Stephen Strother - A Personal History of Statistical Modelling in Neuroimaging
It is 40 years since the importance of mean differences versus correlation and covariance measures in neuro-imaging were first debated in the literature (Clark et al., 1984). Two years after that I finished my PhD on PET scanners at the Montreal Neurological Institute where I had the luck of working with Dr. Keith Worsley, and then helped to coauthor one of the earliest multivariate data analysis models (Moeller et al., 1986). Within the next decade the first papers on neural networks and predictive modeling in PET (Kippenham et al., 1992; Lautrup et al., 1995) and fMRI (Morch et al., 1997) appeared. Drawing on this rich 40-year history of statistical modelling in PET and fMRI I will briefly discuss the importance of multivariate data analysis in behavioural and disease modelling, and the role of predictive resampling and optimisation of neuroimaging pipelines. I will end with some speculation about data standards and the importance of collaborative big science to the future of multivariate statistical modelling in neuroimaging.
Biography: Until he retired in 2022, from 2004 Dr. Strother was a senior scientist at the Rotman Research Institute, Baycrest, and Professor of Medical Biophysics at University of Toronto, Canada. He trained at Auckland University (BSc/MSc) and McGill University, Montreal, Canada (PhD., Electrical Engineering, 1986). Following a fellowship at MSK Cancer Center, New York, USA, from 1989 he was a neuroimaging Physicist, and Assistant Professor of Radiology, University of Minnesota. In 2004 he moved to Toronto. His research interests include neuro-informatics and data science for neuroimaging and big clinical data sets using statistical and machine learning techniques. Until 2022 he was a senior investigator in the Canadian Open Neuroscience Platform (comp.ca), Ontario neurodegeneration (ondri.ca) and Canadian national depression (canbind.ca) research programs. He is a past chairman of the international neuroinformatics standards committee, at the National Institutes of Health, USA, and a past Canadian representative to INCF (https://www.incf.org/). In 2012 he cofounded ADMdx, Inc. in Chicago, USA (admdx.com), a medical analysis and diagnostics company, and he sits on the boards of InDoc Research (indocresearch.org), a medical informatics company in Toronto, Canada, He currently enjoys retired life in Nelson, New Zealand.
Tentative Program -- subject to change! (especially sessions/speaker order)
Chair : Anqi Qiu
Tianye Jia - A longitudinal DNA methylation atlas and its link to brain structure and mental health
We identified 18 longitudinal blood DNA methylation (DNAm) clusters—9 linked to brain-related functions—using a combined WGCNA and longitudinal approach. These patterns were robustly validated across three diverse lifespan datasets (IMAGEN, PPMI, ADNI), spanning adolescence to older adulthood. Longitudinal DNAm changes (notably clusters 3 and 9) correlated with brain maturation, particularly in neurodegeneration-enriched regions, while clusters 1 and 7 were associated with depression and psychosis. Importantly, DNAm clusters contributed to mental health symptoms (especially depression) independently of polygenic risk scores, highlighting their unique role in neurodevelopment and psychiatric disorders.
Aurina Arnatkeviciute - Null models for imaging transcriptomics
The emergence of brain-wide transcriptional atlases, quantifying the expression of thousands of genes across multiple locations in the brain, opened new opportunities for investigating the molecular correlates of brain network organization. Spatial gene expression data are highly multidimensional, therefore, evaluating associations with neuroimaging measures requires additional considerations across multiple levels of analysis. In this talk I will briefly introduce general concepts of imaging transcriptomics analyses and then will focus on the three main areas for applying different null models: i) dealing with spatial autocorrelation when evaluating the associations between gene expression and neuroimaging data; ii) assessing the specificity of identified associations for the selected sets of genes; and iii) evaluating the functional implications through gene enrichment analyses. I will conclude by outlining the recommendations for best practices and providing an overview of freely available open-source toolboxes implementing current best-practice procedures.
Petra Vertes - From cortical gradients to gene regulatory networks
In this talk, I will discuss the opportunities and challenges of using multi-omic data to explain how molecular mechanisms may give rise to neuroimaging features. I will focus on recent work that integrates bulk transcriptomics, single-nuclei RNAseq, and imaging genetics to triangulate a gene regulatory program that drives the development of cortical anatomy. This gene regulatory program activates in adolescence and is enriched for the genetic risk of both schizophrenia and depression. I will also review the methodological advances in data filtering and analytic algorithms that proved necessary to uncover robust and biologically meaningful signals across diverse multi-omic datasets.
Sarah Medland - on behalf of the ENIGMA GWAS working group - Examining the impact of common genetic variants on the gross morphology of the cerebellum within the ENIGMA consortium.
