DeepFDR

The deepfdr package provides a fully unsupervised, deep learning based FDR control method designed for spatial multiple testing. We utilize the W-Net and the concept of local index of significance (LIS) to connect unsupervised image segmentation to a 3D multiple testing problem that enables false discovery rate (FDR) control with a significant boost in statistical power. We tested our methodology using test statistics generated from FDG-PET data available at ADNI database. For specific details on methodology or training, please refer to our paper:

T. Kim, H. Shu, Q. Jia, and M. de Leon (2023). DeepFDR: A Deep Learning-based False Discovery Rate Control Method for Neuroimaging Data. Proceedings of The 27th International Conference on Artificial Intelligence and Statistics (AISTATS), PMLR 238:946-954, 2024.

Usage

The software offers two distinct training modalities: an interactive training interface leveraging Dash and Plotly for a dynamic, web-app-based training, and a standalone training option for those preferring operation without the web application. If one wants to activate the GUI, use the train_gui module:

python train_gui.py --labelpath {optional groundtruth file if using for simulation}
                    --datapath {input test statistics file}
                    --savepath {directory path for saving results}

If one wants to train without visualization, use the train module:

python train.py --labelpath {optional groundtruth file if using for simulation}
                --datapath {input test statistics file}
                --savepath {directory path for saving results}

The following figure illustrates a training session, where the dashed red line indicates the nominal FDR threshold:

Requirements and Installation

This package was developed using Python 3.9 and Pytorch 1.10.1 - please install the compatible version of Pytorch at https://pytorch.org/. To install the package, please run the following lines:

# Optional: upgrade pip
pip install --upgrade pip setuptools wheel

git clone https://github.com/kimtae55/DeepFDR
cd $PATH_TO_DeepFDR$
python setup.py install

Citing our work using bibtex

@inproceedings{kim2024deepfdr, title={DeepFDR: A Deep Learning-based False Discovery Rate Control Method for Neuroimaging Data}, author={Kim, Taehyo and Shu, Hai and Jia, Qiran and de Leon, Mony}, booktitle={International Conference on Artificial Intelligence and Statistics}, pages={946--954}, year={2024}, organization={PMLR} }

Updates

New updates to improving the model, app stability and user experience will be posted here.