Contact: Yunan Wu
To develop a graph convolutional neural networks (GCNNs) method to predict fluid intelligence from the shape of the cortical ribbon and subcortical structures and reveal brain regions involved in the prediction.
Materials and Methods
T1-weighted MRI and Fluid Cognition Composite scores (fIQ) from 1097 healthy subjects from the Human Connectome Project (HCP) were included (596 women; mean age, 28.81 ± 3.71). For each subject, the inner and outer cortical surfaces were extracted using Freesurfer, and seven subcortical structures were modeled as surface meshes. Nodes of the graphs were the surface vertices, and edges of the graph were the links across vertices. The input features were defined as the Cartesian coordinates of surface vertices. GCNNs with residual learning blocks were trained and evaluated using six-fold nested cross validations. Models were composed of a pre-convolutional layer, three residual blocks, a post residual block and followed by a single fully connected layer with one output, that corresponds to fIQ score estimate. Models’ performances were measured with Mean Square Error (MSE) and Pearson’s coefficient of correlation (R). A gradient backpropagation-based visualization method (Grad-CAM) was applied to map cortical and subcortical areas the most involved in fIQ score predictions.
Combining cortical and subcortical surfaces (HCP-Combined) yielded the best predictions (MSE = 0.834, R = 0.454), followed by using only the cortical surfaces (MSE = 0.886, R = 0.381), and only the subcortical structures (MSE = 1.014, R = 0.155). The mapping demonstrated the discriminative locations for predicting fluid intelligence were mainly on the left temporal and parietal lobe and partly involved the right hippocampus and amygdala.
The GCNNs methods outperformed the prediction of fluid intelligence scores compared to current literature and the visual interpretability of the network suggested the significant potential for our method to help further define the inter-relationship brain structure and functions.
Teams: Yunan Wu,, Pierre Besson, Emanuel A. Azcona, S. Kathleen Bandt, Hans C. Breiter, Todd B. Parrish, Aggelos K. Katsaggelos