While the cerebellum is best known for motor related functions, it plays an important role in many aspects of cognition including language and attention. Most genetic studies of the cerebellum have focussed on identifying pathological variants impacting ataxias and neurodegenerative disorders. Recent GWAS within the UK Biobank have show common genetic variants can explain substantial variation in cerebellar volume. Here we present results from the cerebellar GWAS within the ENIGMA consortium bringing together from 45 cohorts and approximately 100,000 people. Importantly, unlike the UK Biobank analyses these participants are not restricted in age and cover the lifespan from adolescence to late adulthood. The analyses also look beyond European ancestry populations and consider the generalisation of findings to other ancestries. Results will be presented at the region of interest (28 regions) regional summary (6 summary variables) and multivariate levels. Relationships with other brain structures, psychiatric and neurological traits will be presented focussing on genetic correlations and evidence for causation. These results, which take a more inclusive approach and include more diverse samples than previous studies, markedly increase our understanding of the genetic architecture of the cerebellum.
Alina Tetereva - Enhancing the Prediction of Individual Differences in Cognitive Abilities Using Multimodal Brain Imaging and Stacked Models
In this talk, I will discuss how multimodal neuroimaging can be leveraged to improve the prediction of individual differences in cognitive abilities. Brain-wide association studies have long aimed to link brain phenotypes with cognitive outcomes but often face limitations in predictability, reliability, and generalisability. I will present a machine-learning “stacking” framework that integrates diverse MRI features—including task-based activations, resting-state and task-based functional connectivity, and structural measures—into a single predictive model. Using data from the Human Connectome Project (Young Adult and Aging cohorts) and the Dunedin Study, I will highlight how stacking enhances cognitive prediction performance across datasets and age ranges, and discuss the implications for reproducible, generalisable brain-behaviour modelling.
Christian Beckmann - From activity to connectivity to connectopic mapping: Evolving Paradigms in Functional Brain Mapping
Functional neuroimaging has undergone a conceptual transformation over the past two decades, moving from detecting locality of activation to characterising distributed systems. A more recent development—connectopic mapping—extends this trajectory by characterizing fine-grained, topographically organized gradients of connectivity within brain areas.
In this talk, I will outline this methodological evolution, highlighting how advances in statistical modeling, dimensionality reduction, and manifold learning have enabled new ways to understand the brain’s intrinsic functional architecture that captures continuous spatial organization—such as somatotopy, retinotopy, and abstract cognitive gradients—that are invisible to traditional parcellation or seed-based methods. I will also discuss recent work on individual-specific mapping and the implications of connectopic analysis for precision neuroscience, developmental studies, and clinical translation.
Biography: Christian Beckmann is Professor for Statistics in Imaging Neuroscience at the Donders Institute at the Radboud University Medical Centre in Nijmegen, The Netherlands and is a research fellow at the University of Oxford. He is a co-developer of several core tools in FSL (FMRIB Software Library) and his work bridges methodological rigor with real-world applications in cognitive neuroscience and psychiatry.
Chairs: Baptiste Couvy-Duchesne, Thomas Nichols
Paul Thompson - Generative AI and Vision-Language Models in Neuroimaging: Some Advances & Open Questions
Generative AI models can create photorealistic images and videos from text prompts, sparking great interest in adapting them for medicine and neuroscience. We focus on models based on statistical thermodynamics and optimal transport (latent diffusion and Schrödinger bridges), which enable us to train deep neural networks to model natural variations in brain images and how they depend on age, sex, disease, and imaging modality [1]. We train a vision-language model to link natural language and brain images, with promising applications to disease diagnosis via visual question answering [2]. We discuss open questions on generalization [3], memorization, privacy, and training needs for this new type of model. This is joint work with Nikhil Dhinagar, Tamoghna Chattopadhyay, and our team at USC.
[1] https://pubmed.ncbi.nlm.nih.gov/40039528/ [2] https://www.biorxiv.org/content/10.1101/2025.02.15.638446v1.full.pdf [3] https://arxiv.org/html/2402.18491v1
Dan Kessler - Predicting Responses from Weighted Networks with Node Covariates in an Application to Neuroimaging
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.
Eardi Lila - Disentangling Genetic Contributions to Human Brain Connectivity Using an Efficient Estimator of Variance Components in Multivariate Random Effects Models
We introduce a novel estimator for the genetic and environmental covariance matrix of multidimensional phenotypes. This estimator forms the basis of a regression framework that allows us to characterize phenotype relationships driven only by genetic and environmental factors. We apply the framework to identify genetically and environmentally driven associations between structural and functional connectomes.
