Inter-fraction variations in motion modeling using patient 4D-cone beam CT images

The purpose of this paper is to quantify the variability of patient-specific motion models derived from 4D cone beam CT (4DCBCT) images captured at different radiation therapy treatment fractions. Motion models are derived from 4DCBCT images by 1) applying deformable image registration on each 4DCBCT image with respect to a reference image, resulting in a set of displacement vector fields (DVFs), and 2) performing dimensionality reduction, principal component analysis (PCA) in this work, on the DVFs to derive a motion model. PCA-based motion models are composed of a few (up to five) eigenvectors associated with eigenvalues sorted in descending order. To quantify inter-fraction variability in motion models, the motion model from the first treatment fraction is compared to motion models from each successive treatment fraction. The eigenvectors are compared using two criteria; the absolute variation and the directional similarity. Three clinical patient treatments were obtained retrospectively and used in this study. Results showed that the first two eigenvectors have smaller absolute variation (1.8 × 10
^-4
) and larger directional similarity (0.47) than the last three eigenvectors (absolute variation: 2.5 ×10
^-4
, directional similarity: 0.04). The study showed that motion models derived from different treatment fractions have variations but the eigenvectors associated with the greatest eigenvalues were shown to be more stable across treatment fractions than others. Further studies are planned to determine whether the inter-fraction variations may lead to significant changes in motion reconstruction.