Lung tumor tracking, trajectory reconstruction, and motion artifact removal using rotational cone-beam projections

Management of lung tumor motion is a challenging and important problem for modern, highly conformal radiotherapy. Poorly managed tumor motion can lead to imaging artifacts, poor target coverage, and unnecessarily high dose to normal tissues. The goals of this dissertation are to develop a real-time localization algorithm that is applicable to rotational cone-beam projections acquired during regular (6̃0 seconds) cone-beam computed tomography (CBCT) scans, and to use these tracking results to reconstruct a tumor's trajectory, shape and size immediately prior to treatment. Direct tumor tracking is performed via a multiple template matching algorithm where templates are derived from digitally reconstructed radiographs (DRRs) generated from four-dimensional computed tomography (4DCT). Three- dimensional (3D) tumor trajectories are reconstructed by binning two-dimensional (2D) tracking results according to their corresponding respiratory phases. Within each phase bin a point is calculated approximating the 3D tumor position, resulting in a 3D phase-binned trajectory. These 3D trajectories are used to construct motion blurring functions which are in turn used to remove motion blurring artifacts from reconstructed CBCT volumes with a deconvolution algorithm. Finally, the initial direct tracking algorithm is combined with diaphragm-based tracking to develop a more robust "combined" tracking algorithm. Respiratory motion phantoms (digital and physical), and example patient cases were used to test each technique. Direct tumor tracking performed well for both phantom cases, with sub-millimeter root mean square error (erms) in the axial and tangential imager dimensions. In patient studies the algorithm performed well for many angles, but exhibited large errors for some projections. Accurate 3D trajectories were successfully reconstructed for patients and phantoms. Errors in reconstructed trajectories were smaller than the errors in the direct tracking results in all cases. The deblurring algorithm performed excellently in phantom studies. Deblurring was also effective on an example patient case, though the benefits were less stark. Finally, the combined tracking algorithm performed equally to or better than direct tumor tracking in the phantom and patient cases examined. While the preliminary results for each technique are promising, the algorithms must be tested on a larger data set with well defined ground truth to investigate potential clinical application