Accelerated ordered-subsets algorithm based on separable quadratic surrogates for regularized image reconstruction in X-ray CT

Iterative algorithms for X-ray CT image reconstruction offer the possibility of reduced dose and/or improved image quality, but require substantial compute time. Reducing the time will likely require algorithms that can be massively parallelized. Ordered subsets (OS) algorithms update all voxels simultaneously and thus are amenable to such parallelization. We present an new monotonic algorithm for regularized image reconstruction that is derived using optimization transfer with separable quadratic surrogates (SQS). The new algorithm accelerates the convergence rate by adapting reduced curvature values for the regularizer that were proposed by Yu et al. [1] for coordinate descent algorithms. We further accelerate the algorithm using ordered subsets. Simulation results show that the proposed OS algorithm converges faster than the traditionalOS algorithmforX-ray CT reconstruction from a limited number of projection views.