Contributed Poster Presentations: Section on Statistics in Imaging

Magnetic resonance imaging (MRI) data is frequently used to monitor brain regions for the effects of neurodegenerative conditions like Alzheimer's disease (AD). AD exhibits substantial heterogeneity between patients, with this variability frequently described using disease subtypes defined by distinct pathological characteristics. Thus, MRI data from studies of AD often has a nested structure, with several images collected for each participant, who in turn are grouped by disease subtype. Statistical methods that do not account for this structure may be unable to fully capture the relationships present in the data. To address this problem, we adapt the principal manifold estimation algorithm using an additive spline model to obtain manifold estimates of brain region structure that vary at each level of a nested hierarchy. A hypothesis testing framework allows testing for significant differences between group- or individual-level manifold estimates. The proposed method is compared to existing approaches using simulated data and applied to estimate the surfaces of hippocampi of participants in the Alzheimer's Disease Neuroimaging Initiative study. 

Functional magnetic resonance imaging (fMRI) plays a crucial role in investigating the human brain's responses to stimuli. Addressing a key challenge that individuals differ in their brain's underlying functional topography, the Hyperalignment method aligns functional brain representations across individuals by projecting neural responses into a shared high-dimensional space. Through the orthogonal Procrustes transformation, one can search on the Stiefel manifold and determine the optimal rotation for aligning images to a common template. However, unconstrained optimization disregards natural topological structure and may result in highly variable loadings for adjacent brain locations in the estimated projection, potentially causing location flips for distant voxels in extreme cases. To address this issue, we propose a spatially-aware Hyperalignment model that incorporates penalties to encourage smoothness in the projection. By constraining nearby voxel loadings, our approach restricts the search space on the Stiefel manifold leading to a more spatially coherent alignment. We illustrate the benefits of the methods through simulations and application to data from fMRI experiments. 

Understanding how functional magnetic resonance imaging (fMRI)-based functional connectivity underlies complex cognitive functions is crucial for identifying disease-related changes in cognition. Traditional approaches to quantifying functional connectivity (FC), such as Pearson's correlation and Tikhonov regularization, have been used to predict fluid and crystallized cognitive scores with limited success. Recent advancements have leveraged latent space representations to enhance prediction, but at the cost of FC network interpretability. We propose using graphical LASSO with a cross-validated tuning parameter to measure FC networks. Graphical LASSO estimates the inverse covariance matrix under a multivariate normal model by maximizing the l1-penalized log-likelihood, providing a sparse solution in high-dimensional settings. We demonstrate that this approach improves the prediction of individual crystallized, fluid and total cognitive scores in 997 healthy young adults from the Human Connectome Project. Results are compared to Pearson's correlation and Tikhonov regularization across different high-dimensional grey matter parcellation choices. 

Sub-visible particles (ranging from 2-100 µm in size) within liquid injectable sterile biopharmaceutical formulations require quantification and analysis as a critical component of formulation development. Micro-flow imaging (MFI) enables high-resolution visualization of sub-visible particles, allowing scientists to develop sophisticated classification models to distinguish between high-risk proteinaceous particles and low-risk non-proteinaceous particles, such as silicone oil droplets. Previous work has demonstrated that deep learning models for image classification can enable rapid, automated classification of sub-visible particles in MFI. In this study, we enhance existing models developed for this purpose by enriching the annotated dataset, fine-tuning a pre-trained model architecture, and adjusting the training regime to improve model generalizability. We demonstrate that our new model, SVRNet, shows improved classification performance over existing methods and provides superior assessment of sub-visible particles.