Xi-Nian Zuo - Charting Brain Health Fostered by Chinese Color Nest Data Community at Science Data Bank
Quantifying individual deviations of brain morphology from normative references is essential for understanding neurodiversity and facilitating personalized brain health management. We established Chinese brain normative references using morphological imaging scans of 24,061 healthy volunteers from 105 sites, revealing later peak ages of lifespan neurodevelopmental milestones (1.2-8.9 years) than the European/North American references. We then modeled individual brain deviation scores in 3,932 patients with different neurological disorders from the population references to evaluate three key missions of charting brain health using machine learning approaches: (1) estimating disease propensity, (2) predicting cognitive and physical outcomes, (3) assessing treatment effects with distinct disability progression. The norm-deviation scores outperformed raw structural measures in these evaluations, and their utilities and potentials were confirmed by longitudinal and sensitivity analyses. As part of the Global Research Data Ark Initiative, Chinese Color Nest Data Community at Science Data manages data sharing of the large-scale neuroimaging resource and presents how the Chinese-specific normative brain references may foster personalized diagnosis and prognosis in neurological diseases, enabling clinically applicable assessments of brain health.
Wes Thompson - Design and Analysis of the HEALthy Brain Child Development (HBCD) Study.
The HEALthy Brain and Child Development (HBCD) Study represents a pioneering nationwide effort aimed at comprehensively mapping early human brain and behavioral development from prenatal stages through childhood. Recognizing the rapid neurodevelopmental processes occurring during infancy—marked by dramatic brain growth, synaptogenesis, pruning, and myelination—the HBCD study strategically collects extensive longitudinal data across biological, psychological, and environmental domains. HBCD is recruiting a large, socio-demographically diverse sample of over 7,000 pregnant individuals and their infants, including targeted subgroups experiencing prenatal substance exposure and socioeconomic adversity. The HBCD employs sophisticated multimodal methodologies, including neuroimaging (MRI, EEG), biospecimen collection, wearable sensor data, and detailed developmental assessments, to identify biomarkers of early risk and resilience factors. Its adaptive recruitment strategy and statistical design enable robust causal inference regarding the impact of prenatal substance exposure, maternal psychosocial stress, and early caregiving environments on child outcomes.
Gang Chen - Neuroimaging modelling with data-generating processes for robust inference
Neuroimaging data exhibit hierarchical complexity, requiring models that reflect their data-generating processes and causal mechanisms. This approach strengthens validity and tackles critical fMRI issues: multiple comparisons, limited sample sizes, test-retest reliability, heritability estimation, result interpretation, and reproducibility assessment. By addressing these issues, we enhance the robustness and generalizability of fMRI findings in an integrated framework.
Biography: Gang Chen joined the SSCC in 2003 as a Mathematical Statistician. He provides consulting on any statistics-related issues, and keeps developing new statistical tools in AFNI. SSCC is the Scientific and Statistical Computing Core of the NIMH Intramural Research Program. The primary function of this core is to support functional neuroimaging research at the NIH. This includes development of new data analysis techniques, their implementation in the AFNI software, advising researchers on the analysis methods, and instructing them in the use of software tools
Chair: Jean-Baptiste Poline
Vince Calhoun - Functional brain signatures of mental health
The brain is intrinsically dynamic and prior work shows sensitivity to mental health. Existing studies are hard to compare and generalize due to the lack of correspondence among dynamic features. Here, we present a dual-constrained approach to adaptively estimate time-resolved features, revealing a wealth of information about brain dynamics and disorder in four mental disorders. We also evaluate the manifestation of this ‘dynamic fingerprint’ in typically developing individuals and those at clinical high risk. Using a dual-constrained spatio-dynamic ICA model, along with replication analyses, we present canonical brain networks jointly with their canonical dynamic states. We move beyond discrete functional states, allowing continual interactions of changing functional patterns. We next evaluate these dynamic features within a large (N>5K) dataset spanning 4 neuropsychiatric conditions and several clinical high risk data sets. Results highlight the rich potential of expanding the dynamic connectivity repertoire. We find evidence of unique, statistically separable, ‘dynamic fingerprints’ of mental health. Consistent with prior studies, we find reduced spatial dynamic variability and inter-network coupling is linked to neuropsychiatric disorders. These signatures also manifest to varying degrees within clinical high-risk cohorts. It is clear that dynamic interactions play a critical role in illuminating the links between the brain and neuropsychiatric disorders. Capturing time-resolved information can yield intuitive and novel insights into mental illness. Evidence highlights the potential for dynamics to be a major player in characterizing and predicting mental health, including in the context of predicting individuals who might transition to more several mental health conditions.
Habib Ganjgahi - Bayesian Causal Forest on longitudinal data
The rise of large-scale longitudinal neuroimaging studies creates a need for scalable predictive models that handle long-term outcomes, scanner induced non-biological variability and causal inference. Bayesian Additive Regression Trees (BART) are widely used for their scalability, built-in uncertainty estimation, and ability to model complex non-linearity. However, existing BART-based models, assume independent outcomes, making them unsuitable for longitudinal data. we develop a hierarchical model that preserves BART’s strengths while extending it for longitudinal.
