Modified U-Net (MU-Net). MU-Net is modified from U-Net. All max pooling layers and up-sampling layers are removed. The number of kernels is constant across all layers. The input is a noisy image volume while the output is the residual volume between the input and the high-signal-to-noise ratio (SNR) target image volume (i.e., the noise). Network parameter k = 128 is adopted in this study (~9.3 million parameters).
Modified U-Net with dropout (MDU-Net). MU-Net is modified from U-Net. The number of kernels is k during the encoding and 2_k_ during the decoding. The input is a noisy image volume with 10% randomly selected voxels replaced by the average of their six neighbouring voxels (i.e., average masking (AM)), while the output is another noisy image volume with the same selected 10% voxels with raw image intensities, on which losses are computed. Network parameter k = 128 is adopted in this study (~5.6 million parameters).
Image results. Exemplar image slices of a representative subject from Standard-MPRAGE data (a, i), raw Wave-MPRAGE data (a, ii), and the raw Wave-MPRAGE data denoised by BM4D (a, iii), AONLM (a, iv), supervised denoising using a MU-Net trained on simulation data from HCP subjects (a, v) and the same MU-Net fine-tuned on empirical data from another subject (a, vi), as well as Self2Self-AM denoising using a MDU-Net trained on simulation data from a HCP subject (c, i), a MDU-Net trained on empirical data from another subject (c, ii), and the same MDU-Net from another subject fine-tuned on the data of the subject for denoising for 1 epochs (c, iii), 5 epochs (c, iv) and 10 epochs (c, v), and a randomly initiated MDU-Net on the data of the subject for denoising for 100 epochs (c, vi), with magnified views of basal ganglia (b, d). The structural similarity index (SSIM) compared to Standard-MPRAGE images and whole-brain averaged vertex-wise gray–white matter boundary sharpness are listed to quantify image quality. The arrowheads highlight the claustrum (magenta) and caudolenticular gray bridges (blue) with fine textures.
Image quality quantification. The group means (± group standard deviations) of the structural similarity index (SSIM) of the raw and denoised Wave-CAIPI images compared to Standard-MPRAGE images as well as the whole-brain averaged vertex-wise gray–white matter boundary sharpness from all 10 subjects. The red color represents the highest metrics while the green color represents the second highest metrics.
install packages in requirements.txt Run: python s_S2S_cnnTrainS2s0.9Blur1x1.py python s_S2S_cnnApplyS2sSw1_0.9Blur1x1.py
Step-by-step Python tutorial for training the MDUnet in S2S-am.
Utility functions
-
s2s_unet1x1.py: create MDUnet model
-
qtlib.py: create custom loss functions to only include loss within brain mask, extract blocks from whole brain volume data, normalize brain volume in the brain mask.
Output
-
unet-sw1-blur0.91x1-flip/unet-sw1-blur0.91x1-flip_ep99.h5: MDUnet model trained for 100 epoches
-
unet-sw1-blur0.91x1-flip/unet-sw1-blur0.91x1-flip_ep99.mat: L2 losses for the training and validation
Step-by-step Python tutorial for training the MDUnet in S2S.
Utility functions
-
s2s_unet1x1.py: create MDUnet model
-
qtlib.py: create custom loss functions to only include loss within brain mask, extract blocks from whole brain volume data, normalize brain volume in the brain mask.
Output
-
unet-sw11x1/unet-sw11x1_ep99.h5: MDUnet model trained for 100 epoches
-
unet-sw11x1/unet-sw11x1_ep99.mat: L2 losses for the training and validation
Step-by-step Python tutorial for applying the MDUnet in S2S-am.
Utility functions
- qtlib.py: de-normalize brain volume in the brain mask and recover data volume from blocks.
Output
- unet-sw1-blur0.91x1-blur-pred/unet-sw1-blur0.91x1-flip_ep99_img_block_pred0019avg_pad90.nii.gz: s2s-am denoised result by averaging input with 20 random masks.
Step-by-step Python tutorial for applying the MDUnet in S2S-am.
Utility functions
- qtlib.py: de-normalize brain volume in the brain mask and recover data volume from blocks.
Output
- unet-sw11x1-pred/unet-sw11x1_ep99_img_block_pred0019avg_pad90.nii.gz: s2s denoised result by averaging input with 20 random masks.
The Wave-MPRAGE and Standard-MPRAGE T1-weighted MRI data of 10 healthy subjects acquired at the MGH Martinos Center will be made freely and publicly available.
[1] Tian Q, Li Z, Lo W, Li Z, Bilgic B, Polimeni J, Huang S. Improving the accessibility of deep learning-based denoising for aceelerated brain MRI using self-supervised learning and/or transfer learning. [2] Tian Q. Improving The Accessibility Of Deep Learning-Based Denoising For MRI Using Transfer Learning And Self-Supervised Learning, Online Power Pitch Oral Presentation. The 2022 Annual Scientific Meeting of ISMRM.