Purpose: In kilo-voltage (kV) imaging-based motion tracking, kV images acquired during treatment are registered to digitally reconstructed radiographs (DRRs) to determine the target motion. The purpose of this study is to investigate the impact of DRR resolution on the accuracy of mutual information (MI) -based registration, and to develop a metric for measuring the uncertainty in registration.
Methods: Three paraspinal treatments were retrospectively analyzed. Before each treatment, a CBCT scan was acquired and retrospectively reconstructed with various resolutions, ranging from 0.25mm to 2.0mm voxel size. DRRs were calculated from these CBCT reconstructions for planned beam angles, resulting in different spatial resolutions, and were registered to on-treatment kV images. MI values for the kV images and the DRRs were calculated over a 4mm-by-4mm searching range. The peak in the range was identified, and the width of the peak plateau, defined as the width of curve at the magnitude of noise level below the local maximum, was considered as the registration uncertainty because with the presence of noise, registration by searching local maxima may result at any point in this region. The registration uncertainty was evaluated for DRRs with different spatial resolutions at various gantry angles.
Results: The preliminary results showed that the registration uncertainty generally narrows as DRR resolution increases. At AP/PA directions, DRRs computed from CBCT with (0.25mm)3 voxel size showed an average uncertainty of 0.4mm, while the average uncertainty increases to 0.7mm for DRRs computed from CBCT with 1mm x 1mm x 2mm voxel size. An average uncertainty of 1.0mm was observed for all DRR resolutions at lateral directions.
Conclusion: A metric was developed to evaluate the uncertainty in kV-to-DRR registration and was applied to clinical data. It reveals that DRRs with higher spatial resolution is preferable to achieve submillimeter accuracy for motion tracking in spinal treatments.