The proposed method enables a multi-resolution solution using wavelet decomposition for DWI denoising. See the system pipeline below.

Our main contributions are summarized as following:

• We propose a multi-resolution deep learning framework for DWI image denoising.
• We propose a novel wavelet channel transfer framework to reconstruct higher-frequency wavelet coefficients from lower-frequency wavelet coefficient based on 3D pix2pix style transfer network and 3D-SPADE.
• We show that our method outperforms our implementation of DeepDTI for dMRI scans corresponding to randomly chosen gradient directions, and the proposed wavelet channel transfer network outperforms the conventional CNN model.

• Clone this repo;
• Prepare the environment by:

cd Multi-channels-DWIs-Denoising
pip -r requirements.txt

• Download the dataset (HCP dataset);
• Preprocess the dataset (generate the clean DWI, group as subset, wavelet transformation, etc.);

The train.py provides the train function for one wavelet channel. The training set should be organized as:

|--train_dir
    |--subject id
        |--aaa
            |--b1data0.nii.gz
            |--b1clean0.nii.gz
            |--b1data1.nii.gz
            |--b1clean1.nii.gz
            |--mask.nii.gz (optional)
        |--aad
            ...

Note the input of the synthesis net comprises 4 lower-frequency coefficients from the same volume, and the output is the translated higher coefficient.

• Train the lower-frequency denoising channel aaa as example:

python train.py --train_dir train/ --val_dir val/ -e 100 -b 4 -l 0.00005 --dropout 0.2 --wt aaa --net dncnn

• Or use UNet+SPADE as synthesis net to train the higher-frequency channel ddd:

python train.py --train_dir train/ --val_dir val/ -e 100 -b 4 -l 0.0001 --dropout 0.2 --wt add --net unet+spade

• Organize the directory for 8 models in one folder, see detailed arrangement here;

• (optional) Obtain brain mask and normalization parameters for 8 wavelet coefficients;

• You can download the trained models, test examples and normalization parameters (include b=1000, 3000 ms/mm^2) here:

  • test examples
  • trained models
  • normalization parameters

• Test the b1/b3 DWIs using command:

python test.py --test_dir test/146331/b1data.nii.gz --output_dir test/146331/b1denoised.nii.gz --gt_dir test/146331/b1clean.nii.gz --mask_dir test/146331/wavelet_mask.nii.gz --bval b1 --metrix --denoise_in 7 --denoise_out 7
python test.py --test_dir test/146331/b3data.nii.gz --output_dir test/146331/b3denoised.nii.gz --gt_dir test/146331/b3clean.nii.gz --mask_dir test/146331/wavelet_mask.nii.gz --bval b3 --metrix

Junyan Wang, Jinnan Hu, Deep Multi-resolution Diffusion Image Denoising via Low-Frequency-to-High-Frequency Channel Translation, under review, 2022

Please feel free to contact us if you have any problem. [email protected]