sfDLN

This repository includes the implementation of structure-function discrepancy learning which is proposed in "Graph Neural Networks Identify An Increased Discrepancy Between Structural and Functional Brain Connectomes Over the Cognitive Decline Continuum"

Hyperparameter Experiments on ADD-SCI Classification

ChebyNet(snet-lnet) sfDLN is used for experiments given below.

1 ) Impact of the k for the k-NN Classifier

k	Accuracy	F1 Score
3	0.853 ± 0.011	0.832 ± 0.007
5	0.872 ± 0.009	0.836 ± 0.007
7	0.879 ± 0.007	0.845 ± 0.009
9	0.879 ± 0.010	0.844 ± 0.013

2 ) Impact of the K for the ChebyNet

K	Accuracy	F1 Score
0	0.512 ± 0.045	0.465 ± 0.086
1	0.879 ± 0.007	0.845 ± 0.009
2	0.834 ± 0.018	0.792 ± 0.020

sfDLN on ADD-MCI and MCI-SCI Classification

ChebyNet(snet-lnet)-sfDLN is used for experiments given below. ChebyNet- sfDLN - I refers to the sfDLN trained using inverted (alternative) hypothesis of decreasing discrepancy.

1 ) ADD-MCI Classification

Model	Accuracy	F1 Score
ChebyNet-sfDLN-I	0.655 ± 0.023	0.510 ± 0.046
ChebyNet-sfDLN	0.712 ± 0.027	0.636 ± 0.040

2 ) MCI-SCI Classification

Model	Accuracy	F1 Score
ChebyNet-sfDLN-I	0.520 ± 0.036	0.458 ± 0.040
ChebyNet-sfDLN	0.545 ± 0.029	0.478 ± 0.027