Penn Statistics in Imaging and Visualization Endeavor (PennSIVE)

Population reference ranges accounting for age, sex, and other covariates are increasingly important for identifying clinically meaningful outliers, and the World Health Organization recommends GAMLSS for this task because of its flexibility in modeling complex, non-linear trajectories for a wide range of outcome distributions. However, traditional GAMLSS requires large, centralized datasets, which are often impractical due to privacy and data-sharing constraints. We developed distributed GAMLSS (dGAMLSS), a new privacy-preserving method that enables population reference chart modeling across multiple institutions without sharing patient-level data to address this problem. dGAMLSS overcomes this barrier by allowing sites to collaboratively fit GAMLSS models through federated learning, while keeping data local. The method includes tailored strategies for fitting smooth functions with different communication-efficiency tradeoffs. We demonstrate that dGAMLSS can accurately construct population reference charts across diverse domains, including clinical, microbiome, and neuroimaging applications, opening the door to broader, more generalizable normative modeling.

Fengling Hu, Jiayi Tong, Margaret Gardner …
DOI: 10.1093

New 2024 McDonald diagnostic criteria rely on white matter lesions as well as emerging markers such as the central vein sign (CVS) and paramagnetic rim lesions (PRLs), which have been shown to substantially improve specificity but are difficult to segment and categorize manually at scale. We introduced ALPaCA (Automated Lesion, PRL, and CVS Analysis), a fully automated MRI-based method for jointly segmenting and identifying multiple sclerosis lesions and MS-specific imaging biomarkers. ALPaCA combines voxel-wise lesion segmentation with a multi-contrast, multi-label 3D convolutional neural network to classify lesions, PRLs, and CVS in a single framework. Trained and evaluated using cross-validation in adults with MS and radiological mimics, ALPaCA significantly outperforms existing open-source, fully-automated approaches at both the lesion and subject level. Automated lesion and PRL scores correlate with expert manual counts and show associations with MS diagnosis. Together, these results demonstrate that ALPaCA enables scalable assessment of MS-specific imaging biomarkers, providing a new tool to MS researchers and clinicians.

Fengling Hu, Zheng Ren, Luyun Chen …
DOI: 10.1162

Estimating longitudinal brain biomarker trajectories is challenging due to biological variability, inconsistent measurement protocols, and sparse data. We present a personalized deep kernel regression framework that combines a population-level model with a subject-specific model, balanced through Adaptive Shrinkage Estimation, to forecast regional brain volumetric measurements. Our approach outperforms state-of-the-art methods in predictive accuracy and uncertainty quantification, and demonstrates robustness in external neuroimaging studies.

Vasiliki Tassopoulou, Haochang Shou, Christos Davatzikos …
Publication Link

In neuroimaging research, it is oftentimes of interest to interpret brain maps of associations between imaging metrics and behavioral outcomes through the lens of anatomical parcellations or functional networks. For instance, one might want to know whether the association between cortical thickness and age is stronger (i.e., enriched) in dorsolateral prefrontal cortex than the rest of the brain.  To address this question, we developed a permutation-based statistical framework called Network Enrichment Testing using Ordinal Dominance Curves (NETDOM). Notably, NETDOM controls the type I error rate – unlike other potential methods for network enrichment testing – while also exhibiting greater statistical power when enrichment occurs at only a subset of the in-network locations.

Noah Hillman, Sarah M. Weinstein, Joëlle Bagautdinova …
DOI: 10.1002

We investigated how genetic factors influence Alzheimer’s disease (AD)-like brain atrophy through physical activity as a potential mediator using data from 13,425 UK Biobank participants. The study integrated genetic variants with structural MRI volumetric data and multivariate physical activity patterns extracted from wearable device sensors via high-dimensional Bayesian mediation analysis. 22 mediation pathways were identified through which genetic risk variants affect AD-like brain atrophy by altering physical activity. These findings advance our understanding of AD pathogenic mechanisms and highlight the value of integrating multi-omics data with high-dimensional mediation methods.

Hanxiang Xu, Shizhuo Mu, Jingxuan Bao, Christos Davatzikos, Haochang Shou …
DOI: 10.1186

We developed a new method called LaxKAT (linear maximum kernel association test) that can identify both global and local signal in high-dimensional data, including neuroimaging data. The LaxKAT statistic maximizes the sequence kernel association test (SKAT) statistic over a pre-specified subspace of linear kernels. We demonstrated via simulations that it exhibits improved global and local power compared to previous methods while controlling the family-wise error rate (FWER). We also applied LaxKAT to neuroimaging data from cognitively normal controls in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) to identify brain regions exhibiting sex-specific differences in cortical thickness patterns.

Christina Chen, Jeremy Rubin, Lior Rennert…
DOI: 10.1080

In neuroimaging research, volumetric data contribute valuable information for understanding brain changes during both healthy aging and pathological processes. Extracting these measures from images requires segmenting the regions of interest (ROIs), and many popular methods accomplish this by fusing labels from multiple expert‐segmented images called atlases. We proposed a novel method that estimates the precision of the measured ROI volume for each subject due to the multi‐atlas segmentation procedure. We demonstrated in real data that weighting by these estimates markedly improves the power to detect a mean difference in hippocampal volume between controls and subjects with mild cognitive impairment or Alzheimer’s disease.

Christina Chen, Sandhitsu Das, M Dylan Tisdall…
DOI: 10.1002