on varying coefficient models that allows the effects of the explanatory variable on both the mediator and the outcome to change smoothly with covariates. Estimation is carried out using B-spline approximations, with an Adaptive Lasso penalty to enable variable selection and control model complexity in high-dimensional settings. We establish convergence rates and asymptotic distributions for the proposed estimators. For inference, we develop an F-test for direct effects and a partially penalized Wald test for indirect effects, both of which are shown to have chi-square limiting distributions under the null, with noncentral chi-square behavior under local alternatives for the direct effect test. Simulation studies demonstrate the validity and robustness of the proposed methods. An application to Alzheimer's Disease Neuroimaging Initiative MRI data identifies region-of-interest–based volumetric mediators and reveals the dynamic influence of age on Alzheimer's disease. Overall, the proposed method provides a flexible tool for causal inference in high-dimensional environments.

Accurately modeling longitudinal trajectories of structural MRI (sMRI)–derived biomarkers along Alzheimer's disease (AD) progression is critical for understanding disease dynamics, yet remains challenging because available datasets typically capture only fragmentary segments of an individual's decades-long disease course. As a result, standard longitudinal analyses are often limited to subsets of participants with observed clinical conversion, and therefore, suffer from reduced statistical power and selection bias. To address these challenges, we previously proposed a double anchoring events–based sigmoidal mixed model (DSMM) and applied it to characterize cognitive decline trajectories using time to incident AD diagnosis as a clinically meaningful time scale, while uniquely enabling inclusion of participants without an observed incident AD diagnosis. In this work, we extend the DSMM framework to the characterization of longitudinal sMRI-derived biomarkers in participants with AD. Using harmonized multi-cohort data from the Alzheimer's Disease Sequencing Project Phenotype Harmonization Consortium (ADSP-PHC), the proposed approach aligns heterogeneous imaging trajectories relative to AD onset and integrates information from participants without observed AD conversion. This unified modeling strategy reduces selection bias, provides excellent clinical interpretability, and enables estimation and comparison of the population-level temporal ordering of changes in multimodal measures along the AD progression continuum.

Within-individual coupling between measures of brain structure and function evolves across development and may underlie differential risk for neuropsychiatric disorders. Despite growing interest in the developmental trajectories of structure–function relationships, rigorous methods to quantify and test individual differences in coupling remain limited. In this talk, we outline a framework for defining intermodal coupling and parameterizing individual differences, and show how several existing approaches can be interpreted as statistical inference via cluster enhancement under specific assumptions. We then discuss methods for estimating the genetic heritability of intermodal coupling and for conducting valid statistical inference, highlighting key methodological challenges. We illustrate these approaches using data from the Philadelphia Neurodevelopmental Cohort (PNC) and the Human Connectome Project (HCP). Overall, our findings raise a cautionary note regarding intermodal coupling analyses, particularly with respect to proper control of false positives. This is joint work with Ruyi Pan and Ruilin Bai.

Work on multiple networks has typically focused on estimating their shared structure. Two-sample tests for networks have also been developed, testing the hypothesis of two samples of networks coming from the same distribution. However, scientifically relevant hypotheses rarely take this form: for example, in neuroimaging, a common application for multiple network analysis where each network represents a patient's brain connectome, it is rarely of interest to compare whole brains of patients and healthy controls; more often, the focus is on a particular brain region. Beyond comparisons, it is also of interest to estimate structures that are specific to a disease, or to some other trait in patients. One could always do that using just the patients with that trait, but using all available samples allows us to better estimate structures that are shared by all, which in turn helps separate out the structure associated with a trait. This talk will introduce two methods that help address these challenges: mesoscale testing on networks, which allows for formal hypothesis testing on a subset of edges (like a brain region) which leverages the rest of the network to increase power; and group MultiNeSS, a method that takes a sample of networks and estimates structures that are shared by all, specific to groups corresponding to a trait, or just unique to an individual. In both cases, we leverage the assumption of low-rank expectation of adjacency matrices which has been observed widely in practice. Based on joint work with Peter MacDonald, Alexander Kagan, and Ji Zhu.

