Search Abstracts | Symposia | Slide Sessions | Poster Sessions
A probabilistic, network-based approach to voxel-based morphometry
Poster Session A, Wednesday, September 30, 11:00 am - 1:00 pm, Wangari Maathai
G. Lynn Kurteff1, Grant M. Walker2, Maria Luisa Mandelli1, Siddarth Ramkrishnan1, Zachary A. Miller1, Howard J. Rosen1, William W. Seleey1, Maria Luisa Gorno-Tempini1; 1University of California, San Francisco, 2University of California, Irvine
INTRODUCTION – Mapping brain-behavior relationships in complex datasets poses a challenge for statistical interpretability and generalization of results. The issue is compounded in neurodegenerative disease, where there are additional challenges concerning clinical and anatomical heterogeneity. Furthermore, tissue cannot be easily binarized as impaired or preserved as it can in stroke, and additional steps must be taken to quantify the extent of damage. Conventional Voxel-based morphometry (VBM), which typically relies on mass-univariate voxelwise general linear models (GLMs) to assess brain-behavior relationships, requires extensive multiple-comparison correction and often produces summary maps that may not fully capture phenotypic diversity across patients. Here, we introduce a probabilistic, network-based framework designed to improve interpretability and characterize heterogeneous brain-behavior relationships in neurodegenerative disease. METHODS – Critical Network VBM (CN-VBM) combines low-dimensional regression and cross-validated out-of-sample prediction to identify distributed brain networks associated with specific behavioral features. Rather than producing a single “optimal” map, CN-VBM uses an ensemble-based framework to characterize multiple plausible brain-behavior relationships and better capture phenotypic heterogeneity in neurodegenerative disease. The multiple comparisons problem of VBM is simplified through region-based, rather than voxel-based, statistical testing. In this study, we validated CN-VBM by modeling brain networks associated with core speech-language features in primary progressive aphasia (PPA), and comparing these results with conventional VBM in both real and simulated neurodegenerative datasets. RESULTS – Across the three primary subtypes of PPA (semantic, nonfluent, and logopenic), CN-VBM identified distinct network-level patterns associated with specific speech-language impairments. For example, in nfvPPA, CN-VBM dissociated networks associated with apraxia of speech, dysarthria, and agrammatism: apraxia of speech-related atrophy localized to inferior precentral and anterior insular cortex, dysarthria-related atrophy localized to dorsal precentral and medial frontal cortex, and agrammatism-related atrophy spanned a broader perisylvian network. CONCLUSION – CN-VBM provides a complementary framework for neurodegenerative brain-behavior mapping that emphasizes interpretability and network-level characterization. We believe an ensemble-based approach such as CN-VBM has utility in explaining the phenotypic heterogeneity of language impairments within each variant of PPA. An open-source Python implementation of CN-VBM is available on GitHub.
Topic Areas: Methods, Disorders: Acquired