4487

High-performance 3D Compressive Sensing MRI reconstruction

Daehyun Kim, J.D. Trzasko, M. Smelyanskiy, C.R. Haider, A. Manduca, P. Dubey
Throughput Computing Lab, Intel Corporation, USA
Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2010

@inproceedings{kim2010high,

   title={High-performance 3D Compressive Sensing MRI reconstruction},

   author={Kim, D. and Trzasko, J.D. and Smelyanskiy, M. and Haider, C.R. and Manduca, A. and Dubey, P.},

   booktitle={Engineering in Medicine and Biology Society (EMBC), 2010 Annual International Conference of the IEEE},

   pages={3321–3324},

   year={2010},

   organization={IEEE}

}

Source Source   

728

views

Compressive Sensing (CS) is a nascent sampling and reconstruction paradigm that describes how sparse or compressible signals can be accurately approximated using many fewer samples than traditionally believed. In magnetic resonance imaging (MRI), where scan duration is directly proportional to the number of acquired samples, CS has the potential to dramatically decrease scan time. However, the computationally expensive nature of CS reconstructions has so far precluded their use in routine clinical practice – instead, more-easily generated but lower-quality images continue to be used. We investigate the development and optimization of a proven inexact quasi-Newton CS reconstruction algorithm on several modern parallel architectures, including CPUs, GPUs, and Intel’s Many Integrated Core (MIC) architecture. Our (optimized) baseline implementation on a quad-core Core i7 is able to reconstruct a 256x160x80 volume of the neurovasculature from an 8-channel, 10x undersampled data set within 56 seconds, which is already a significant improvement over existing implementations. The latest six-core Core i7 reduces the reconstruction time further to 32 seconds. Moreover, we show that the CS algorithm benefits from modern throughput-oriented architectures. Specifically, our CUDA-base implementation on NVIDIA GTX480 reconstructs the same dataset in 16 seconds, while Intel’s Knights Ferry (KNF) of the MIC architecture even reduces the time to 12 seconds. Such level of performance allows the neurovascular dataset to be reconstructed within a clinically viable time.
No votes yet.
Please wait...

* * *

* * *

HGPU group © 2010-2017 hgpu.org

All rights belong to the respective authors

Contact us: