DeepVBM: A fully automatic and efficient voxel-based morphometry via deep learning-based segmentation and...

Voxel-based morphometry (VBM) using T1-weighted magnetic resonance imaging is a pivotal tool for assessing brain structure and identifying subtle morphological changes associated with various neurological conditions. Conventional VBM workflows, however, face significant computational challenges, particularly during the nonlinear deformable registration stage, which impedes analysis of large-scale neuroimaging databases. In this study, we introduce FuseMorph, a deep learning-based registration method that refines initial zero-shot predictions from a pretrained model via iterative inference and targeted parameter search. By eliminating the need for full backpropagation and additional model retraining, FuseMorph significantly reduces computational demands, achieving registration accuracy comparable to state-of-the-art methods even in CPU-only environments. FuseMorph is integrated into DeepVBM, a fully automated VBM pipeline streamlines the processing of high-resolution MRI datasets and substantially reduces computation time compared to traditional pipelines, thereby facilitating the efficient analysis of large multi-center studies. The proposed approach was validated on multiple datasets, including an Alzheimer's disease cohort where DeepVBM successfully detected characteristic patterns of gray matter atrophy in regions such as the hippocampus, entorhinal cortex, and amygdala. These findings not only underscore the clinical relevance of the method but also demonstrate its potential for early detection and monitoring of neurodegenerative changes. This work contributes an accessible, efficient, and scalable solution for neuroimaging research, with potential applications extending to various neurodegenerative disease studies.