Taming dimensionality in big neuroimaging data: Efficient orthonormal projective NMF via stochastic learning and data compression Authors: Abdalla Bani, Sung Min Ha, Thomas Earnest, Braden Yang, Pan Xiao, John Lee, Janine Bijsterbosch, Aristeidis Sotiras Published: 2026 ABSTRACT AbstractLarge-scale neuroimaging datasets present remarkable opportunities for advancing our understanding of human brain structure and function. Data-driven pattern analysis methods such as Orthonormal Projective Non-negative Matrix Factorization (opNMF) are particularly well suited to uncover multivariate relationships within these data, offering greater interpretability and reproducibility than more conventional approaches such as principal component analysis (PCA) and independent component analysis (ICA). Despite its utility in clinical computational neuroscience, the application of opNMF in large cohort studies has been impeded by computational challenges and scalability limitations. In this work, we address these issues by introducing a stochastic optimization strategy that processes mini-batches of the data, substantially improving scalability. We further accelerate computation through random data compression and leverage repulsive point processes to diversify mini-batches, reducing redundancy and the variance of updates. We first evaluated our method on gray matter tissue density maps from 1,000 participants in the Open Access Series of Imaging Studies (OASIS). Compared with the original approach, it achieved similar approximation accuracy and factor interpretability while greatly reducing computational cost. To demonstrate practical utility, we then applied the framework to 10,000 participants from the UK Biobank, identifying 20 patterns of structural covariance (PSCs) and examined associations between visceral adipose tissue (VAT) and PSC loadings, finding significant relationships for 13 PSCs in females and 11 PSCs in males, most of which were negative. We further show how these patterns refine in higher-rank decompositions with 40 and 60 components. This enhanced opNMF framework opens new possibilities for large-scale neuroimaging analyses, facilitating deeper insights into brain structure in both health and disease. Introduction Advancements in imaging technology and streamlined acquisition workflows have spurred the emergence of numerous large-scale neuroimaging studies of heterogeneous populations. This shift has driven neuroimaging initiatives to broadly supply high-resolution data for thousands of subjects (Glasser et al., 2013; LaMontagne et al., 2019; Sudlow et al., 2015), placing the field firmly within the domain of “Big Data.” The accessibility of these datasets presents opportunities for conducting big-data neuroimaging analyses, with greater accuracy and statistical power. This, in turn, holds the promise of advancing our understanding of brain structure in health and disease. Developing data-driven, hypothesis-free analysis tools with a focus on large-scale studies can complement traditional methods for assessing brain structure, such as anatomical atlases based on regions of interest (ROIs). Unsupervised machine learning techniques such as principal component analysis (PCA) and independent component analysis (ICA) have been commonly employed in neuroimaging studies to derive data-driven representations of brain structure and function. However, these methods rely on complex interactions between components with opposite signs, leading to representations that are difficult to interpret, lack clear regional specificity, and often fail to generalize well to new, unseen datasets (Sotiras et al., 2015). In contrast, non-negative matrix factorization (NMF) (Lee & Seung, 1999) offers an alternative approach by factorizing the data under a non-negativity constraint. This constraint leads to a parts-based representation, where parts combine additively to form a cohesive whole. A variant of the vanilla NMF, known as orthonormal projective NMF (opNMF) (Z. Yang & Oja, 2010), has demonstrated superior interpretability and reproducibility than principal component analysis (PCA) and independent component analysis (ICA) in structural brain analyses (Sotiras et al., 2015, 2017), supporting its increasing use in neuroimaging research. In particular, prior work showed that opNMF yields more localized, additive, and neuroanatomically meaningful components, whereas PCA and ICA tend to produce spatially diffuse patterns with mixed-sign weights (Sotiras et al., 2015). The orthogonality constraint in opNMF further identifies spatially