Q-space Deep Learning

https://user-images.githubusercontent.com/1512443/147379103-f3d80ff6-dcac-440b-b829-2111cfb685ae.mov

This repository provides implementation codes for "Jointly estimating parametric maps of multiple diffusion models from undersampled q-space data: A comparison of three deep learning approaches."

Setting up a Python env and installing required packages

To run Python codes provided in this repository, create a Python environment: $ conda create -p /path_to_env/env_name python=3.x and install the follwoing packages $ conda install -p /path_to_env/env_name/ -c anaconda numpy $ conda install -p /path_to_env/env_name/ -c conda-forge matplotlib $ conda install -p /path_to_env/env_name/ -c conda-forge nibabel $ conda install -p /path_to_env/env_name/ -c anaconda h5py $ conda install -p /path_to_env/env_name/ -c conda-forge tqdm $ conda install -p /path_to_env/env_name/ -c conda-forge argparse $ conda install -p /path_to_env/env_name/ pytorch torchvision torchaudio cudatoolkit=x.x -c pytorch where cudatoolkit=x.x in the last command depends on the installed cude version, e. g. 10.2 or 11.3. After these activate your environment using: $ conda activate /path_to_env/env_name

Training and Testing the 1D-qDL network

The 1D-qDL, as shown in the above block diagram, uese fully connected layers to jointly estimate parametric diffusion maps on a per-voxel basis. The implementation codes are under ./1D-qDL/. The training data are not provided in this repository, but one test dataset, trained models for three undersampling patterns, and the test results can be found at ./Data/.

Training: ``` usage: python train_1d.py [-h] [--sampling SAMPLING] [--batchsize BATCHSIZE] [--numofchannels NUMOFCHANNELS] [--numoflayers NUMOFLAYERS] [--numofhidden NUMOFHIDDEN] [--dropout DROPOUT] [--epochs EPOCHS] [--lr LR] [--datapath DATAPATH]

Training 1D-qDL

optional arguments: -h, --help show this help message and exit --sampling SAMPLING Q-space undersampling pattern name --batchsize BATCHSIZE Training batch size --numofchannels NUMOFCHANNELS Number of qDL input channels --numoflayers NUMOFLAYERS Number of qDL layers --numofhidden NUMOFHIDDEN Number of hidden nodes in qDL --dropout DROPOUT Drop out probability in qDL --epochs EPOCHS Number of training epochs --lr LR Initial learning rate --datapath DATAPATH Path to the data

``` Testing:

``` usage: python test1d.py [-h] [--sampling SAMPLING] [--batchsize BATCHSIZE] [--numofchannels NUMOFCHANNELS] [--numoflayers NUMOFLAYERS] [--numofhidden NUMOFHIDDEN] [--dropout DROPOUT] [--datapath DATAPATH] [--testcases TEST_CASES]

1D-qDL testing

optional arguments: -h, --help show this help message and exit --sampling SAMPLING Q-space undersampling pattern name --batchsize BATCHSIZE Testing batch size --numofchannels NUMOFCHANNELS Number of qDL input channels --numoflayers NUMOFLAYERS Number of qDL layers --numofhidden NUMOFHIDDEN Number of hidden nodes in qDL --dropout DROPOUT Drop out probability in qDL --datapath DATAPATH Path to the data --testcases TESTCASES List of test subjects ids ```

## Training and testing the 2D-CNN network

The 2D-CNN uses convolutional blocks with residual connections to jointly estimate diffusion parametric maps on a per-slice basis. The implementation codes are under ./2D_CNN/.

Training: ``` usage: python train2d.py [-h] [--sampling SAMPLING] [--batchsize BATCHSIZE] [--numofchannels NUMOFCHANNELS] [--epochs EPOCHS] [--lr LR] [--datapath DATA_PATH]

Training 2D-CNN

optional arguments: -h, --help show this help message and exit --sampling SAMPLING Q-space undersampling pattern name --batchsize BATCHSIZE Training batch size --numofchannels NUMOFCHANNELS Number of CNN input channels --epochs EPOCHS Number of training epochs --lr LR Initial learning rate --datapath DATAPATH Path to the data ``` Testing:

``` usage: python test_2d.py [-h] [--sampling SAMPLING] [--batchsize BATCHSIZE] [--numofchannels NUMOFCHANNELS] [--datapath DATAPATH] [--testcases TESTCASES]

2D-CNN testing

optional arguments: -h, --help show this help message and exit --sampling SAMPLING Q-space undersampling pattern name --batchsize BATCHSIZE Testing batch size --numofchannels NUMOFCHANNELS Number of CNN input channels --datapath DATAPATH Path to the data --testcases TESTCASES List of test subjects ids ```

## Training and Testing the MESC-SD network

The MESC-SD uses a dictionary-based sparse coding representation to jointly estimate parametric diffusion maps using 3D input patches. The implementation codes are under ./3D_MESC_SD/.

Training:

``` usage: python train_3d.py [-h] [--sampling SAMPLING] [--batchsize BATCHSIZE] [--numofchannels NUMOFCHANNELS] [--numofvoxels NUMOFVOXELS] [--numhiddenlstm NUMHIDDENLSTM] [--numhiddenfc NUMHIDDENFC] [--epochs EPOCHS] [--lr LR] [--datapath DATAPATH]

Training MESC-SD

optional arguments: -h, --help show this help message and exit --sampling SAMPLING Q-space undersampling pattern name --batchsize BATCHSIZE Training batch size --numofchannels NUMOFCHANNELS Number of MESCSD input channels --numofvoxels NUMOFVOXELS Number of voxels in 3D input patches --numhiddenlstm NUMHIDDENLSTM Number of hidden nodes in LSTM units --numhiddenfc NUMHIDDENFC Number of hidden nodes in FC layers --epochs EPOCHS Number of training epochs --lr LR Initial learning rate --datapath DATA_PATH Path to the data ```

Testing:

``` usage: python test3d.py [-h] [--sampling SAMPLING] [--batchsize BATCHSIZE] [--numofchannels NUMOFCHANNELS] [--numofvoxels NUMOFVOXELS] [--numhiddenlstm NUMHIDDENLSTM] [--numhiddenfc NUMHIDDENFC] [--datapath DATAPATH] [--testcases TEST_CASES]

MESC-SD testing

optional arguments: -h, --help show this help message and exit --sampling SAMPLING Q-space undersampling pattern name --batchsize BATCHSIZE Testing batch size --numofchannels NUMOFCHANNELS Number of MESCSD input channels --numofvoxels NUMOFVOXELS Number of voxels in 3D input patches --numhiddenlstm NUMHIDDENLSTM Number of hidden nodes in LSTM units --numhiddenfc NUMHIDDENFC Number of hidden nodes in FC layers --datapath DATAPATH Path to the data --testcases TEST_CASES List of test subjects ids`

```