Towards tDCS Digital Twins using Deep Learning-based Direct Estimation of Personalized Electrical Field Maps from T1-Weighted MRI

Transcranial Direct Current Stimulation (tDCS) is a non-invasive brain stimulation method that applies neuromodulatory effects to the brain via low-intensity, direct current. It has shown possible pos- itive effects in areas such as depression, substance use disorder, anxiety, and pain. Unfortunately, mixed trial results have delayed the field’s progress. Electrical current field approximation provides a way for tDCS researchers to estimate how an individual will respond to specific tDCS parameters. Publicly available physics-based stimulators have led to much progress; however, they can be error-prone, susceptible to quality issues (e.g., poor segmentation), and take multiple hours to run. Digital functional twins provide a method of estimating brain function in response to stimuli using computational methods. We seek to implement this idea for individualized tDCS. Hence, this work provides a proof-of-concept for generating electrical field maps for tDCS directly from T1-weighted magnetic resonance images (MRIs). Our deep learning method employs special loss regularizations to improve the model’s generalizability and calibration across individual scans and electrode montages. Users may enter a desired electrode montage in addition to the unique MRI for a custom output. Our dataset includes 442 unique individual heads from individuals across the adult lifespan. The pipeline can generate results on the scale of minutes, unlike physics-based systems that can take 1-3 hours. Overall, our methods will help streamline the process of individual current dose estimations for improved tDCS interventions.

Paper

This repository provides the official implementation of training and evaluation of the model as described in the following paper:

Towards tDCS Digital Twins using Deep Learning-based Direct Estimation of Personalized Electrical Field Maps from T1-Weighted MRI

Skylar E. Stolte¹, Aprinda Indahlastari^2,3, Alejandro Albizu^2,4, Adam J. Woods^2,3,4, and Ruogu Fang^1,2,5*

¹ J. Crayton Pruitt Family Department of Biomedical Engineering, Herbert Wertheim College of Engineering, University of Florida (UF), USA
² Center for Cognitive Aging and Memory, McKnight Brain Institute, UF, USA
³ Department of Clinical and Health Psychology, College of Public Health andHealth Professions, UF, USA
⁴ Department of Neuroscience, College of Medicine, UF, USA
⁵ Department of Electrical and Computer Engineering, Herbert Wertheim College ofEngineering, UF, USA

International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) 2024
[paper](TO BE ADDED) | [code]https://github.com/lab-smile/tDCS-DT | [poster](TO BE ADDED)

Major results from our work

• Custom dataset of T1-weighted MRIs and tDCS current maps for 442 individual heads across the lifespan
• Dual training system and loss regularization that also incorporates calibration
• Proof-of-concept results for the generation of tDCS current estimations directly from MRIs using deep learning

fig. 1: Method Pipeline.

Usage

Data Preparation

Our dataset consisted of unprocessed T1-weighted structural magnetic resonance images (MRIs) from our own in-house dataset and the T1-weighted MRIs from the 1200 Subject Release from the Human Connectome Project's (HCP) Young Adult Study (https://www.humanconnectome.org/study/hcp-young-adult/document/1200-subjects-data-release). Our dataset is still private awaiting the results of the finished parent study, but you can obtain the HCP data from filling out their permissions forms in accordance with their website.

The Data should be arranged within a root_dir of any name with the default as "Data". "Data" needs to contain folders named "Train", "Val", "Test". Each of these folders contain an additional three folders "T1", "J_C3Fp2", and "J_F3F4". They also need a .npy file with the image names. Samples of "train.npy", "val.npy", and "test.npy" are in each of the associated folders (Train, Val, Test, respectively) for reference.

ROAST

The reference data is generated using the Parra lab's ROAST tool (https://github.com/andypotatohy/roast). We used the electrical field outputs from this program to train our model. We used the below parameters when running ROAST.

% Recipe 
recipe = {'F3',-2,'F4',2};  
% I used F3-F4 and C3-Fp2 (one variation of M1-SO) as proof-of-concept for this paper

% Mesh Options
rb = 5; % Radial Bound
ab = 30; % Angular Bound
db = 0.3; % Distance Bound
rr = 3; % Radius Edge Ratio
mv = 10; % Max Elem Volume

You will need to use recipe of "F3-F4" and "C3-Fp2" once per participant to replicate this paper exactly. Rename files to be the subject number but in each appropriate folder. For instance, you have participant ID 100 in training. Then, you should put the T1.nii for participant 100 in "Train/T1/100.nii", the electrical field map results from ROAST with F3-F4 for participant 100 as "Train/J_F3F4/100.nii", and the electrical field map results from ROAST with C3-Fp2 for participant 100 as "Train/J_C3Fp2/100.nii".

Conda Environment

Initiate a conda environment to establish the necessary libraries and toolboxes.

%Create a new conda environment named "tdcs"
conda create --name tdcs

%Activate the "tdcs" environment
conda activate tdcs

%Install packages listed in the requirements.txt file
pip install -r requirements.txt

%Close the environment
conda deactivate

Training

The training code can be run as below:

conda activate tdcs

python train.py -x 256 -y 256 -z 256 -batch_size 2 -root_dir 'Data' -num_epochs 100 -save_path 'model.pth'

Change x, y, and z to the dimensions of your image size or your desired image size after cropping. The training code will crop the images automatically if you select a smaller size than your total image size. The batch_size can be changed depending on the memory of your gpu. The root_dir is where your source data is located. The num_epochs can be any number. The save_path is where your trained model saves.

Testing

The training code can be run as below:

conda activate tdcs

python test.py -x 256 -y 256 -z 256 -batch_size 2 -root_dir 'Data' -save_path 'model.pth'

The testing code needs to use the same arguments as the training code to work correctly.

Acknowledgement

This work was supported by the National Institutes of Health/National Institute on Aging (NIA RF1AG071469, NIA R01AG054077), the National Science Foundation (1842473, 1908299, 2123809), the University of Florida McKnight Brain Institute, the University of Florida Center for Cognitive Aging and Memory, and the McKnight Brain Research Foundation.

Source Networks -

SwinUNETR: https://arxiv.org/abs/2201.01266 https://github.com/Project-MONAI/research-contributions/tree/main/SwinUNETR

EfficientNet: https://arxiv.org/pdf/1905.11946 https://github.com/lukemelas/EfficientNet-PyTorch

Citation

If you reference our work, please cite:

@inproceedings{stolte2024towards,
  author    = {Skylar E. Stolte and Aprinda Indahlastari and Alejandro Albizu and Adam J. Woods and Ruogu Fang},
  title     = {Towards tDCS Digital Twins Using Deep Learning-Based Direct Estimation of Personalized Electrical Field Maps from T1-Weighted MRI},
  booktitle = {Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2024, Part II},
  series    = {Lecture Notes in Computer Science},
  volume    = {15002},
  publisher = {Springer Nature},
  year      = {2024},
}

Contact

For any discussion, suggestions and questions please contact: Skylar Stolte, Dr. Ruogu Fang.

Smart Medical Informatics Learning & Evaluation Laboratory, Dept. of Biomedical Engineering, University of Florida