Environment for Linear and non-linear Image reconstruction using GPU Acceleration
- Kristian Bredies (University of Graz)
- Gerald Buchgraber (Graz University of Technology, now at Datenkraft IT-Consulting)
- Manuel Freiberger (Graz University of Technology, now at Anton Paar)
- Andreas Huber (Graz University of Technology)
- Florian Knoll (CAI2R, New York School of Medicine)
AGILE (Environment for Linear and non-linear Image reconstruction using Gpu Acceleration) is an open source library for GPU accelerated reconstruction problems in medical imaging. It includes reconstruction code for Fluorescence Tomography and Magnetic Resonance Imaging.
If you use this library or the provided data in your publications, please cite: Knoll, F.; Freiberger, M.; Bredies, K.; Stollberger, R.: AGILE: An open source library for image reconstruction using graphics card hardware acceleration. Proc. Intl. Soc. Mag. Reson. Med. 19:2554 (2011)
This work is funded and supported by the Austrian Science Fund (FWF) in the context of project 'SFB F3209-19' (Mathematical Optimization and Applications in Biomedical Sciences) MOBIS
AGILE was developed and tested on Linux (OpenSUSE and Ubuntu) systems. It uses CMake in order to generate the Makefiles in a flexible manner. You also need the third-party libraries Boost (http://www.boost.org/) and OpenMPI (http://www.open-mpi.org/) to compile the library.
It is suggested to make an out-of-source build where the resultant object files, libraries and executables will be placed separated from the source files. This helps to keep the source directories clean.
To build the library, create a new directory (e.g. 'build') and execute the 'cmake' command in there providing the root directory of AGILE (which should be the current directory) as argument. Afterwards 'make' can be executed in the build directory to generate the library.
Copy/paste instructions for the lazy to be executed right in this directory:
mkdir build
cd build
cmake ..
At this step you need to set the path where you installed the CUDA SDK:
ccmake .
Toggle to advanced mode (press 't') and enter the path where you installed the CUDA SDK in the variable "CUDA_SDK_ROOT_DIR". At this point, you can also specify to enable double-precision floating point support on the GPU. This can be done by setting the variable ENABLE_GPU_DOUBLE to ON. Note that this is a requirement to build the TGV-MRI example which will be skipped if the flag is not set.
Then compile the library using
make
####Note:
AGILE was tested on 64 bit platforms. If you want to compile it on a
32 bit system, you have to change two entries in
/apps/tgv_radial_image/recon/CMakeLists.txt
"link_libraries("-L/${CUDA_SDK_ROOT_DIR}/C/lib -lcutil_x86_64")" to
"link_libraries("-L/${CUDA_SDK_ROOT_DIR}/C/lib -lcutil")"
"target_link_libraries(nfft2d_tgv_recon_nogfx
If CMake cannot determine the location of the CUDA toolkit, you have to provide an variable called CUDA_TOOLKIT_ROOT_DIR that is set to the root directory of the CUDA toolkit. Normally this is the parent directory of the directory in which nvcc, the CUDA compiler, resides in. The following commands can be used on a system where the full path to the CUDA compiler is /opt/cuda/bin/nvcc:
mkdir build
cd build
cmake -DCUDA_TOOLKIT_ROOT_DIR=/opt/cuda ..
make
To install the compiled library, simply call
make install
On Linux machines the directory defaults to /usr/local/.
You can specify a different directory for installation by compiling and installing the library with
cmake -DCMAKE_INSTALL_PREFIX=/prefix/for/the/library ..
make install
The commands above will install the files into the directory /prefix/for/the/library/.
If you are new to the AGILE library, the tutorials are a good place to start. Run 'make doc' in the build directory. Afterwards open the file 'doc/html/index.html' in your favourite browser and proceed to the tutorials.
If you are especially interested in iterative TGV based MR image reconstruction, please see Florian Knolls AGILE-TGV implementation.
IRGN-T(G)V for Cartesian image reconstruction: The IRGN executable is available in the 'bin' folder. Type './IMT_IRGN --help' to see all available command line options. Example: ./IMT_IRGN --input=braindata.bin --output=recon.dcm --format=2 --calc=1
- version 1.4: 06.06.2018: Added command line functionality for imt_irgn application
- version 1.3: 01.01.2016: Included necessary extensions for dynamic MRI reconstrucion (AVIONIC) by Andreas Schwarzl
- Refactoring to support latest CUDA architecture
- Finite forward and backward differences for 3-D data vectors
- Forward and backward FFT operations for vector based data sets
- Extended vector utilities (l1/l2-norm, min-max element computation, element-wise compare)
- Extended thread assignment for large data sets
- version 1.2: 01.01.2014: Included IRGN-T(G)V for Cartesian image reconstruction (apps/imt_irgn) by Herbert Heigl
- version 1.1: 16.01.2012: Included CPU reference implementation for TGV.
- version 1.0: 28.07.2011: First release of AGILE.