GCAN

Code for GCAN: Generative CounterfactualAttention-guided Network for Explainable Cognitive Decline Diagnostics based on fMRI Functional Connectivity. This paper has been accepted by MICCAI2024

Environment

torch ==1.13.1
numpy == 1.22.3  
nibabel == 1.10.2  
torchcam == 0.3.2  
torchvision == 0.14.1  
einops == 0.6.0  
python == 3.9.0  
imageio == 2.31.1

extract the imaging features

To run the model, you need to extract the fc by Matlab. You can use the batch operation of spm12 to finish this. Additionally, the input data shape might influence the kernel size of avgpooling in ResNet, you need to change the kernel size, if has bugs. TO get the counterfactual map, you need to train a pre-trained classifier first (set the mode as 'pretrain'). Then, use the image generator to create target label FC (set the mode as 'image_generator'). After subtracting the target label FC with the source label FC, you get the counterfactual map. Finally, set the mode as 'region-specific', train a new classifier to validate the contirbution of counterfatucal map to MCI diagnosis performance.

run the model

1. Create k fold csv file

generate_csv.py

2. pretrain classifier model

set the mode_net as pretrained classifier in opt.py

run
main.py

3. get counterfactual attention

set the mode_net as image_generator in opt.py

run
main.py

4. train final classifier

set the mode_net as region-specific in opt.py

run
main.py