Abstract
Human visual cortex is comprised of a number of areas, nearly all of which are organized retinotopically. This retinotopic mapping is similar across individuals; however, considerable inter-subject variation does exist, and this variation has been linked to cortical folding and other anatomical features. It was our aim to develop a neural network capable of learning the complex relationship between the functional organization of visual cortex and the underlying anatomy. Importantly, we used a geometric approach able to directly interact with data represented on cortical surface models, as many properties only make sense considering their location with respect to the various sulci and gyri.
To build our network, we used the most comprehensive retinotopy dataset openly available – that from the Human Connectome Project. This dataset includes 7T fMRI data from 181 participants. The data serving as input to our network included curvature and myelin values as well as the connectivity among vertices forming the cortical surface and their spatial disposition. The output of the network was the retinotopic mapping value for each vertex of the surface model. Our final network included 12 spline-based convolution layers, interleaved by batch normalization and dropout.
Our neural network accurately predicted the main features of both polar angle and eccentricity maps. More impressive yet, we show that the network was able to predict nuanced variations in the retinotopic maps across individuals. We further showed how disruption of the spatial organization of the input features increases the error and reduces the individual variability of the predicted maps.
In conclusion, we were able to predict the detailed functional organization of visual cortex from anatomical properties alone using geometric deep learning. Although we demonstrate its utility for visual neuroscience, geometric deep learning is flexible and well-suited for a range of other applications involving data structured in non-Euclidean spaces.