Poster template
Sparse Parallel Acceleration for full 3D Golden Angle Radial Whole Breast DCE-MRI 1 Saleem ,
2 Kompan ,
3 Mann ,
3 Vreemann
1 Vicari
M.A. I.N. R. S. , M. 1Bremen,DE,Fraunhofer MEVIS, Medical Physics, 2Heidelberg,DE,Mediri Gmbh, Medical Imaging Research Institute, 3Nijmegen,NL,Radboud University Nijmegen, Diagnostic Image Analysis Group Introduction •High spatial and temporal resolution is required to access changes in Contrast Agent (CA) uptake in DCE MRI and can be achieved by golden angle radial Sparse Parallel (GRASP) MRI technique[1]. •This work aims to extend the GRASP technique to full 3D golden angle (GA) radial acquisitions [2] (Figure 1) in order to allow for an optimized retrospective adaption of spatiotemporal resolution to improve model fitting [3]. A high temporal resolution is required when the curve is quickly rising (upslope) while a high spatial resolution is required when it plateaus (washout).
Results Figure 2 shows results acquired after 24 iterations of 3D GRASP. Improved temporal resolution can be appreciated in Figure 3.
Figure 1: Full 3D golden
angle radial acquisition of k-space
Figure 4: CA uptake curve for different weightings: The plot for = 0.01M0 follows the NUFFT curve more faithfully and consequently is less smoother than those for higher values of
Materials and methods • Sparse parallel MRI is obtained by minimizing the following objective function:
Conclusions Figure 3: Various temporal frames for a 32x accelerated Figure 2: 3D GA radial whole breast MRI. Each row shows three different
The L2 minimization provides coil by coil data consistency. The undersampled Fourier operator F was based on the NUFFT toolbox [4] and coil sensitivity maps S were computed from the full dataset by the phased array adaptive reconstruction technique (5, 6). The L1 minimization enhances image sparsity in the temporal total variation domain ·D. The regularization factor λ controls the tradeoff between data consistency and image sparsity. • A whole breast 3D GA radial DCE-MRI data was acquired at 3T (Skyra, Siemens Erlangen Germany) at the Radboud University Nijmegen, as an additional non-diagnostic scan of a breast screening protocol during contrast agent inflow. Scan parameters were: TR/TE=5.19/2.25ms, voxel size=1.17x1.17x1.17 mm3, FA=15o, BW=500Hz/pixel and number of profiles=25000. • Acquisition time for undersampling factor of 32x as to the Nyquist criteria (λ = 0.05M0) was 1hr 48mins (4 core i7 4th generation, 8GB RAM, Windows operating system) after coil compression technique [7]. This acceleration factor is twice that used in original GRASP implementation and serves to compare the efficiency of recovery by 3D GRASP.
slices of a) the full dataset, b) and c) a 32x undersampled dataset with only NUFFT and with the 3D GRASP reconstruction, respectively. 3D GRASP reconstruction removes undersampling artifacts and restores spatial resolution efficiently.
whole-breast full 3D GA radial MRI. 3D GRASP reconstruction (b) allows for an improved visualization in comparison with the NUFFT only reconstruction (a) with temporal resolution Δt=4.17s. The arrows show enhancement in the heart due to contrast agent uptake
Table 1 shows results from RMSE (root mean square error) analysis done on some selected case. The error reduces for all acceleration factors after applying 3D GRASP. Figure 4 shows Contrast Agent uptake curves plotted at different values of . Higher value of increases the sparsity constraint and leads to a stronger denoising. Acceleration Factor 18x 32x 50x
Before 3D GRASP 2.944 3.799 4.829
= 0.01 M0 2.231 -
= 0.05 M0 2.366 2.287 2.143
= 0.09 M0 2.230 -
References [1] Li Feng et al. Golden angle radial sparse parallel mri: combination of compressed sensing, parallel imaging, and golden angle radial sampling for fast and flexible dynamic volumetric mri. Magnetic resonance in medicine [2 I. N. Kompan and et al, Retrospective Resolution Adaption for DCE MRI Using 3D Golden Angle Radial Acquisition, Proceedings of the 23rd ISMRM AM, Toronto, Canada, 2015 [3] Alan Jackson et al. Dynamic contrast-enhanced magnetic resonance imaging in oncology. Medical radiology. Springer, Berlin, 2005. ISBN9783540423225. [4] Jeffrey A. Fessler and Bradley P. Sutton. Non uniform fast fourier transforms using minmax interpolation. Signal
Table 1: RMSE analysis for different cases: A significant decrease in error is seen for all acceleration factors after implementing 3D GRASP
•A temporal resolution of 4.17s was achieved at 32x acceleration factor with image quality comparable to original fully sampled image. The true limit of pushing spatiotemporal resolution can be even beyond these preliminary results. •Improvement in the technique can be expected by incorporating various combinations of sparsity transforms such as discrete wavelet transform, temporal Fourier transform etc.
• Acquisition time was reduced by coil compression and can be further improved by using more powerful workstation in addition to code parallelization.
Processing, IEEE Transactions on, 51(2):560574, 2003. [5] D. O. Walsh, A. F. Gmitro, and M. W. Marcellin. Adaptive reconstruction of phased array MR imagery. Magnetic resonance in medicine [6] Mark A. Griswold, D. Walsh, Robin M. Heidemann, Axel Haase, and Peter M. Jakob. The use of an adaptive reconstruction for array coil sensitivity mapping and intensity normalization, 2002. [7] F. Huang, S. Vijayakumar, Y. Li, S. Hertel, and G. R. Duensing. A software channel compression technique for faster reconstruction with many channels. Magnetic resonance imaging, 26(1):133–141, 2008. ISSN 0730-725X. doi: 10.1016/j.mri.2007.04.010.
