Keywords:
Radiation physics, Computer applications, Cone beam CT, CT-Angiography, Physics, Computer Applications-3D, Image guided radiotherapy, Artifacts
Authors:
M. Elter, T. D. O. Lang, N. Maaß; Erlangen/DE
DOI:
10.1594/ecr2013/C-0496
Methods and Materials
An automatic geometry alignment correction method that is based on comparing redundant rays (line integrals) was implemented.
The proposed method is capable of correcting alignment errors regarding horizontal detector offsets u0,
rotation axis tilts a,
and rotation axis skews b.
Assuming a full circle scan,
many rays within the plane of the source rotation is measured twice.
The L2 norm of the measured line integrals of a ray p and its complementary ray q is used to quantify the geometry parameter quality: c = ||p - q||2.
Typically,
a standard method for multi-dimensional optimization like gradient decent would be applied to determine optimal parameters given a cost function c.
However,
these approaches are not indicated in the case of misalignment correction because the proposed cost function has many local minima.
Using a gradient decent optimization with a variable step size,
typically the closest local minimum is found but not the global minimum.
Furthermore,
a sequence of one-dimensional optimizations instead of a multi-dimensional optimization was found to be applicable for misalignment correction if the parameters are optimized in the order of their relative importance.
Therefore,
a sequence of simple hierarchical grid-searches is employed to determine the optimal horizontal detector offset,
rotation axis tilt,
and rotation axis skew.
Five CT datasets,
based on the FORBILD phantom definitions (6),
were simulated: head,
hip,
jaw,
abdomen,
and thorax.
The proposed misalignment correction approach was evaluated on these datasets.
In general a horizontal detector offset of two pixels and rotation axis tilts and skews of one degree were assumed.
In a first experiment,
individual misalignments were assumed and corrected with individual hierarchical grid searches.
In a second experiment,
joint misalignments of the three geometry parameters were assumed and jointly corrected using a sequence of hierarchical grid searches.