This section demonstrates how to run a standalone non-Cartesian parallel MRI reconstruction similar to the one that was previously shown using the streaming framework infrastructure in section?. More details can be found in [Non-Cartesian 2D Parallel MRI (SENSE)].
In addition to a regularized linear least squares solution to the reconstruction problem, we furthermore use the Split Bregman solver to obtain the solution with minimum total variation subject to the constraint of the encoding operator (compressed sensing).
Download a free-breathing cardiac MRI sample dataset from http://sourceforge.net/projects/gadgetron/files/testdata/standalone/fb_data.zip
Source code is found in the files main_cg.cpp and main_sbc.cpp in directory
$(GADGETRON_SOURCE)/apps/standalone/gpu/MRI/sense/noncartesian/radial/2d_golden_ratio.
Both command lines below produce a 2D image sequence, each image with a matrix size of 1922 (-m). 32 projections are used for each frame (-p) for a frame rate of roughly 32 profiles/frame * 2.5 ms/profile = 80 ms ms/frame or 12.5 frames/s. The reconstruction results are written out in both complex and magnitude data as result.cplx and result.real respectively.
If sufficient device memory is available on your GPU (i.e. you are in possession of a high-end card) all frames in the sequence can be reconstructed concurrently (as a 3D volume). On systems that do not hold enough device memory to reconstruct all frames in parallel, they can instead be reconstructed in several batches. The -f option to the command lines below indicate the number of frames that are reconstructed per batch. A negative value indicates "all". If the command below fails to complete due to lack of device memory, try running with argument -f 8 (or an even smaller number) instead.
The following output was obtained on a Geforce GTX 480 GPU.
user@host$ radial_sense_cg -d fb_data.cplx -m 192 -o 256 -p 32 -K 0.01
Running reconstruction with the following parameters:
----------------------------------------------------
Sample data file name : fb_data.cplx
Result file name : result.cplx
Matrix size : 192
Oversampled matrix size : 256
Profiles per frame : 32
Frames per reconstruction (negative meaning all) : -1
Number of iterations : 10
Kernel width : 5.5
Kappa : 0.01
----------------------------------------------------
Loading data: 18.339 ms
#samples/profile: 256
#profiles/frame: 32
#profiles: 2560
#coils: 4
#frames/reconstruction: 80
#profiles/reconstruction: 2560
#samples/reconstruction: 655360
Filling rhs buffer: 283.675 ms
Estimating csm: 3.435 ms
Computing regularization: 0.319 ms
Computing preconditioning weights: 0.081 ms
Iterating...
Iteration 0. rq/rq_0 = 0.453177
Iteration 1. rq/rq_0 = 0.132643
Iteration 2. rq/rq_0 = 0.0413432
Iteration 3. rq/rq_0 = 0.0144378
Iteration 4. rq/rq_0 = 0.00681063
Iteration 5. rq/rq_0 = 0.00450857
Iteration 6. rq/rq_0 = 0.00342872
Iteration 7. rq/rq_0 = 0.00240418
Iteration 8. rq/rq_0 = 0.00146108
Iteration 9. rq/rq_0 = 0.000903398
GPU Conjugate Gradient solve: 2115.7 ms
Full SENSE reconstruction.: 2188.68 ms
Writing out result: 50.111 ms
user@host$
user@host$ radial_sense_sbc -d fb_data.cplx -m 192 -o 256 -p 32
Running reconstruction with the following parameters:
----------------------------------------------------
Sample data file name : fb_data.cplx
Result file name : result.cplx
Matrix size : 192
Oversampled matrix size : 256
Profiles per frame : 32
Frames per reconstruction (negative meaning all) : -1
Number of cg iterations : 20
Number of sb inner iterations : 1
Number of sb outer iterations : 20
Kernel width : 5.5
Mu : 1.0
Lambda : 2.0
----------------------------------------------------
Loading data: 16.082 ms
#samples/profile: 256
#profiles/frame: 32
#profiles: 2560
#coils: 4
#frames/reconstruction 80
#profiles/reconstruction 2560
#samples/reconstruction 655360
CSM and regularization estimation: 288.983 ms
...
GPU constrained Split Bregman solve: 57257.4 ms
Full SENSE reconstruction with TV regularization.: 57330.3 ms
Writing out result: 50.421 ms
user@host$
As all 80 frames are reconstructed in parallel it is straightforward to add temporal regularization to the reconstructions. We leave this as a suggested exercise for the reader.
For the interested reader, an implementation of kt-Sense can be found in directory
$(GADGETRON_SOURCE)/apps/standalone/gpu/MRI/sense/noncartesian/radial/2d_golden_ratio_kt.
Additionally, the source code for the user interface demonstrated in the movie accompanying ? can be found (if you configured cmake to inlude Qt support) in direcotry
$(GADGETRON_SOURCE)/apps/standalone/gpu/MRI/sense/noncartesian/radial/2d_golden_ratio_gui.