Stable and Robust Geometric Self-Calibration for Cone-Beam CT Using Mutual Information

We present a novel geometric calibration method for cone beam computed tomography, which is based on an iterative multiresolution 2-D-3-D image registration framework. Our technique does not require a dedicated calibration phantom or any other specific markers inside the scanned object, so that the geometric parameters can be computed online from the acquired data itself. In contrast to existing registration-based calibration methods, our approach uses the mutual information for maximization of the similarity between the original and the reprojected projections. Furthermore, we incorporated a consistent volume update scheme into our algorithm that supports the registration to accurately retrieve the geometrical misalignment of the scanner. In addition to this, stochastic ray sampling in combination with forward and backprojection operators implemented on the graphics card make the overall runtime comparable to current registration-based approaches. However, we can reach as accurate results as the current phantom-based techniques. Our results will show that our algorithm converges stably to the correct solution, that it is robust against projection noise and that it is independent of the underlying system geometry. It can be applied to circular and to helical trajectories with arbitrary cone angles. Moreover, our studies on simulated and on real data demonstrate that our algorithm can deal with arbitrary imaging objects.