Accelerated Line Search for Coordinate Descent Optimization

Iterative reconstruction (IR) methods show promise for image quality improvements in helical X-ray CT. Coordinate descent methods have good convergence properties for these problems, but include a one-dimensional, nonlinear optimization for each voxel´s update. This paper presents two methods for speeding computation by replacing the 1D line search with one-step updates. Both methods greatly reduce this portion of IR computation, at no loss of convergence speed as measured by full iteration count. Experimental examples include both phantom and clinical CT scans.