Skip to main content
. 2022 Feb 4;15:828512. doi: 10.3389/fnins.2021.828512

Figure 2.

Figure 2

The proposed ID-GCN architecture. The selected features are trained in three invertible blocks. A fully connected (FC) layer is finally used to obtain the output scores for ASD classification. The whole network is reversible before the FC layer, meaning that we can reconstruct the informative disease-related brain connectivity patterns by selecting important output features of the network.