Modern Statistical Methods for Advancing Biomedical and Neuroimaging Research

The brain is a high-dimensional object. Graphical models and network analysis methods are promising tools to characterize complex brain structures at different disease stages. Community detection is particularly valuable to enhance our understanding of the intrinsic connections between different brain modules. In this study, we introduce a new method that integrates brain topology and crucial node-level information for community detection. The estimation is done in a Bayesian framework, where a scalable algorithm based on variational Bayes is proposed. Extensive simulations are conducted to evaluate the performance of the proposed method and related algorithms. The method is applied to brain networks generated from functional magnetic resonance imaging (fMRI) data of Alzheimer's disease for a case study. 

The brain is a network of interconnected neural circuits. Numerous studies show that neural circuits are key to understanding brain function, development and aging, and neuropsychiatric disorders. However, delineating neural circuits from complex, high-dimensional brain networks is challenging. Blind source separation (BSS) offers a powerful, data-driven approach to uncover neural circuits from brain data, with each latent source corresponding to one circuit. Recent advances in BSS allow direct decomposition of connectivity data instead of raw imaging, yielding new insights into the brain connectome. However, a crucial limitation is that these methods are typically developed without integrating any domain knowledge such as neuroanatomical information. While their status as independent statistical tools demonstrates statistical effectiveness, they miss the opportunity to reveal more reliable scientific findings. To address this gap, we propose a Bayesian hierarchical decomposition model with a novel domain-knowledge-informed prior for brain connectivity data. Our method further enables joint analysis of data from multiple subject groups, facilitating the identification of differences in neural circuits across groups with varying covariates, such as age or sex. Applied to functional connectivity data from the Lifespan Human Connectome Project in Development (HCP-D) study, our method identifies several scientific meaningful and reliable neural circuits, along with age- and sex-related variations within them. Our method also uncovers novel neural circuits that have not been previously reported, offering new insights into the development of the functional connectome.  

Generative artificial intelligence (AI) has transformed biomedical imaging through synthetic image creation, addressing challenges of data availability, privacy, and dataset diversity in medical applications. In this paper, we propose a novel nonparametric method within the functional data framework to estimate signals and discern significant differences between the mean functions of real and synthetic biomedical imaging data, thereby enhancing model training. Focusing on surface-based synthetic imaging data, our approach employs spherical splines on triangulations to address spatial heterogeneity. A key innovation is our method for constructing simultaneous confidence corridors (SCCs), which quantify uncertainty in real-synthetic data differences and guide AI model refinement. We establish the asymptotic properties of our estimators and SCCs, proving that the SCCs are asymptotically equivalent to those constructed from noise-free images. Numerical studies validate the SCCs' coverage probability and demonstrate and our method's effectiveness in improving AI-driven image synthesis when integrated into generative AI training. The effectiveness of this approach is demonstrated through an application comparing original and synthetic cerebral spinal fluid functional MRI (cs-fMRI) data from the Human Connectome Project. 

Deep learning models have demonstrated strong predictive performance in neuroimaging tasks, but their "black box" nature often limits their utility in scientific interpretation and statistical inference. In this talk, I will present our recent developments in interpretable AI models tailored for functional MRI (fMRI) time series data. Our approach combines high predictive accuracy with mechanisms for identifying clinically relevant brain regions at multiple levels of analysis. We demonstrate that the identified region importance is robust across datasets and experimental conditions, offering a promising path toward more transparent and reliable use of AI in neuroimaging research. 

Functional connectivity (FC) provides insights into multiple psychiatric disorders, yet the substantial inter-subject variability of FC hinders its effectiveness in distinguishing between various psychiatric disorders. Therefore, we propose a novel graph learning framework that integrates FC with behavioral characteristics to better differentiate between psychiatric disorders. Additionally, applying Grad-CAM enhances model interpretability by identifying key regions of interest (ROIs) involved in distinguishing individuals with psychiatric disorders from those without. Preliminary experiments using the ABCD dataset revealed two key findings: first, critical regions including the thalamus, putamen, and pallidum, along with nodes from the somatomotor and cingulo-opercular networks, are essential for distinguishing psychiatric disorders. Additionally, visualization of latent representations indicated that individuals with externalizing disorders, specifically Oppositional Defiant Disorder (ODD), are notably distinguishable from healthy controls. These results highlight the potential of our graph learning framework in identifying psychiatric disorders.