In this paper, a novel regularization approach for (non-statistical) iterative reconstruction is developed. In our
implementation, the update equation of iterative reconstruction is based on Filtered Backprojection (FBP) and the
solution is stabilized using nonlinear regularization priors. It is well known that the usage of nonlinear regularization
priors can reduce image noise at the same time preserving image sharpness . The final noise level can be adjusted by
dedicated choice of regularization priors, regularization strength and the total number of iterations. In contrast to
conventional CT using convolution kernels, image characteristics can not be further manipulated. This might cause
artificial image texture.
We present a new class of (non-local) 3D-regularization priors, which gives us control over image characteristics similar
to that obtained with conventional CT convolution kernels. In addition, efficient noise reduction at constant sharpness is
obtained. Due to the manipulation of the low-frequency components of the regularization filter, the filter is non-local.
The regularization strength becomes a 3D-matrix with contrast-dependent entries, which gives us control over contrastdependent
sharpness. The contrast edges are estimated using a 3D Laplacian kernel. High contrast edges get a low
regularization weight and vice versa. We demonstrate the potential of noise reduction on basis of clinical CT data. Also,
it is shown, that radiation exposure to the patient can be reduced by 60% in general purpose radiological CT applications
and cardiac CT at the same time maintaining image quality. Moreover, for a 128-slice detector with 0.6 mm collimation,
it is shown, that cone-beam and spiral artifacts caused by non-exact image reconstruction can be fairly removed. Putting
all together our iterative reconstruction approach substantially improves image quality in cone-beam CT, and thus has
the potential to enter routine clinical CT.