distinct structural networks that are more stable and reproducible across initializations and cohorts (Sotiras et al., 2015). These properties make opNMF especially attractive for neuroimaging studies in which interpretability, regional specificity, and reproducibility are central goals. For example, studies have applied these methods across multiple imaging modalities, including T1- and T2-weighted MRI (Kaczkurkin et al., 2019; Nazeri et al., 2022; Sotiras et al., 2015, 2017), fluid-attenuated inversion recovery (FLAIR) imaging (Habes et al., 2018), diffusion MRI (Bagautdinova et al., 2023; Nazeri et al., 2025; Ouyang et al., 2019), flortaucipir PET imaging (Earnest et al., 2024), functional data (Fürböck, 2020; Ghanbari et al., 2017), cortical thickness covariance analyses (Jirsaraie et al., 2019), and multimodal imaging frameworks (Ochi et al., 2022; Patel et al., 2020; Robert et al., 2022). These applications span diverse populations across the lifespan, including neonates (Nazeri et al., 2022; Ouyang et al., 2019), adolescents and young adults (Jirsaraie et al., 2019; Sotiras et al., 2017), adults (Patel et al., 2020; Robert et al., 2022), and clinical populations with neuropsychiatric and neurodegenerative conditions (Habes et al., 2018; Kaczkurkin et al., 2019; Shui et al., 2024; Varikuti et al., 2018; J. Yang et al., 2024). Importantly, these applications have demonstrated that opNMF-derived representations capture biologically, psychologically, and clinically meaningful variation. In developmental studies, opNMF has identified structural networks aligned with functional systems and evolutionary cortical organization (Sotiras et al., 2017), coordinated patterns of preterm cortical maturation (Ouyang et al., 2019), and reproducible patterns of early postnatal white matter maturation whose developmental trajectories corresponded with microstructural indices and histological features of brain reorganization (Nazeri et al., 2022). In adolescents, NMF-derived structural covariance networks also revealed that greater impulsive choice was associated with reduced cortical thickness in networks encompassing the ventromedial prefrontal and orbitofrontal cortices, temporal pole, and temporoparietal junction, and that these networks predicted delay discounting beyond cognitive performance (Pehlivanova et al., 2018). In multimodal analyses, opNMF has revealed hippocampal microstructural organization associated with demographic and behavioral variation (Patel et al., 2020), as well as striatal microstructural components whose interindividual variability was primarily related to age and sex and further characterized through functional annotation (Robert et al., 2022). Clinical applications have identified regional white matter lesion patterns associated with vascular risk and cognition (Habes et al., 2018), cortical thickness covariance networks associated with irritability and accelerated thinning in frontal and temporal regions involved in emotion regulation (Jirsaraie et al., 2019), and widespread cortical network disruptions associated with posttraumatic stress disorder diagnosis and symptom severity, particularly within prefrontal, cingulate, insular, motor, and sensory-processing regions (J. Yang et al., 2024). In aging and neurodegenerative disease, NMF-derived gray matter representations have supported age prediction across healthy and clinical populations, demonstrated transferability across datasets, and provided localized components that facilitate interpretation of brain-aging patterns (Varikuti et al., 2018). opNMF has also identified Alzheimer’s disease-related atrophy patterns linked to cerebrospinal fluid biomarkers and longitudinal cognitive decline (Shui et al., 2024), as well as reproducible tau covariance patterns in flortaucipir PET data that overlapped with Alzheimer’s disease-related gene-expression maps and were associated with dementia status, neuropsychological performance, and amyloid burden (Earnest et al., 2024). Collectively, these findings establish opNMF as a robust and well-validated framework for deriving biologically, psychologically, and clinically meaningful representations. Large-scale neuroimaging has also motivated scalable adaptations of methods other than opNMF, particularly in the fMRI literature. For example, Multivariate Exploratory Linear Optimized Decomposition into Independent Components’ (MELODIC) Incremental Group-PCA (MIGP) (Smith et al., 2014) was developed to approximate group PCA for very large fMRI datasets by updating a low-rank PCA representation incrementally