Jeong-Won Jeong1, Soumyanil Banerjee2, Min-Hee Lee1, Nolan O'Hara3, Csaba Juhasz1, Eishi Asano1, and Ming Dong2
1Pediatrics, Wayne State University, Detroit, MI, United States, 2Computer Science, Wayne State University, Detroit, MI, United States, 3Translational Neuroscience Program, Wayne State University, Detroit, MI, United States
1Pediatrics, Wayne State University, Detroit, MI, United States, 2Computer Science, Wayne State University, Detroit, MI, United States, 3Translational Neuroscience Program, Wayne State University, Detroit, MI, United States
Global reasoning and aggregation of DWI connectome
features improved the prediction of language impairment in children with focal
epilepsy without relying on a specific DWI connectome node configuration when compared
to other deep learning methods.
Figure 3. Examples of activation maps highlighting the important Ωi for
prediction of expressive and receptive language score. Each 3D visualization
shows the nodes of Ωi and their contributions (i.e., Z scores of attenuation
weights) to increase the prediction accuracy, quantified by the thickness of an
individual tube. Thicker tubes indicate that a node is more predictive of observed
language scores. Relatively small attention weights (i.e., Z-scores less than
2.5 standard deviations of each map) were omitted for clarity.
Figure 1. The architecture diagram of the proposed CNN+GCN model, which takes the connectome
matrix S as input, to predict an output score t. Here, “**” indicates 2D
convolution, and “*” indicates 1D convolution. In the final layer, a linear
unit is used for language score regression.
