Purpose: Model-based iterative reconstruction (MBIR) for cone-beam CT (CBCT) yields better noise-resolution tradeoffs compared to analytic methods but carries a major increase in computational burden. We report a morphological pyramid (mPyr) framework that leverages coarse-to-fine sampling, dynamic ordered subsets, and momentum-based acceleration for speedup in MBIR in support of clinical implementation.
Methods: mPyr was implemented for penalized weighted least squares (PWLS) minimization in CBCT. The optimization is divided in a cascade of layers with evolving reconstruction parameters to yield faster evolution in the first layers, followed by layers to guarantee convergence. The framework was tested in CBCT on a robotic C-arm (Artis Zeego, Siemens) (496 projections, 100 kV, and 700 mAs) for an abdomen phantom with realistic anatomy (Kyoto Kagaku). A PWLS volume (640x800x512 voxels, 0.4 mm isotropic voxel size) was reconstructed with a four-layer pyramid in: voxel size (1.6/0.8/0.4/0.4 mm); number of subsets (M = 31/16/16/8); number of iterations (5/5/10/10); and momentum (Nesterov update). Performance was assessed in terms of computation time and root mean squared deviation (from a converged image computed with a single pyramid). Performance was compared among four variations: (i) nominal, (ii) fixed M = 16 subsets; (iii) fixed voxel size = 0.4 mm; and (iv) no momentum.
Results: The mPyr framework achieved convergence (RMSD < 2x10â?»â?´) in 120 s with RMSD = 1.7x10â?»â?´, compared to ~2800 s for non-accelerated PWLS, a 24x speedup. The method achieved larger speedup compared to the three alternatives, which showed higher RMSD at equivalent run time (~120 s): 2.4x10â?»â?´, 3.1x10â?»â?´, and 3.9x10â?»â?´, for configurations (ii), (iii), and (iv), respectively.
Conclusion: Acceleration of MBIR in CBCT to clinically acceptable runtime appears achievable through a combination of morphological pyramid, dynamic ordered subsets, and a momentum-based approach with guaranteed convergence.
Funding Support, Disclosures, and Conflict of Interest: This research was supported by academic-industry partnership with Siemens Healthineers (AX Division, Forcheim, Germany).