eduzhai > Applied Sciences > Engineering >

Accelerating MRI Reconstruction on TPUs

  • Save

... pages left unread,continue reading

Document pages: 9 pages

Abstract: The advanced magnetic resonance (MR) image reconstructions such as thecompressed sensing and subspace-based imaging are considered as large-scale,iterative, optimization problems. Given the large number of reconstructionsrequired by the practical clinical usage, the computation time of theseadvanced reconstruction methods is often unacceptable. In this work, we proposeusing Google s Tensor Processing Units (TPUs) to accelerate the MR imagereconstruction. TPU is an application-specific integrated circuit (ASIC) formachine learning applications, which has recently been used to solvelarge-scale scientific computing problems. As proof-of-concept, we implementthe alternating direction method of multipliers (ADMM) in TensorFlow toreconstruct images on TPUs. The reconstruction is based on multi-channel,sparsely sampled, and radial-trajectory $k$-space data with sparsityconstraints. The forward and inverse non-uniform Fourier transform operationsare formulated in terms of matrix multiplications as in the discrete Fouriertransform. The sparsifying transform and its adjoint operations are formulatedas convolutions. The data decomposition is applied to the measured $k$-spacedata such that the aforementioned tensor operations are localized withinindividual TPU cores. The data decomposition and the inter-core communicationstrategy are designed in accordance with the TPU interconnect network topologyin order to minimize the communication time. The accuracy and the high parallelefficiency of the proposed TPU-based image reconstruction method aredemonstrated through numerical examples.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×