Abstract to be announced

Abstract to be announced

Abstract to be announced

Abstract to be announced

Abstract to be announced

Abstract to be announced

Age-related changes in brain structural networks reflect ongoing remodeling of white matter pathways across the lifespan. However, empirical findings on age-related changes in structural networks are often inconsistent across studies, likely due to methodological heterogeneity in diffusion MRI preprocessing, tractography, and network construction. In this study, we assess the reproducibility of estimated age-related changes in structural connectivity (SC) across large neuroimaging cohorts and analytic choices. Using diffusion MRI data from five major datasets, including the Human Connectome Project Young Adults (HCP-YA), Lifespan Human Connectome Project in Development (HCP-D), Lifespan Human Connectome Project in Aging (HCP-A), UK Biobank, and the Philadelphia Neurodevelopmental Cohort (PNC), we examine how preprocessing pipelines, tractography algorithms, and brain parcellation schemes influence estimates of lifespan SC changes. We systematically vary preprocessing pipelines, tractography algorithms, and brain parcellation schemes to construct structural connectomes. Our results highlight the substantial impact of methodological choices on estimated age-related changes in SC and provide guidance for improving the robustness and interpretability of SC analyses.

Abstract to be announced

We consider the setting where many networks are observed on a common node set, and each observation comprises edge weights of a network, covariates observed at each node, and an overall response. The goal is to use the edge weights and node covariates to predict the response while identifying an interpretable set of predictive features. Our motivating application is neuroimaging, where edge weights encode functional connectivity measured between brain regions, node covariates encode task activations at each brain region, and the response is disease status or score on a behavioral task. We propose an approach that constructs feature groups based on assumed community structure (naturally occurring in neuroimaging applications). We propose two feature grouping schemes that incorporate both edge weights and node covariates, and we derive algorithms for optimization using an overlapping group LASSO penalty. Empirical results on synthetic data show that our method, relative to competing approaches, has similar or improved prediction error along with superior support recovery, enabling a more interpretable and potentially more accurate understanding of the underlying process. We also apply the method to neuroimaging data from the Human Connectome Project. Our approach is widely applicable in neuroimaging where interpretability is highly desired.

Is it possible to read the content of human thought out using recordings of brain activity? We use non-invasive functional MRI and machine learning methods based on large language models to investigate the relationship between brain activity and the content of thought. These methods reveal complex spatial and temporal patterns of activity that relate to specific semantic categories. We show that this information can be read out as language, even when the stimulus evoking it is from another modality. We also use self-supervised speech models to show that fine temporal information can be deduced even from slow non-invasive recordings. These results point to a future of neuroscience that strongly integrates modern neural network models.

To enhance prostate cancer risk stratification through explainable artificial intelligence (XAI) by developing a pipeline that integrates principal component analysis (PCA) for dimensionality reduction and restores interpretability in radiomics-based classification using multiparametric MRI. We used the 2021 PI-CAI dataset from RadMLBench, comprising 3,045 radiomics features extracted from T2-weighted, apparent diffusion coefficient (ADC) maps, and diffusion-weighted imaging (DWI) MRI sequences of 969 patients. The task involved binary classification of csPCa (Gleason >3+3, n=637) vs. non-csPCa (n=332). We applied 10-fold nested cross-validation using LightGBM and compared baseline performance with a dimensionality-reduced variant utilizing PCA preserving 95% of variance. The baseline 10-fold AUC without dimensionality reduction was 0.701. Incorporating PCA improved performance, achieving an AUC of 0.753. On a representative 60/20/20 split, the test set yielded an AUC of 0.735, accuracy of 0.727, and F1-score of 0.819. To overcome PCA's limitation in post-hoc interpretability, a feed-forward neural network with a single hidden layer (256 units) was trained to reconstruct the full radiomics space. Feature importance scores from LightGBM were projected through this mapping and transformed into the original feature domain using the stored eigenvector matrix. Post-hoc interpretability analysis identified top contributing features in the original radiomics space. This study addresses the challenge of interpretability in high-dimensional radiomics datasets by reconciling dimensionality reduction with explainable machine learning, supporting transparent prostate cancer risk modeling.

Ventricular and atrial fibrillation are major causes of morbidity and mortality with enormous clinical-societal burden. Recent experimental and computational research studies suggest that fibrillation can be sustained by functional re-entrant patterns of excitation, also termed rotors, but their dynamics remain unclear. In fundamental terms, the heart is considered an excitable medium whose spatio-temporal patterns of excitation are governed by a reaction-diffusion process. Thus, we propose a physics-informed video prediction framework for modeling rotor-driven excitation dynamics using simulated voltage-map sequences generated from FitzHugh-Nagumo reaction-diffusion systems. The model embeds physics-inspired inductive bias through a latent space representation, and then performs reaction-diffusion-based temporal updates, and spatial coupling via a fixed Laplacian operator, enabling interpretable forecasting of excitation dynamics from spatiotemporal voltage maps.