across subjects, thereby reducing the memory burden of full temporal concatenation. Multi-power iteration PCA (MPOWIT) (Rachakonda et al., 2016) addresses a similar scalability problem from a different angle, using memory-efficient subspace iteration to estimate the dominant PCA space for large group-ICA workflows. In contrast, neighborhood preserving embedding (NPE) (Zhao et al., 2022) methods approach dimensionality reduction as a nonlinear manifold-learning problem, seeking to preserve local neighborhood structure in the reduced representation rather than recover a PCA subspace. These methods illustrate that scalable neuroimaging analysis can be achieved through several distinct strategies, including incremental PCA, iterative subspace estimation, and manifold-based embedding. However, they differ fundamentally from opNMF. Both MIGP and MPOWIT are designed to scale the PCA-based reduction and, therefore, optimize for maximal variance, yielding mixed-signed, spatially overlapping components that can limit interpretability. In contrast, opNMF enforces a non-negativity constraint to produce a “parts-based” representation, ensuring that the resulting structural networks are purely additive and spatially localized. Stochastic versions of opNMF, which process mini-batches of data, offer a potential solution for managing large datasets by reducing memory demands (Wen et al., 2023; Z. Yang et al., 2012). However, these variants often lack theoretical convergence guarantees and typically rely on random sampling, which can introduce redundancy and bias due to non-representative mini-batches (LeWinn et al., 2017; Sankar et al., 2016). Another computational strategy involves data compression, which accelerates updates in traditional NMF by projecting the data onto a lower-dimensional subspace that retains most of the critical information (Tepper & Sapiro, 2016). To identify this subspace, compression methods utilize structured random projections (Halko et al., 2011). The compression procedure entails obtaining the QR decomposition of a matrix product. In large-scale data settings, the matrix product becomes too extensive to fit into memory, necessitating advanced parallel computing architecture to execute the QR decomposition (tall-and-skinny-QR [TSQR]) (Tepper & Sapiro, 2016). To address the above limitations and facilitate the application of opNMF in large-scale studies, we propose an algorithm specifically designed to scale efficiently to big neuroimaging datasets. Building on prior algorithmic solutions for stochastic and compressed NMF, our approach integrates the advantages of both paradigms into a unified framework. The algorithm adopts a stochastic learning strategy that performs opNMF updates on compressed mini-batches of data. We refer to this method as compressed stochastic opNMF (csopNMF). The stochastic component alleviates memory constraints by processing data in small batches, while the compression accelerates updates, thereby reducing run-time requirements (Fig. 1). To further enhance robustness, we introduce an alternative sampling strategy within the stochastic learning framework. Specifically, we incorporate repulsive sampling based on determinantal point processes (DPPs) (Kulesza & Taskar, 2012), which encourage diversity among selected samples and yield mini-batches that better represent the population. DPP sampling offers two key advantages over random sampling: (1) it can incorporate covariate information (e.g., age, sex, pathology status) to promote balanced, representative sampling and mitigate bias (Bani et al., 2023; Zhang et al., 2017) and (2) by discouraging redundant sample selection, it reduces the variance of gradient updates, enabling faster and more stable convergence. Finally, our framework provides a practical solution to the big-data QR decomposition bottleneck encountered in traditional compression algorithms. Because the method operates stochastically on mini-batches, QR decomposition is performed on smaller data subsets—eliminating the need for complex distributed computing architectures such as tall-and-skinny QR (TSQR). Theoretical RAM and compute-time scaling for opNMF and csopNMF across decomposition ranks k and increasing sample size. Here, k denotes the factorization rank, that is, the number of components to be estimated. Curves were generated assuming 1,200,000 voxels per subject, consistent with volumetric data registered to MNI-152 space, and subject counts ranging from 1,000 to 100,000. For csopNMF, calculations assume a fixed compressed mini-batch size of 100, so