IC-GAN fMRI Reconstruction

Official repository for the IJCNN 2022 (Accepted Oral) paper "Reconstruction of Perceived Images from fMRI Patterns and Semantic Brain Exploration using Instance-Conditioned GANs" by Furkan Ozcelik, Bhavin Choksi, Milad Mozafari, Leila Reddy, Rufin VanRullen.

Results

The following are a few reconstructions obtained :

Create conda environment using environment.yml in ic_gan directory by entering conda env create -f environment.yml . You can also create environment by checking requirements yourself.
For preparation of Kamitani images you should also include some required libraries pip install pandas scikit-image imageio
Before loading ICGAN model you should download the pretrained model and required library using:

cd ic_gan
chmod u+x download-weights.sh
sh ./download-weights
pip install pytorch-pretrained-biggan

For copyright reasons, we cannot share images used in this study. You can request access to Imagenet images used in Generic Object Decoding study by applying this form as stated in KamitaniLab/GenericObjectDecoding repository. Downloaded "images" directory should be added to KamitaniData dir.
Inverted noise and dense vectors are provided in this link together with extracted instance features but you can extract the instance features your self by applying first 2 steps stated below. (Noise and Dense vector inversion codes will be provided in future.)

After setting up the environment and downloading the images provided with form;

Preprocess images using

cd KamitaniData 
python kamitani_image_prepare.py

Extract instance features of Kamitani images using python extract_features.py
Train regression models using python train_regression.py -sub 3 (You can change the subject num between 1-5) You need extracted features in order to run this code successfully.
Reconstruct images from test fMRI data using python reconstruct_images.py -sub 3
Explore ROI semantics by ROI maximization using python explore_roi_semantics.py -sub 3