Deep Learning Analysis of fMRI Data for Alzheimer's Disease Prediction

This repository contains the implementation of deep learning models for Alzheimer's Disease prediction using fMRI data, as described in our paper Deep learning analysis of fMRI data for predicting Alzheimer's Disease: A focus on convolutional neural networks and model interpretability.

Requirements

Environment Setup

module load miniconda
module load cuDNN/8.0.5.39-CUDA-11.1.1
conda activate fmgene

Required Packages

pip install torch matplotlib numpy scipy pydicom shap pandas

Data Access

The data used in this study is from the Alzheimer's Disease Neuroimaging Initiative (ADNI). To access the data:

1. Submit an application at ADNI
2. Complete ADNI's Data Use Agreement
3. Once approved, download the fMRI scans

Models

This repository includes implementations of several models:

1. CNN (Main Model)

   • 3D Convolutional Neural Network
   • Achieves 92.8% accuracy on test set
   • Implementation in main.py

2. Baseline Models

   • RNN: main_rnn.py
   • MLP: main_mlp.py
   • Merged (CNN + genetic data): main_merged.py
   • Gupta's CNN implementation: main_gupta.py

Usage

python main.py  # For the primary CNN model

Configuration

• Model parameters can be modified in config.json
• Default train/validation/test split ratio: 60%/20%/20%
• Dataset split can be modified in the dataloader file

Results

Results will be displayed for the test dataset by default. To view results for other splits:

• Modify the evaluation section in the respective main files
• Available splits: train, valid, test

Last Updated

Dec. 15, 2024

Citation

If you use this code, please cite our paper:

@article{zhou2024deep,
  title={Deep learning analysis of fMRI data for predicting Alzheimer’s Disease: A focus on convolutional neural networks and model interpretability},
  author={Zhou, Xiao and Kedia, Sanchita and Meng, Ran and Gerstein, Mark},
  journal={PloS one},
  volume={19},
  number={12},
  pages={e0312848},
  year={2024},
  publisher={Public Library of Science San Francisco, CA USA}
}

License

Contact

[email protected]

Note: This implementation requires CUDA-capable hardware and appropriate CUDA/cuDNN installations.