Hemispheric multi-dimension features extraction analysis based on decoupled representation learning

Abstract

The predominant approach in investigating brain structural asymmetry relies on predefined regions of interest, assessing variations between homologous brain regions through a single indicator, which is local, univariate, and relative. In response to this challenge, we employ decoupled representation learning from deep learning to extract hidden features containing hemisphere-specific information at a hemispheric systemic level. This novel approach enables a global and multivariate analysis of brain structural asymmetry. Our findings indicate a significant association between left-hemisphere-specific hidden features and language-related behavioral metrics, as well as a correlation between right-hemisphere-specific hidden features and social-related behavioral metrics. Tensor-based Morphometry results find the impact of left-hemisphere-specific features on the left inferior frontal sulcus within Brocas area, a crucial region for language processing. Additionally, right-hemisphere-specific features influenced the right rostral hippocampus, a region implicated in emotion regulation and spatial navigation. The findings from Neurosynth indicate that significant regions caused by left-hemisphere-specific features are correlated with language, while significant regions caused by right-hemisphere-specific features are associated with behaviors primarily governed by the right hemisphere. Furthermore, our study establishes a link between structural changes induced by hemisphere-specific features and several genes. Such findings demonstrate that the application of deep learning techniques allows for precise capture of hemisphere-specific information within individual hemispheres, offering a new perspective for future research on brain structural asymmetry.