Material for: Minimal Linear Networks for MR image reconstructiom
The main topology, Complex-valued neural linear networks, a k and I layers with location-independent (variable-density) kernels with several "time-segments" is able to produce artifact-free images where standard advanced (e.g. compressed sensing, etc.) reconstruction fails.
Example on a spiral trajectory, compared vs. ESPIRiT with an augmented signal model with B0 inhomogeneities correction:
The difference is more pronounced in the case of a multi-band spiral trajectory with with incoherent CAIPI blips, vs. BART with augmented signal model:
Vs. common methods on our benchmark data:
Summary of proposed and reference reconstruction methods on real data for all 4 subjects can be found summarized at https://doi.org/10.6084/m9.figshare.7007777 and detailed for every slice at https://doi.org/10.6084/m9.figshare.7007774 .
NUFFT with B0 field inhomogeneities correction by the time-segments method is implemented in tensorflow to run on the GPU.
The code is based on https://github.com/david-gpu/srez, so in case of errors it might be useful to verify srez is working. Installation should take no more than several minutes.
You will need Python 3 with Tensorflow, numpy, scipy, h5py and moviepy.
See srez -requirements.txt for details.
The "CUDA_VISIBLE_DEVICES" is set at the beginning of srez_main1.py. The code uses some forced 'with /gpu:0'. Remove soft allocation (on srez_main1.py) to verify the main computation is done entirely on the GPU.
They system was developed on python 3.6.5, tensorflow 1.10.1, on a desktop machine with a GTX 1080 Ti GPU. MATLAB R2018a scripts are used for calling the system. Running pretrained nets should work well on CPU-only. 32GB RAM is advised.
The system is highly configurable from the human-readable params.txt . Most parameters are rather self-explanatory; some additional information can be found here: https://docs.google.com/document/d/18lZOREQs4aX6HWqjV1Dn5tCwnAgcu9XmiUsXeqp5uRQ/edit?usp=sharing
The dataset used for the real data and benchmark test is a collection of randomly chosen slices from the HCP. It can be downloaded from https://figshare.com/s/4e700474da52534efb30 . The data is augmented with random cropping, flipping and 90deg rotation, and a random 2D phase is added. The following parameters deternine the strength of the added phase: RandomPhaseLinearFac, RandomPhaseQuadraticFac, RandomPhaseScaleFac
For the real data, the acquired signal, the trajectory, MIRT-based NUFFT coefficients and time-segments data are included here. For the benchmark test, the poisson-disc masks and the images used are provided, as well as the reconstructed images using various methods.
wget -O HCPData_256x256_int16.mat https://ndownloader.figshare.com/files/12420215?private_link=4e700474da52534efb30
The download may take a moment. ~560MB
wget -O ./RealData/RegridTry3C2_7TS_S01_Sli06.tar.xz https://ndownloader.figshare.com/files/13198547?private_link=61f25dc380af32a12da3
tar xf RealData/RegridTry3C2_7TS_S01_Sli06.tar.xz -C RealData
The download may take several minutes. ~ 4.3GB
wget -O P2DF.tar.xz https://ndownloader.figshare.com/files/13198568?private_link=65c0e9f77f23c664aabe
tar xf P2DF.tar.xz
Training can be done by calling python srez_main1.py. The dataset used is set by the parameter DatasetMatFN in params.txt.
If ShowRealData=1, the output every summary_period (minutes) will include a run on the data in RealDataFN.
Running a trained network on a series of .mat files, given in the format Path/Prefix_XX.mat, can be done my setting the following parameters in params.txt:
LoadAndRunOnData=1,LoadAndRunOnData_checkpointP,LoadAndRunOnData_Prefix
LoadAndRunOnData_OutP, HowManyToRun - see example in current params_ex.txt
For running training on the example real data, from the base folder, after downloading the dataset, run:
source ~/tensorflow/bin/activate (or any other way to activate the tensorflow environment)
cp ./RealData/ParamsUsed.txt ./Params.txt
<Tensorflow-python> srez_main1.py
Where <Tensorflow-python> is e.g. ~/tensorflow/bin/python. The result should slowly accumulate in the folder RegridTry3C2_7TS_S01_Sli06__<DATE>
Download the trained network before.
source ~/tensorflow/bin/activate (or any other way to activate the tensorflow environment)
cp ./RealData/ParamsUsedForRun.txt ./Params.txt
<Tensorflow-python> srez_main1.py
The results will appear in RealData/OnRealData0{1,2}.mat. Note that real and imaginary channels are separated.
Output images are produced during the run. In the output images, each row contain the reconstruction on a different training image, in the order real, imaginary, magnitude, phase. Image used for training on the right, reconstructd image on the left.
Expected output can be seen in the image in the RealData folder: the reonstructed image from real data is shown on the first row (mismatching with non-existing training image).
RunTFForMatlab.sh should be edited with the correct folders, CUDA, etc. Execution should be enabled:
sudo chmod +777 ./RunTFForMatlab.sh
The MLN_Benchmark_Run_Trained_Net.m (download benchmark trained networks first) reconstructs images using the trained MLN. The script has some folder definitions that need to be adjusted.