the matrix processed at each update remains constant as the total dataset size grows. Consequently, the per-update memory and computational requirements of csopNMF are approximately independent of the total number of subjects, whereas those of opNMF increase with the full sample size. Both panels are shown on a logarithmic scale to facilitate comparison across the wide range of values. Population-scale neuroimaging enables multivariate mapping of brain organization, and the UK Biobank (UKB) provides an ideal setting to do so because of its large imaging cohort with harmonized acquisition and deep phenotyping. In this work, we used UKB as a testbed to move beyond region-wise summaries and localize population patterns with csopNMF and investigate their hierarchical organization in higher-order models. We then ask how those patterns relate to visceral adiposity. A growing literature links adiposity to brain structure. In a population sample, higher total body fat was associated with smaller gray matter volume, alterations in white matter microstructure, and subcortical differences, underscoring a brain–body connection observable at scale (Dekkers et al., 2019). More recently, a midlife cohort study (N=10,001) showed that both visceral and subcutaneous abdominal fat predicted lower brain volumes, with widespread lobar associations and indications of sex differences; these effects were observed even after normalization to intracranial volume (Raji et al., 2024). However, prior analyses of adiposity and brain structure have largely relied on region- or lobe-based reporting schemes. Such predefined anatomical units can obscure distributed and overlapping effects, impose arbitrary boundaries, and fail to capture coordinated structural variation across the brain. Here, we extend adiposity–brain analyses in the UK Biobank using data-driven csopNMF-derived patterns of structural covariance. By testing pattern loadings against visceral adipose tissue, we provide an anatomically interpretable, population-scale map of VAT–brain relationships that complements and overcomes key limitations of region- and lobe-based approaches. Note that a preliminary version of this work was presented in Bani et al. (2023). The present work extends the previous publication in multiple ways: We proposed an efficient and scalable opNMF variant called csopNMF, which leverages random compression and stochastic learning in a unified framework to overcome opNMF’s computational challenges.We enhanced csopNMF by incorporating a diverse sampling approach to reduce update variances.We conducted an extensive performance comparison of csopNMF with both stochastic opNMF (sopNMF) and standard opNMF as baselines. Our experiments spanned two different batch sizes and three different compression scenarios providing a broad evaluation of the algorithms’ performance.We further extended the evaluation by thoroughly assessing not only the computational accuracy but also the computational time and memory.Furthermore, we applied csopNMF factors in a typical downstream machine learning task (age prediction) and analyzed its performance across various setups against opNMF and sopNMF. The results demonstrated that our proposed method achieved comparable accuracy with the baseline opNMF while significantly reducing computational cost.We demonstrated population-scale utility by applying csopNMF to 10,000 UK Biobank participants to derive 20 components, referred to as patterns of structural covariance (PSCs), and testing brain–body links via associations between PSC loadings and VAT.Finally, we extended the analysis to higher model orders of 40 and 60 components to illustrate how broader structural covariance patterns resolve into finer-grained organization. We proposed an efficient and scalable opNMF variant called csopNMF, which leverages random compression and stochastic learning in a unified framework to overcome opNMF’s computational challenges. We proposed an efficient and scalable opNMF variant called csopNMF, which leverages random compression and stochastic learning in a unified framework to overcome opNMF’s computational challenges. We enhanced csopNMF by incorporating a diverse sampling approach to reduce update variances. We enhanced csopNMF by incorporating a diverse sampling approach to reduce update variances. We conducted an extensive performance comparison of csopNMF with both stochastic opNMF (sopNMF) and standard opNMF as baselines. Our experiments spanned two different batch sizes and three different compression scenarios providing a broad evaluation