Vous êtes sur la page 1sur 6

Directed Study Report :

Medical Image Processing on GPU


Bita Kazemi Zahrani
Supervision of : Prof. Bhandarkar

1 INTRODUCTION
The use of GPU for Medical Image Processing is getting widespread in recent years. Many image
modalities namely MRI, CT, PET, SPECT, DTI and a lot more have been exploiting the GPU speedup in
recent years. The objective of this report is to provide a vivid insight over the applications of GPU in
medical image processing and explain the tools and methodologies in a nutshell.

2 OVERVIEW
NVIDIA new architecture has turned GPU into a massively parallel multithreaded machine which
features flexible memory access, fast user-controllable on-chip memory and integer, bitwise operations.
CUDA has been the C-extension for GPU-programming without API graphics overhead and support for
different development tools.
Over the recent years GPU computing has been applied to different medical image processing tools like
ITK, different computed tomography tools, MRI auto calibration tools, segmentation, registration and
rendering tools.
In the following sections we will discuss each adaptation to GPU programming paradigm with more
detail and introduce the toolboxes. The report is organized as follows. Section 3, the methodologies and
different toolboxes are discussed in detail. A comparison of GPU speedup in different toolbox and
methodologies is followed in section 4. Conclusion to the report and the references are organized in
section 5 and 6 respectively.

3 METHODOLOGY
In this section we will discuss about different methodologies and tools taking advantage of GPU
paradigm. Using ITK in CUDA, utilization of GPU in computed tomography and its application in CT
modality is discussed. Buffalo toolkit for high speed 3D tomography is introduced. Application of CUDA
in MRI is further discussed. Later we will take a brief overview on how different image processing
techniques such as segmentation and registration and rendering are performed and enhanced by using
GPUs.
3.1 ITK ON CUDA :
GOU acceleration for ITK has been proposed by Harvard University and University of Utah. This feature
keeps the application of CPU/GPU transparent to the user. IT supports image filters and pipelining
methods. The project provides a GPU data, context and kernel manager to help users to use OpenCL
with ITK as well. New classes are added to ITK are GPUImage, GPUDataManager, GPUImageDataManager,
GPUContextManager, GPUKernelManager, GPUImageToImageFilter, GPUMeanImageFilter,
GPUInPlaceImageFilter and GPUFiniteDifferenceImageFilter. We will discuss these classes and some of their
speedups throughout this report [8].

Registration and image processing functions are computationally expensive but highly parallizable thus
they can benefit from GPU enabled capabilities. CUDA code is successfully integrated into ITK so that it
can be transparent from the users and there is no need to change the code using ITK. The only thing that
an image processor should be concerned about in exploiting GPU is to set the ITK_CUDA environment
variable that should be set to 1 in order to run ITK code on GPU.
ITK image filters enabled on GPU are convolution filters; Mean, Gaussian, Derivative, Hessian of
Gaussian, Statistical filters as Median filters, and PED-based filters.

3.2 COMPUTED TOMOGRAPHY
Dmitri et al, explored cone-beam backprojection on GPU-CPU devices in computed tomography field. In
CT imaging the three dimensional reconstruction is computationally intensive. The GPU support for
pipelining and its high speed for parallel processing make it suitable for the back projection step in
image reconstruction which is the part with highest volume of computations. They utilized CPU-GPU
hybrid to accelerate backprojection. They resided projections in GPU and reconstructed each volume
slice in GPU whereas each GPU loop uploads its slice to CPU memory [2].


H. Scherl, Keck et al, proposed a method to use CUDA for fast CT reconstructions. They divide
FDK method for standard CT reconstruction in 3 steps, generating weighed projection data, ramp
filtering the projection row-wise and back projecting the filtered data into the volume. They have
used CUFFT (CUDA library for FFT) to implement filter at kernel level. Each CUDA kernel is
responsible for computing the total rows of a projection concurrently. They claim that use of
CUDA-enabled architecture is highly desired in cone-beam CT reconstruction [3].



3.3 MRI ON CUDA
Mathias Shneider, worked on MRI modality shifted on CUDA. Current MR systems has been using
parallel image acquisition techniques like GRAPPA which involves computational intensive process. For
boosting the performance it is wise to take advantage of GPUs. The autocallibration stage of GRAPPA
algorithm is done by GPU and takes about 25% of the time on CPU [4].
Gregerson implemented fast MRI gridding on GPU with use of CUDA. The rapid improvement of NVIDIA
GPUs for floating point operations leaded the medical imaging researchers to adapt their algorithms to
GPUs. The work has implemented the convolution interpolation algorithm in CUDA with optimization
strategies. The GPU device showed a performance gain [5].

3.4 IMAGE RENDERING
For high quality real-time rendering of volumes, GPU accelerated ray-casting technique has been
proposed. it is implemented as a part of OpenGL c++ library. Floating point renderers were used for full
precision. Local illumination techniques and models are computed on GPUs. For F+C rendering the
computational intensive calculations should be taken into considerations. Small memory of GPU might
become a delimiter however as much texture, render and vertex buffers can be shared.




4 COMPARISON
4.1 ITK :
Median filter speedup on CUDA ITK has been 140 x where as Median filter has been 25X, Gaussian for
60X and Anisotropic diffusion for 70 x faster than CPUs.


4.2 CT :

4.3 MRI:
Device



5 CONCLUSION

In the paper different applications of GPU in medical image processing were exposed. Different
toolboxes and incorporations taking advantage of GPUs were discussed. The overall result of each
methodology and application, the overall speedup was shown.

6 ACKNOWLEDGMENT
Special thanks to Prof. Bhandarkar for leading me through preparation of this report.
This is a review of the works done by different researchers and the author has no claim, nor novelty over
the information provided in this report.
7 REFERENCES
1. http://www.cmake.org/Wiki/ITK/Release_4/GPU_Acceleration
2. Riabkov, Dmitri, et al. "Accelerated cone-beam backprojection using GPU-CPU hardware." Proceedings of the 9th
International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine. 2007
3. Scherl, Holger, et al. "Fast GPU-based CT reconstruction using the common unified device architecture (CUDA)."
Nuclear Science Symposium Conference Record, 2007. NSS'07. IEEE. Vol. 6. IEEE, 2007.
4. http://wwwcip.informatik.uni-erlangen.de/~simtschn/thesis/thesisHref.pdf
5. http://www.nvidia.com/docs/IO/47905/ECE757_Project_Report_Gregerson.pdf
6. Anders Eklund, Paul Dufort, Daniel Forsberg, Stephen M. LaConte, Medical image processing on theMedical Image
Analysis, Volume 17, Issue 8, December 2013, Pages 1073-1094, ISSN 1361-8415,
7. http://campar.in.tum.de/twiki/pub/AMIARCS08/ProceedingsOnline/104-113.pdf
8. http://campar.in.tum.de/twiki/pub/AMIARCS08/ProceedingsOnline/104-113.pdf

Vous aimerez peut-être aussi