2 Kompan ,
3 Mann ,
3 Vreemann
1 Vicari
M.A. I.N. R. S. , M. 1Bremen,DE,Fraunhofer MEVIS, Medical Physics, 2Heidelberg,DE,Mediri Gmbh, Medical Imaging Research Institute, 3Nijmegen,NL,Radboud University Nijmegen, Diagnostic Image Analysis Group Introduction •High spatial and temporal resolution is required to access changes in Contrast Agent (CA) uptake in DCE MRI and can be achieved by golden angle radial Sparse Parallel (GRASP) MRI technique[1]. •This work aims to extend the GRASP technique to full 3D golden angle (GA) radial acquisitions [2] (Figure 1) in order to allow for an optimized retrospective adaption of spatiotemporal resolution to improve model fitting [3]. A high temporal resolution is required when the curve is quickly rising (upslope) while a high spatial resolution is required when it plateaus (washout).
Results Figure 2 shows results acquired after 24 iterations of 3D GRASP. Improved temporal resolution can be appreciated in Figure 3.
Figure 1: Full 3D golden
angle radial acquisition of k-space
Figure 4: CA uptake curve for different weightings: The plot for = 0.01M0 follows the NUFFT curve more faithfully and consequently is less smoother than those for higher values of
Materials and methods • Sparse parallel MRI is obtained by minimizing the following objective function:
Conclusions Figure 3: Various temporal frames for a 32x accelerated Figure 2: 3D GA radial whole breast MRI. Each row shows three different
The L2 minimization provides coil by coil data consistency. The undersampled Fourier operator F was based on the NUFFT toolbox [4] and coil sensitivity maps S were computed from the full dataset by the phased array adaptive reconstruction technique (5, 6). The L1 minimization enhances image sparsity in the temporal total variation domain ·D. The regularization factor λ controls the tradeoff between data consistency and image sparsity. • A whole breast 3D GA radial DCE-MRI data was acquired at 3T (Skyra, Siemens Erlangen Germany) at the Radboud University Nijmegen, as an additional non-diagnostic scan of a breast screening protocol during contrast agent inflow. Scan parameters were: TR/TE=5.19/2.25ms, voxel size=1.17x1.17x1.17 mm3, FA=15o, BW=500Hz/pixel and number of profiles=25000. • Acquisition time for undersampling factor of 32x as to the Nyquist criteria (λ = 0.05M0) was 1hr 48mins (4 core i7 4th generation, 8GB RAM, Windows operating system) after coil compression technique [7]. This acceleration factor is twice that used in original GRASP implementation and serves to compare the efficiency of recovery by 3D GRASP.
slices of a) the full dataset, b) and c) a 32x undersampled dataset with only NUFFT and with the 3D GRASP reconstruction, respectively. 3D GRASP reconstruction removes undersampling artifacts and restores spatial resolution efficiently.
whole-breast full 3D GA radial MRI. 3D GRASP reconstruction (b) allows for an improved visualization in comparison with the NUFFT only reconstruction (a) with temporal resolution Δt=4.17s. The arrows show enhancement in the heart due to contrast agent uptake
Table 1 shows results from RMSE (root mean square error) analysis done on some selected case. The error reduces for all acceleration factors after applying 3D GRASP. Figure 4 shows Contrast Agent uptake curves plotted at different values of . Higher value of increases the sparsity constraint and leads to a stronger denoising. Acceleration Factor 18x 32x 50x
Before 3D GRASP 2.944 3.799 4.829
= 0.01 M0 2.231 -
= 0.05 M0 2.366 2.287 2.143
= 0.09 M0 2.230 -
References [1] Li Feng et al. Golden angle radial sparse parallel mri: combination of compressed sensing, parallel imaging, and golden angle radial sampling for fast and flexible dynamic volumetric mri. Magnetic resonance in medicine [2 I. N. Kompan and et al, Retrospective Resolution Adaption for DCE MRI Using 3D Golden Angle Radial Acquisition, Proceedings of the 23rd ISMRM AM, Toronto, Canada, 2015 [3] Alan Jackson et al. Dynamic contrast-enhanced magnetic resonance imaging in oncology. Medical radiology. Springer, Berlin, 2005. ISBN9783540423225. [4] Jeffrey A. Fessler and Bradley P. Sutton. Non uniform fast fourier transforms using minmax interpolation. Signal
Table 1: RMSE analysis for different cases: A significant decrease in error is seen for all acceleration factors after implementing 3D GRASP
•A temporal resolution of 4.17s was achieved at 32x acceleration factor with image quality comparable to original fully sampled image. The true limit of pushing spatiotemporal resolution can be even beyond these preliminary results. •Improvement in the technique can be expected by incorporating various combinations of sparsity transforms such as discrete wavelet transform, temporal Fourier transform etc.
• Acquisition time was reduced by coil compression and can be further improved by using more powerful workstation in addition to code parallelization.
Processing, IEEE Transactions on, 51(2):560574, 2003. [5] D. O. Walsh, A. F. Gmitro, and M. W. Marcellin. Adaptive reconstruction of phased array MR imagery. Magnetic resonance in medicine [6] Mark A. Griswold, D. Walsh, Robin M. Heidemann, Axel Haase, and Peter M. Jakob. The use of an adaptive reconstruction for array coil sensitivity mapping and intensity normalization, 2002. [7] F. Huang, S. Vijayakumar, Y. Li, S. Hertel, and G. R. Duensing. A software channel compression technique for faster reconstruction with many channels. Magnetic resonance imaging, 26(1):133–141, 2008. ISSN 0730-725X. doi: 10.1016/j.mri.2007.04.010.