of the algorithms’ performance. We conducted an extensive performance comparison of csopNMF with both stochastic opNMF (sopNMF) and standard opNMF as baselines. Our experiments spanned two different batch sizes and three different compression scenarios providing a broad evaluation of the algorithms’ performance. We further extended the evaluation by thoroughly assessing not only the computational accuracy but also the computational time and memory. We further extended the evaluation by thoroughly assessing not only the computational accuracy but also the computational time and memory. Furthermore, we applied csopNMF factors in a typical downstream machine learning task (age prediction) and analyzed its performance across various setups against opNMF and sopNMF. The results demonstrated that our proposed method achieved comparable accuracy with the baseline opNMF while significantly reducing computational cost. Furthermore, we applied csopNMF factors in a typical downstream machine learning task (age prediction) and analyzed its performance across various setups against opNMF and sopNMF. The results demonstrated that our proposed method achieved comparable accuracy with the baseline opNMF while significantly reducing computational cost. We demonstrated population-scale utility by applying csopNMF to 10,000 UK Biobank participants to derive 20 components, referred to as patterns of structural covariance (PSCs), and testing brain–body links via associations between PSC loadings and VAT. We demonstrated population-scale utility by applying csopNMF to 10,000 UK Biobank participants to derive 20 components, referred to as patterns of structural covariance (PSCs), and testing brain–body links via associations between PSC loadings and VAT. Finally, we extended the analysis to higher model orders of 40 and 60 components to illustrate how broader structural covariance patterns resolve into finer-grained organization. Finally, we extended the analysis to higher model orders of 40 and 60 components to illustrate how broader structural covariance patterns resolve into finer-grained organization. The remainder of this paper is structured as follows. In section 2, we define the csopNMF framework, beginning with opNMF and introducing its stochastic and compressed variants. This section also describes the datasets used and outlines the validation and application experiments along with their parameter settings. In section 3, we present the findings from both the validation and application analyses. In section 4, we summarize the main results, limitations, and conclusions of our work. Materials and Methods Proposed methodology In this section, we first provide a brief overview of opNMF and outline the key challenges in its optimization. Then, we describe our proposed solution, including the stochastic optimization scheme, mini-batch diversification approach, and compression framework. An outline of the proposed approach is shown in Figure 2. Overview of the proposed approach. (a) (1) A large-scale neuroimaging study of interest is selected for analysis. (2) Subjects meeting the inclusion criteria are identified, and their T1-weighted (T1w) images are processed using standard neuroimaging tools. (b) (3) Mini-batches are selected from the full sample either stochastically or using determinantal point processes (DPP) to promote diversity. (4) Images within each batch are vectorized and arrayed column wise to form mini-batch data matrices. (5) Mini-batches are compressed using structured random projections. (6) csopNMF is performed on compressed mini-batches, and model updates are accumulated across batches. Orthonormal projective non-negative matrix factorization (opNMF) Let x1,x2,...,xn∈ℝ+m be vectorized images of n different subjects (or samples) acquired through medical imaging and m be the number of non-negative features (voxels) of each image. The data matrix X is then constructed by arraying these images column wise to obtain X∈ℝ+m×n , equivalently X=[x1,x2,...,xn], xi​ ∈ℝ+m. Our goal is to decompose the data matrix X into a user-specified number of components k that group features that covary consistently across the subjects. Toward this goal, we aim to solve the orthogonal non-negative matrix factorization problem (Yuan & Oja, 2005): ℒ=minW≥0‖X−WWTX‖F2 s.t. WTW=I, (1) where ℒ is the loss, W​ ∈ℝ+m×k is the components matrix, ∥·∥F is the Frobenius norm, and I denotes the identity matrix. The previous minimization problem can be solved using the iterative multiplicative update (MU) rule (Z. Yang & Oja, 2010; Yuan & Oja, 