GPU Accelerated MRI Reconstruction Professor Kevin Skadron Computer Science, School of Engineering...

1
GPU Accelerated MRI Reconstruction Professor Kevin Skadron Computer Science, School of Engineering and Applied Science University of Virginia, Charlottesville, VA 22904 Supported by a grant from the NSF Algorithm and Approach The current speed of MRI equipment hinders the fast acquisition of data. In order to compensate for this bottleneck, a multi-channel spiral acquisition process must be implemented. This process requires an algorithm that can recover planar data across multiple channels which can become costly as image resolution increases. Drawbacks: Reconstruction of this acquisition style is extremely computationally intense. In order to even reconstruct the data in an acceptable timeframe requires multi- thousand dollar clusters. Advantages: Spiral acquisition trajectories decrease acquisition times to realtime rates (~ 50fps in some cases). If the reconstruction time can be reduced, doctor’s will be able to diagnose patients using MRI scanners while the patient is still under examination. Certain parts of the GPU are highly optimized for calculations meaningful to MRI reconstruction. These include: Matrix operations Parallel execution of data blocks High memory-read (sometimes memory- write) bandwidth Exploiting the GPU Specific Applications to MRI Reconstruction: Combining multiple channels requires solving larger matrix systems which can be represented in arrays of “pixels” Each channel can be operated on independently as long as all channels are stored in local RAM (as textures) Project Goals Reduce monetary costs of MRI Reconstruction solutions to allow more ubiquitous use in the medical community. Reduce reconstruction time to realtime rates thereby allowing doctors to more accurately and effectively diagnose patients Exploit the new NVIDIA architecture in order to determine its capabilities and uses. Introduce new parallel computing design and programming paradigms to allow for more effective implementations of existing single-core oriented code. Sean M. Arietta New Research Aim: Exploit the parallelizability of the reconstruction technique using modern Graphics Processing Units (GPU’s), allowing: Cheaper reconstruction solutions Faster reconstruction Easier maintenance By collaborating the University of Virginia Medical College, we are able to implement fast reconstruction algorithms that could one day run close to realtime. Although up until just recently a less intuitive programming practice was needed to exploit GPU architecture, the release of NVIDIA’s CUDA platform has allowed for easier implementation. Overview: Data for each channel is sent to the processing unit as frequencies derived from the response of human tissue to varied EM fields. The data is acquired in a spiral pattern and must be transformed into Cartesian coordinates before processing can begin. Each sampled point from this spirally acquired dataset is accumulated to a point in the normal 2D Cartesian grid. Now the data is in “k- space” or the space within which frequencies are encoded. A system of linear equations is built from each channel and solved to return the fully stitched image, still in k-space. Finally, the stitched image is transformed from k- space to the spatial domain to recover the imaged subject.

Transcript of GPU Accelerated MRI Reconstruction Professor Kevin Skadron Computer Science, School of Engineering...

Page 1: GPU Accelerated MRI Reconstruction Professor Kevin Skadron Computer Science, School of Engineering and Applied Science University of Virginia, Charlottesville,

GPU Accelerated MRI Reconstruction

Professor Kevin Skadron

Computer Science, School of Engineering and Applied ScienceUniversity of Virginia, Charlottesville, VA 22904

Supported by a grant from the NSF

Algorithm and Approach

The current speed of MRI equipment hinders the fast acquisition of data. In order to compensate for this bottleneck, a multi-channel spiral acquisition process must be implemented. This process requires an algorithm that can recover planar data across multiple channels which can become costly as image resolution increases.

Drawbacks: Reconstruction of this acquisition style is extremely computationally intense. In order to even reconstruct the data in an acceptable timeframe requires multi- thousand dollar clusters.

Advantages: Spiral acquisition trajectories decrease acquisition times to realtime rates (~ 50fps in some cases). If the reconstruction time can be reduced, doctor’s will be able to diagnose patients using MRI scanners while the patient is still under examination.

Certain parts of the GPU are highly optimized for calculations meaningful to MRI reconstruction. These include:

Matrix operations Parallel execution of data blocks High memory-read (sometimes memory-write) bandwidth

Exploiting the GPU

Specific Applications to MRI Reconstruction: Combining multiple channels requires solving larger matrix systems which can be represented in arrays of “pixels” Each channel can be operated on independently as long as all channels are stored in local RAM (as textures)

Project Goals Reduce monetary costs of MRI Reconstruction solutions to allow more ubiquitous use in the medical community. Reduce reconstruction time to realtime rates thereby allowing doctors to more accurately and effectively diagnose patients Exploit the new NVIDIA architecture in order to determine its capabilities and uses. Introduce new parallel computing design and programming paradigms to allow for more effective implementations of existing single-core oriented code.

Sean M. Arietta

New Research Aim: Exploit the parallelizability of the reconstruction technique using modern Graphics Processing Units (GPU’s), allowing:

Cheaper reconstruction solutions Faster reconstruction Easier maintenance

By collaborating the University of Virginia Medical College, we are able to implement fast reconstruction algorithms that could one day run close to realtime. Although up until just recently a less intuitive programming practice was needed to exploit GPU architecture, the release of NVIDIA’s CUDA platform has allowed for easier implementation.

Overview: Data for each channel is sent to the processing unit as frequencies derived from the response of human tissue to varied EM fields. The data is acquired in a spiral pattern and must be transformed into Cartesian coordinates before processing can begin. Each sampled point from this spirally acquired dataset is accumulated to a point in the normal 2D Cartesian grid. Now the data is in “k-space” or the space within which frequencies are encoded. A system of linear equations is built from each channel and solved to return the fully stitched image, still in k-space. Finally, the stitched image is transformed from k-space to the spatial domain to recover the imaged subject.