Deep learning approaches to map individual differences in macroscopic neural structure with variations in spatial navigation behavior

Abstract

Understanding the association between structural properties of the human brain and individual differences in behavior is an ongoing endeavor, challenged by the brain's complexity. Past approaches, limited by simplistic neural structure measures like brain volume or cortical thickness, have given way to more advanced modeling approaches. Empirical evidence using these simpler metrics occasionally shows that hippocampal structure relates to individual variation in spatial navigation ability, particularly for older individuals or for expert navigators (like London taxi drivers). Yet high-powered, pre-registered studies in typical younger adults revealed no association between hippocampal volume and navigation ability. Here, we follow a data-driven approach developing and comparing deep learning methods (graph convolution neural networks, GCNN; 3DCNN) to analyze whether complex aspects of brain structure predict spatial navigation ability in young populations. To that end, we trained GCNNs and 3DCNNs on a T1 MRI dataset (N = 90) to predict navigational ability as measured by an objective virtual reality test of spatial memory in which participants created as accurate a map as they could of a highly realistic virtual environment. Across all approaches, we found weak predictive value in held-out test data, despite good fits to training data. These results could indicate the need for much larger datasets, including more comprehensive behavioral measures (as this study was limited to one measure) to improve predictability but may also support the notion that hippocampal structural features may not be a primary factor associated with navigation ability in healthy younger adults.

Keywords: 3D convolutional neural networks; Artificial intelligence; Deep learning; Graph convolutional neural networks; Hippocampus; Spatial navigation.