Meeting Banner
Abstract #0611

Accelerating Compressed-Sensing-Based DCE-MR Image Reconstruction with GPU

Jiangsheng Yu1, Yiqun Xue2, Hee Kwon Song2

1Toshiba Medical Research Institute USA, Cleveland, OH, United States; 2University of Pennsylvania, Philadelphia, PA, United States

Synopsis: Temporally constrained reconstruction based on compressed sensing (CS) has recently been developed for dynamic MR imaging to obtain high temporal resolution without losing image quality. The intensive computation overhead in CS reconstruction has limited the application for clinical data processing where large data sets are generated from multi-slice and multi-channel acquisition. The current work presents a parallelized GPU implementation to accelerate the CS-based image reconstruction in DCE-MRI. The forward and backward gridding operations, which are the most-time consuming part of the conjugate gradient searching, is addressed with a radial-point driven parallelization approach by assigning a thread for each radial point operation. A comparison with the C++ sequential implementation shows an acceleration factor of ~15 was achieved on a moderately GPU-powered laptop computer.

Keywords

accelerated accelerating acceleration access accurate achieve achieved acknowledgments acquisition adjacent advantages agent aims alternatively although application arises arrow assessment assigned atomic avoid background backtracking backward blue called cancer channel clinical compressed computation computer conjugate constrained consuming contrast convolution cores cost denote denotes determination distributes divided driven dynamic effective enhanced enhancement evaluate except fact fidelity fifth forth forward frame frames function functions golden gradient grants graphics implementation increment indexed injection intensity intensive investigated iteratively kernel laptop limited marked measures memory minimize need neighboring norms operation operator optimization overhead overlapping parallel parallelization parallelized part piece powered powerful precise prevents previously problem process processing profiling program quality radial reconstruct reconstructed reconstruction removes reported representative resolution response sampling search searching sensing sequential sequentialize series sets simple slice slices song space spaces sparsity spatial subtraction synchronization temporal temporally term therapeutics thread threads took tool tumor unit utilize utilizing versa vice view views wait whose wise yields