This study evaluates a novel radiomics-based vector database (vectorDB) retrieval pipeline for MRI-driven risk stratification of prostate cancer, aiming to improve the accurate identification of clinically significant prostate cancer (csPCa) lesions amidst the broad spectrum of indolent tumors and benign prostatic conditions, and compares its performance against a supervised learning baseline. We utilized the 2021 PI-CAI dataset from RadMLBench, which includes 3,045 radiomics features extracted from T2-weighted, ADC maps, and DWI MRI sequences for 969 patients. A LightGBM classifier with nested cross-validation (20 outer folds, 19-fold inner loop) served as the baseline. For the proposed method, we implemented a vector database using Facebook AI Similarity Search (FAISS). In each outer fold, the training set was indexed into a vectorDB. For each test case, the top-K nearest neighbors based on Euclidean distance were retrieved, and the final label was determined by majority voting. The baseline LightGBM model achieved an accuracy of 0.716, an F1-score of 0.801, a precision of 0.742, and a recall of 0.873. The proposed vectorDB pipeline yielded an accuracy of 0.677, an F1-score of 0.790, a precision of 0.689, and a notably higher recall of 0.926. While the VectorDB method did not demonstrate statistically significant performance differences overall compared to the supervised baseline (p = 0.185), it achieved a comparable F1-score and recall, reducing computational runtime by over 90%. Radiomics features can be effectively repurposed as embedding vectors for similarity-based classification via vector databases, offering a scalable, explainable, and efficient alternative to traditional supervised models for prostate cancer risk stratification suitable for real-time and large-scale applications.

The #Enzian classification provides a comprehensive 14-compartment framework for describing the anatomical extent of endometriosis on imaging. Despite its clinical value, routine implementation in radiology reporting remains limited due to the complexity of manual extraction and scoring from narrative MRI reports. Large language models (LLMs) may enable automated extraction of structured #Enzian scores from free-text reports and potentially support clinical training and standardized reporting. To evaluate the ability of online and locally deployed LLMs to automatically extract #Enzian scores from pelvic MRI reports and compare their performance with radiology trainees. A single-center retrospective study was conducted using 186 pelvic MRI reports from patients with suspected or confirmed endometriosis. Ground-truth #Enzian scores were established through manual review by a fellowship-trained uroradiologist. Twelve radiology trainees without prior #Enzian experience independently scored a subset of 50 cases. Six online LLMs and six local LLMs were evaluated using one-shot prompting, and classification accuracy was compared across the 14 compartments of the #Enzian system. Online LLMs demonstrated consistently high performance with pooled multiclass accuracies ranging from 89.1% to 92.4%, with Gemini 2.5 Pro achieving the highest accuracy (92.4%) and outperforming trainees in 11 of 14 categories. Local models showed performance strongly dependent on model size; the best-performing local model (GPT-OSS 120B) achieved 87.1% average accuracy, whereas smaller models showed substantially lower performance. Cost analysis revealed considerable variation among online APIs (ranging from $3.82 to $42.26 for 186 cases), while local deployment provided a privacy-preserving alternative. LLMs can reliably extract structured #Enzian scores from pelvic MRI reports. Online models currently demonstrate the highest performance and may support training and reporting workflows, whereas large local models provide a viable privacy-preserving alternative with slightly lower accuracy.

Quantifying spatial heterogeneity in medical imaging is important for understanding structural variability and improving imaging-derived biomarkers, which can be beneficial for machine learning-based modeling. In this work, we introduced a Shannon entropy-based methodology designed to quantify spatial heterogeneity in brain MRI using radiomics-derived traits. In this approach, each brain image was partitioned into regions, and radiomics features were extracted from each region using a standardized radiomics pipeline. Continuous feature values were then transformed into categorical traits, allowing the construction of trait distributions across spatial regions. The proposed methodology summarizes the overall spatial heterogeneity of MRI images by aggregating entropy-based heterogeneity measures derived from these trait distributions. This formulation provides an interpretable metric that captures the diversity of radiomic patterns across brain regions. The proposed framework offers a systematic and reproducible approach for translating radiomics features into quantitative measures of spatial complexity, enabling comparative heterogeneity analysis across medical imaging datasets and supporting future studies in developing heterogeneity-aware machine learning models trained by MRI images. The framework was evaluated on two brain MRI datasets: the IXI healthy adult dataset (n = 270) and the BraTS-PED pediatric brain tumor dataset (n = 251). Using first-order radiomic entropy as an example feature and dividing each brain into four spatial quadrants, the proposed method produced heterogeneity average index (HTA) of 0.820 for the IXI dataset and 0.884 for the BraTS-PED dataset. Individual quadrant-level indices (HTI) exhibited substantial variance (0.618 to 0.995), reflecting the tool's sensitivity to localized structural diversity. These results indicate that the framework provides a standardized scale for comparing spatial complexity across diverse cohorts, offering a necessary precursor for integrating heterogeneity-aware features into predictive machine learning models.