SuperSurfer Tutorial

SuperSurfer pipeline. The input is a low-resolution image volume, and the output is the residual between high-resolution image volume and the input. SuperSurfer has a network architecture comprised of stacked convolutional filters paired with nonlinear activation functions. The n-layer SuperSurfer network consists of a first layer of paired 3D convolution (k kernels of size d × d × d × 1 voxels, stride of 1 × 1 × 1 voxel) and rectified linear unit (ReLU) nonlinear activation, n − 2 middle layers of paired 3D convolution (k kernels of size d × d × d × k voxels, stride of 1 × 1 × 1 voxel) and ReLU activation, and a last 3D convolutional layer (one kernel of size d × d × d × k voxels, stride of 1 × 1 × 1 voxel). Parameter values are set n = 20, k = 64, d = 3.

Comparison of results. SuperSurfer synthesis results improved signal-to-noise ratio, image sharpness, and detailed anatomical information. It improves cortical surface reconstruction, and provides a unique opportunity to re-analyze legacy data acquired at standard 1-millimeter resolution. Quantitative comparison can be found in the NeuroImage paper of SuperSurfer.

s_t1wSim.m

Step-by-step MATLAB tutorial for simulating the low-resolution images from the high-resolution data for convolutional neural network in SuperSurfer. HTML file can be automatically generated using command: publish('s_t1wSim.m', 'html').

Input

• data.mat: high-resolution images

Output

• data_sim.mat: low-resolution images prepared for CNN

s_SuperSurfer_train.py

Step-by-step Python tutorial for training the SuperSurfer using data prepared using the s_Sim.m script.

Utility functions

• supersurfer.py: create SuperSurfer model

• utils.py: create custom loss functions to only include loss within brain mask; extract blocks from whole brain volume data; normalize and denormalize image

Output

• supersurfer_ep60.h5: SuperSurfer model trained for 60 epoches
• supersurfer_ep60.mat: L2 losses for the training and validation

s_SuperSurfer_apply.py

Step-by-step Python tutorial for applying the trained model to synthesize high-resolution image.

Utility function

• utils.py

Output

• data_supersurfer.mat: synthesized high-resolution image

compare.m

Step-by-step MATLAB tutorial for computing the whole-brain averaged MAD, PSNR, SSIM of simulated noisy images and denoised images.

Output

• diff.png: difference map with MAD and PSNR
• SSIM.png: SSIM map with SSIM

HCP data

The example data are provided by the WU-Minn-Oxford Hhuman Connectome Project (HCP) (open access data use term at https://www.humanconnectome.org/study/hcp-young-adult/document/wu-minn-hcp-consortium-open-access-data-use-terms). Please acknowledge the source to the WU-Minn-Oxford HCP. The orginal data is available at https://www.humanconnectome.org/study/hcp-young-adult.

Refereces

[1] Tian Q, Bilgic B, Fan Q, Ngamsombat C, Zaretskaya N, Fultz NE, Ohringer NA, Chaudhari AS, Hu Y, Witzel T, Setsompop K, Polimeni JR, Huang SY. Improving in vivo human cerebral cortical surface reconstruction using data-driven super-resolution. Cerebral Cortex, 2021, 31(1): 463-482. [PDF]