2005) Wt+1←Wt(XXTW)t(WWTXXTW)t. (2) In contrast to the traditional NMF approach (Lee & Seung, 1999), which decomposes the data matrix X into non-negatively constrained components W and coefficients H (min‖X−WH‖F2), opNMF takes a different route by estimating the loading coefficients H​ ∈ℝ+k×n as a projection of the data onto approximated orthogonal components H=WTX . The introduced orthogonality and projectivity constraints enhance spatial specificity, reducing overlap between components and facilitating interpretability (Sotiras et al., 2015, 2017). However, challenges arise in high-dimensional scenarios, such as in MRI with millions of features. The covariance product XXT appearing in the numerator and denominator in Eq. (2) becomes too large to reside in memory. To address this, opNMF implementations have employed a memory-efficient update rule, which changes the order of multiplication and performs the calculation X(XTW) in each iteration. Nevertheless, this strategy encounters further complications in large sample size studies, where the computational expense of computing the entire product XXTW in Eq. (2) becomes impractical. Consequently, these challenges render opNMF impractical for handling large neuroimaging datasets. Compressed stochastic opNMF (csopNMF) Next, we will address the computational bottlenecks of opNMF by presenting our solution. We will begin by discussing the stochastic learning scheme, and then we will build upon this framework by incorporating mini-batch compression through structured random projections. Stochastic opNMF (sopNMF) sopNMF is a variant of opNMF that aims to stochastically approximate the solution of opNMF over mini-batches of the data. By adopting the stochastic approach, we mitigate the impact of the costly term XXTW that appears twice in Eq. (2) by replacing it with XpXpTW , where Xp ∈ℝ+m×p (p≪n ) is a small matrix constructed by randomly sampling subsets of the columns of the data matrix X (Z. Yang et al., 2012). The multiplicative update (MU) rule in Eq. (2) can be rewritten as Wt+1←Wt(XpXpTW)t(WWTXpXpTW)t. (3) The update rule in Eq. (3) can be thought of as the solution to the sequence of surrogate loss functions that minimize the opNMF objective over the mini-batches. fn(W)≜1B∑i=1Bℓ(Xi,W), (4) where ℓ (Xi,W) is a loss function. For example, in the case of a single batch, we can write ℓ (Xp,W) as l(Xp,W)≜minW≥0Xp−WWTXpF 2.. In machine learning applications, we usually seek to minimize the expected cost f(W)≜Ex[ℓ(X,W)]= limn→∞fn(W) a.s. , where the expectation E is taken relative to the probability distribution of the data p(x) (Bottou, 2010; Mairal et al., 2010), instead of minimizing the empirical cost fn(W) with high accuracy. Considering a dataset composed of subjects from a distribution p(x), we draw a mini-batch Xi and update Wt+1 by minimizing the function f^B(W)≜1B∑i=1B(‖Xi−WWTXi‖F2). Proposition 1 (Bani et al., 2023): The sequence of surrogate objective functions converges almost surely. f^t acts as a converging surrogate of f, and f(Wt)−f^B(Wt)→a.s.0 and hence f(Wt)→a.s.0 . This means that f^ is a surrogate and upper-bounds the empirical cost. Given that f^ acts as a surrogate of f, we can iterate over mini-batches of the data. Diversified sopNMF (dsopNMF) Random sampling may lead to batches that contain redundant data that do not represent the population well, which may lead to biased results (Kulesza & Taskar, 2012; Zhang et al., 2017). DPPs are a repulsive sampling technique that aim to obtain samples that are as diverse as possible by encoding negative correlations (Kulesza & Taskar, 2012). The dissimilarity between the samples is measured using a real, symmetric, positive semi-definite (PSD) kernel L∈ℝn×n , where each element Lij is a measure of dissimilarity between samples i and j. In a DPP, subsampling any subset X of {1, …, n} is proportional to the determinant of the submatrix that indexes the subset LX of L: P(X)=det(LX)det(L+I)∝det(LX). We use κ-DPPs (Kulesza & Taskar, 2011) and condition the inclusion probability on the size of the subset P Lk (X)=det(LX)∑|X′|det(LX′), where |X|=p (in keeping with convention, and at the risk of abuse of notation, we use “κ” to refer to the number of subsets generated by the κ-DPPs sampling scheme). In our experiments, we create a linear kernel designed to introduce diversity among the batches, considering the similarities among the images of subjects in the study Lij​=xiTxj. sopNMF and dsopNMF are shown in Algorithm 1. Algorithm 1: sopNMF and dsopNMFData: data matrix X∈ℝ+m×n , target rank k∈ℕ+, batch size p<