Toward Generalizing Visual Brain Decoding to Unseen Subjects

  

Authors: Xiangtao Kong^* , Kexin Huang^* , Ping Li^† and Lei Zhang^†

Abstract

Visual brain decoding aims to decode visual information from human brain activities. Despite the great progress, one critical limitation of current brain decoding research lies in the lack of generalization capability to unseen subjects. Prior works typically focus on decoding brain activity of individuals based on the observation that different subjects exhibit different brain activities, while it remains unclear whether brain decoding can be generalized to unseen subjects. This study is designed to answer this question. We first consolidate an image-fMRI dataset consisting of stimulus-image and fMRI-response pairs, involving 177 subjects in the movie-viewing task of the Human Connectome Project (HCP). This dataset allows us to investigate the brain decoding performance with the increase of participants. We then present a learning paradigm that applies uniform processing across all subjects, instead of employing different network heads or tokenizers for individuals as in previous methods, which can accommodate a large number of subjects to explore the generalization capability across different subjects. We conduct a series of experiments and find the following: First, the network exhibits clear generalization capabilities with the increase of training subjects. Second, the generalization capability is common to popular network architectures (MLP, CNN and Transformer). Third, the generalization performance is affected by the similarity between subjects. Our findings reveal the inherent similarities in brain activities across individuals. With the emerging of larger and more comprehensive datasets, it is possible to train a brain decoding foundation model in the future.

⭐ If TGBD is helpful to your images or projects, please help star this repo. Thanks! 🤗

🔎 Overview

We present a learning paradigm that applies uniform processing across all subjects, instead of employing different network heads or tokenizers for individuals as in previous methods, which can accommodate a large number of subjects to explore the generalization capability across different subjects.

The network exhibits clear generalization capabilities with the increase of training subjects.

📷 Visual Results (click to expand)

⚙️ Dependencies and Installation

## git clone this repository
git clone https://github.com/Xiangtaokong/TGBD.git
cd TGBD

# create an environment
conda env create -f environment.yml

🚀 Test

Setp 1 Download and Preprocess the Dataset

Baidu Drive. Key: fevm

TGBD/movie_frame.zip contains the training images which are extracted from movies. TGBD/hcp_Movie_fmri/XXX contains the corresponding fmri voxels of different participants. You can download a part of all 177 subjects to save disk space. Run Datasets/HCP/read_nii1.py to generate the final fmri voxels uesd in this work.

Setp 2 Download the pre-trained models

Baidu Drive. Key: fevm

TGBD/pretrained_models contains models pre-trained on 1 and 177 subjects.

Setp 3 Edit the test yml file

Edit Brain_decoding/options/test/xxxx.yml to make sure the data/model paths are correct. subj1 or 177.yml means testing the model trained on 1 or 177 subjects.

Setp 4 Run the command

cd Brain_decoding
python test.py --opt options/test/xxxx.yml

⭐ Train

Setp 1 Download and Preprocess the Dataset

Baidu Drive. Key: fevm

TGBD/movie_frame.zip contains the training images which are extracted from movies. TGBD/hcp_Movie_fmri/XXX contains the corresponding fmri voxels of different participants. You can download a part of all 177 subjects to save disk space. Run Datasets/HCP/read_nii1.py to generate the final fmri voxels uesd in this work.

Setp 2 Edit the test yml file

Edit Brain_decoding/options/train/xxxx.yml to make sure the data/model paths are correct. subj1 or 177.yml means training the model on 1 or 177 subjects.

Setp 3 Run the command

cd Brain_decoding
python train.py --opt options/test/xxxx.yml

❤️ Acknowledgmentst

We use Pytorch Lightning template from this project and thank the excellent code of MindEye.

📧 Contact

If you have any questions, please feel free to contact: xiangtao.kong@connect.polyu.hk

🎓Citations

If our code helps your research or work, please consider citing our paper. The following are BibTeX references:

@article{kong2024toward,
  title={Toward Generalizing Visual Brain Decoding to Unseen Subjects},
  author={Kong, Xiangtao and Huang, Kexin and Li, Ping and Zhang, Lei},
  journal={arXiv preprint arXiv:2410.14445},
  year={2024}
}

🎫 License

This project is released under the Apache 2.0 license.

statistics