Cross-modal structural residuals reveal cognitive impairment in cerebral small vessel disease

Abstract

Cerebral small vessel disease (CeVD) is a major cause of vascular cognitive impairment and dementia (VCID), but the relationship between structural brain alterations and heterogeneity in cognitive outcome remains incompletely understood. Existing neuroimaging studies primarily focus on discriminative imaging features, while the correspondence between different structural representations within individuals has received less attention. In this study, we developed a deep cross-modal structural correspondence framework to model the relationship between macrostructural features derived from T1-weighted MRI and microstructural features derived from diffusion tensor imaging (DTI). The model was trained exclusively on cognitively normal CeVD subjects to establish a reference pattern of structural correspondence. Structural residuals were computed as the deviation between predicted and observed DTI representations. The structural residual demonstrated stable performance for identifying VCID from CeVD, achieving an average area under the receiver operating characteristic curves of 0.96. Residual estimates showed high intra-subject consistency. In contrast, models trained using shuffled T1-DTI pairings exhibited reduced stability and diagnostic performance, supporting the importance of preserved within-subject cross-modal correspondence. Higher structural residuals were significantly associated with poorer cognitive performance across global cognition, memory, and executive function domains. Incorporating the structural residual into a multimodal classifier further improved probability separation and clinical net benefit. Model interpretability analyses revealed that posterior cortical regions, particularly occipital and precuneus regions, contributed more strongly to residual computation in VCID. These findings suggest that altered macrostructural-microstructural correspondence may represent an imaging marker associated with cognitive impairment in CeVD. Modelling cross-modal structural correspondence provides a complementary framework for characterizing disease heterogeneity in VCID.