Hemispheric Multidimensional Feature Extraction Analysis Based on Decoupled Representation Learning

2108 

Abstract Submission 

Yuwei Su¹, Sifeng Wang¹, Suyu Zhong¹

¹School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing, China

Yuwei Su  
School of Artificial Intelligence, Beijing University of Posts and Telecommunications
Beijing, China

Sifeng Wang  
School of Artificial Intelligence, Beijing University of Posts and Telecommunications
Beijing, China

Suyu Zhong  
School of Artificial Intelligence, Beijing University of Posts and Telecommunications
Beijing, China

Human brain exhibits asymmetry in both structure and function. The conventional method for studying structural MRI asymmetry entails subtracting a certain metric between homologous brain regions to obtain asymmetry index [1, 2, 3]. However, this approach confines comparisons between corresponding regions/voxels in left and right hemispheres. Here, we expanded our study of structural asymmetry to the systematic level of hemispheres. We assumed that human brain could be disentangled into 3 distinct factors, i.e., left-hemisphere-specific factors, right-hemisphere-specific factors, and left-right-shared factors. Capitalizing on the recent advancements of decoupled representation learning [4, 5], we proposed a contrastive learning model to extract left-hemisphere-specific, right-hemisphere-specific and shared features from human anatomical hemispheres.

907 right-handed subjects with T1w from HCP were included (Male: 397, Handedness > 40, age 22y to 36y). Illustrated in Figure 1, we proposed a deep learning model to extract the hemisphere-specific features. The model was based on contrastive VAE [5] and the input of it was 3D hemisphere imaging data in native space. To ascertain the relationships between the hemisphere-specific features and behavioral data, we employed Partial Least Squares Correlation (PLSC) [6, 7]. Our sole focus was on two hemisphere-dominant cognition performances: language and social. Tensor-based Morphometry (TBM) was utilized to identify loci of hemisphere-specific regions [8]. Specifically, for each subject, we generated three synthetic brains: the first one constructed only from shared features, termed counterfactual brain; the second one composed of left-hemisphere-specific and shared features, termed synthetic left brain; and the third one with right-hemisphere-specific and shared features, termed synthetic right brain. Then a template was created using counterfactual brains from all subjects. 3 Jacobian determinants were calculated between each subject 's 3 synthetic brains and the template. Paired t-tests was then used on Jacobian determinants to capture structural changes separately induced by 2 specific features, followed by multiple comparison correction using 10,000 permutations.

Regarding the relationship between hemisphere-specific features and language, there was a single significant latent variable (LV) explaining 49.7% of the covariance (Fig. 2A). Remarkably, a left-specific feature showed significantly contribution to the language-related LV after the robustness assessment (BSR = 3.462, p < 0.01). For social, the result also revealed a significant LV (explaining 66.8%) and one right-specific feature significantly contributing to the social-related LV (BSR = 2.582, p < 0.01). The reproducibility validation utilizing a randomly selected 2/3 of the participants confirmed the stability of our findings. Figure 2B showed the loci of structural changes associated with hemisphere-specific features. Regions with negative values indicated compression of brain structural volume, while positive values indicated expansion. For left hemisphere, the regions with significant compression in volume were mainly located in the lateral prefrontal, lateral anterior temporal and medial occipital regions, while the regions with significant expansion were in cingulate gyrus. Right-brain-specific features exhibited significantly compression in the default mode network and significantly expansion mainly in the orbitofrontal cortex.

By combining the decoupled representation learning and structural MRI, we extracted the hemisphere-specific structural features of the human brain from the systematic level. Our results demonstrated that the hemisphere-specific multidimensional features extracted by our model revealed correlations with lateralized behavioral performance, indicating that the specific features could be disentangled using a data-driven approach. These findings might provide new insights about the brain asymmetry.

Multivariate Approaches ²

Other Methods

Cortical Anatomy and Brain Mapping ¹

Other - asymmetry; decoupled representation learning; AI for science