Voxelwise Modeling Reveals Selectivity for Body Part Identity and Location in BOLD fMRI Responses to Complex Naturalistic Stimuli
Loading...
Authors
Shinkle, Matthew W.
Issue Date
2024
Type
Thesis
Language
Keywords
Body Representation , Encoding Models , FMRI , Neural Networks
Alternative Title
Abstract
Information about bodies plays an important role in human interaction, and multiple regions in human lateral occipitotemporal cortex (LOTC) represent body parts, poses, and actions. However, different studies have reported varying estimates of the extent and cortical arrangement of body representations in LOTC. We adapted the CDCL part segmentation model to estimate body part locations in the visual field across three sets of naturalistic stimuli: 1) static naturalistic images, 2) naturalistic movie clips, and 3) moving computer-rendered scenes. We used voxelwise modeling to characterize the relationships between these body features and BOLD fMRI responses for multiple participants in each data set. The resulting encoding models accurately predict BOLD responses in LOTC around the hMT+ complex. Variance partitioning against motion energy features and simple semantic labels reveals unique variance explained by our body part model extending around the hMT+ boundary. This constitutes a robust and general test of the selectivity for body parts in naturalistic stimuli. In contrast, unique variance explained by our motion energy model shows a strong pattern of motion selectivity in the center of hMT+. Previous work has suggested the existence of several distinct regions of body part selectivity in LOTC. Contrasts of our regression weights do show differences in selectivity in conventionally face selective regions (OFA, pSTS), and reveal some evidence of other part selective regions around hMT+.