Non-Cartesian 2D Parallel MRI (SENSE)

The Gadgetron includes a real-time implementation of a GPU-based real-time non-Cartesian Sense reconstruction published in IEEE Trans Med Imaging. 2009 Dec;28(12):1974-85 (www.ncbi.nlm.nih.gov). One of the keys to obtaining real-time performance is an efficient GPU implementation of the non-Cartesian Fast Fourier Transform IEEE Trans Med Imaging. 2008 Apr;27(4):538-47 (www.ncbi.nlm.nih.gov). The application reuses several of the gadgets we have seen in use already for the Cartesian Grappa implementation above ([Cartesian 2D Parallel MRI (GRAPPA)]). An overview of the non-Cartesian Sense gadget chain is given in the figure below.

This figure originates from the Gadgetron paper. As described in [Gadgetron Gadgets] this gadget has since been broken up to support both linear SENSE and kt-SENSE and non-linear compressed sensing SENSE. The CGSenseGadget (now the gpuRadialSensePrepGadget and gpuCgSenseGadget) implements the linear non-Cartesian Sense reconstruction. It contains a conjugate gradient solver (Linear Solvers) set up with a nonCartesianSense image encoding matrix and an imageOperator for regularization. Internally it maintains a cyclic buffer of a few seconds of imaging data. It uses this buffer to maintain a fully sampled (i.e. unaliased but blurred) k-space image from which coil sensititivities and regularization images are dynamically estimated. The combination of parallel imaging and image regularization operators allows for alias-suppressed image reconstruction using significant undersampling hereby achieving real-time data acquisition rates per frame. The conjugate gradient solver is able to reconstruct faster than the acquisition time e.g. a 192x192 image from 32 coils using 10 solver iterations on newer graphics hardware.

To test this configuration use the 32 channel radial MRI test dataset (golden_angle.h5), which you can download from https://sourceforge.net/projects/gadgetron/files/testdata/ismrmrd/. We assume that you have added \$(GADGETRON_HOME)/bin to your PATH environment variable. You need a CUDA enable GPU on your system and your Gadgetron distribution should be compiled with CUDA and CULA enabled. Please see ? for details for your specific platform.

To run the reconstruction; start up gadgetron (in its own terminal window) and use the mriclient to send the data from another terminal. First start gadgetron:

user@host$ gadgetron
Configuring services

If asked, allow the gadgetron application to allow incoming network
connection. Next start the mriclient:

user@host:~/temp$ wget http://sourceforge.net/projects/gadgetron/files/testdata/ismrmrd/golden_angle.h5

user@host:~/temp$ mriclient \
       -d golden_angle.h5 \
       -c golden_radial_mode2_realtime.xml

Gadgetron MRI Data Sender
  -- host            :      localhost
  -- port            :      9002
  -- hdf5 file  in   :      golden_angle.h5
  -- hdf5 group in   :      gre_golden_angle
  -- conf            :      golden_radial_mode2_realtime.xml
  -- loop            :      1
  -- hdf5 file out   :      ./out.h5
  -- hdf5 group out  :      2012-05-11 15:47:22
(32608|139797448419136) Connection from 127.0.0.1:9002
32608, 81, GadgetronConnector, Close Message received
(32608|139797376546560) Handling close...
(32608|139797376546560) svc done...
(32608|139797376546560) Handling close...

Your current folder now holds the reconstructed images in the out.h5
HDF5 file. They will look something like the one depicted in the figure below.