CSVR: Combined Surface and Volume Registration for Neonatal Brain MRI

Abstract

Nonlinear image registration is a cornerstone of neuroimaging analysis, supporting both qualitative and quantitative comparisons of brain structures across individuals and over time. While traditional volumetric registration methods, driven by voxel intensities, achieve good alignment of subcortical regions, they generally fail to capture correspondences between highly convoluted and variable cortical shapes. Surface-based methods, which instead regularise mappings as geodesics along the cortical sheet, yield improved cortical alignment but ignore the subcortical domain, limiting their utility for whole-brain analyses. A unified registration framework would address these limitations to enable integrated analysis of cortical and subcortical structures and the neuronal fibres that connect them. However, achieving this is challenging, since matching heterogeneous cortical shapes implies large volumetric displacements local to the cortex. To overcome these challenges, we introduce CSVR, the first deep learning-based framework for combined surface–volume registration of neonatal MRI. By integrating hierarchical registration strategies with discrete optimisation, CSVR achieves accurate, smooth, and anatomically plausible alignment of the entire brain.