Explorations of using a convolutional neural network to understand brain activations during movie watching

Abstract

Neuroimaging studies increasingly use naturalistic stimuli like video clips to trigger complex brain activations, but the complexity of such stimuli makes it difficult to assign specific functions to the resulting brain activations, particularly for higher-level content like social interactions. To address this challenge, researchers have turned to deep neural networks, e.g., convolutional neural networks (CNNs). CNNs have shown success in image recognition due to their different levels of features enabling high performance. In this study, we used pre-trained VGG-16, a popular CNN model, to analyze video data and extract hierarchical features from low-level shallow layers to high-level deeper layers, linking these activations to different levels of activation of the human brain. We hypothesized that activations in different layers of VGG-16 would be associated with different levels of brain activation and visual processing hierarchy in the brain. We were also curious about which brain regions would be associated with deeper convolutional layers in VGG-16. The study analyzed a functional MRI (fMRI) dataset where participants watched the cartoon movie Partly Cloudy. Frames of the videos were fed into VGG-16, and activation maps from different kernels and layers were extracted. Time series of the average activation patterns for each kernel were created and fed into a voxel-wise model to study brain activations. Results showed that lower convolutional layers (1st convolutional layer) were mostly associated with lower visual regions, but some kernels (6, 19, 24, 42, 55, and 58) surprisingly showed associations with activations in the posterior cingulate cortex, part of the default mode network. Deeper convolutional layers were associated with more anterior and lateral portions of the visual cortex (e.g., the lateral occipital complex) and the supramarginal gyrus. Analyzing activation features associated with different brain regions showed the promise and limitations of using CNNs to link video content to brain functions.