Taki Shinohara - Modern neuroimaging data harmonization methods for multi-center studies
While magnetic resonance imaging (MRI) studies are critical for the diagnosis, monitoring, and study of a wide variety of diseases, their use in quantitative analysis can be complex. An increasingly recognized issue involves the differences between MRI scanners that are used in large multi-center studies. To address this, the current state of the art is to "regress out'' or "adjust for'' scanner differences. The field has found these methods to be insufficient, and recent developments are explored to address unique imaging data structures and analytic goals for which specialized harmonization methods are crucial. Available tools are further compared and contrasted for several common neuroimaging data settings.
Trang Cao - Are psychiatric diagnoses associated with a robust neuroanatomical phenotype?
Despite the clinical heterogeneity within psychiatric disorders, a fundamental assumption in psychiatric neuroimaging is that each disorder is associated with some core neural phenotype that should be replicable and consistent across different samples and study locations. If such a consistent phenotype cannot be identified, there may be questionable value in ongoing attempts to examine anatomical group differences between people with psychiatric illness and healthy controls in small, individual studies. Here, we investigated the degree to which five psychiatric disorders--Autism Spectrum Disorder (ASD), Bipolar Disorder (BD), Mood Disorder (MDD), Schizoaffective (SCA), and Schizophrenia (SCZ)--show consistent neuroanatomical phenotypes by examining correlations between disorder- and site-specific maps of cortical thickness alterations.
Michelle Wang - Benchmarking federated learning approaches against siloed and mega-analysis regimes for multi-site analyses
Although neuroimaging is seeing a growing number of datasets, the international adoption of strong data privacy frameworks has led to many of these datasets remaining in so-called “silos”. When data cannot readily be shared, it becomes imperative to develop distributed data processing tools and federated analysis methods to enable large-scale multi-site studies. We present a study comparing a simple federated analysis setup (i.e. sharing only fitted models) with two traditional experimental setups – siloed (no sharing) and mega- (sharing data) analyses. We evaluate the performance of machine learning (ML) models on three neuroimaging datasets of Parkinson’s (PD) and Alzheimer’s disease (AD) on two common prediction tasks in neurodegenerative diseases.
Neda Jahanshad - Challenges and opportunities in large-scale human neuroimaging
Biography: Neda Jahanshad, PhD, is an assistant professor of neurology at the Imaging Genetics Center, part of the Stevens Neuroimaging and Informatics Institute at the University of Southern California. Dr. Jahanshad has over ten years of experience in analyzing brain scans from large groups of individuals, and comparing magnetic resonance imaging (MRI) scans of people with and without brain disorders of all ages. Much of her research explores the brain as a network of interconnected pathways using diffusion-weighted MRI, a technology that reveals physical connections between different functional areas of the brain. Dr. Jahanshad is a key developer for the ENIGMA consortium, which brings together hundreds of researchers around the world. These researchers work together to identify the genetic and environmental risk factors that make the brain susceptible to disease. She holds undergraduate degrees from Johns Hopkins University and a PhD from University of California, Los Angeles.
Chair: Hernando Ombao
Chetan Gohil - Inferring fast dynamic brain networks in electrophysiological data
Electrophysiological recordings (M/EEG) provide a direct measure of neuronal activity at its natural (millisecond) timescale. Recent advances in generative modelling from the field of deep learning allow us to infer fast dynamic functional networks in large datasets. Here, we outline state-of-the-art approaches for dynamic networks analysis in M/EEG.
Ting Ting Zhang - Lifespan Changes in Human Brain Networks
We present a new clustering-enabled regression approach that reveals diverse patterns of lifespan connectivity changes across different regions. We apply our method to diffusion MRI (dMRI) data aggregated from three Human Connectome Project studies. Our analysis reveals that lifespan changes in functional brain networks do not consistently mirror those observed in structural brain networks.
Heather Shappell - A Hidden Semi-Markov Model Approach to State-Based Dynamic Brain Network Analyses
Most functional connectivity (FC) analyses assume that brain FC networks are stationary across time. However, there is interest in studying changes in FC over time. 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.
Anass El-Yaagoubi - Unraveling Information Flow in Brain Networks with Hodge Decomposition
We introduce a modeling framework based on Hodge decomposition to study brain information flow. By breaking down effective connectivity into gradient (causal hierarchy), curl (feedback loops), and harmonic (global integration) components, the model captures distinct patterns of neural communication. Applied to brain signals, it reveals dynamic reorganization of information during brain activity.
Richa Phogat - Unifying Framework for cortico-hippocampal Interactions
Cortico-hippocampal interactions are fundamental to cognitive processes as well as neurological disorders, yet the mechanisms underlying their coordinated dynamics remain poorly understood. Here, we address this knowledge gap by modelling the cortex and the hippocampus as interconnected neural sheets, incorporating their geometric, physiological, and connectivity profiles.