Purpose: Evaluate the accuracy of atlas-based auto-segmentation and intra-patient deformable propagation of organs-at-risk (OARs) on CT and CBCT for head-and-neck (HN) adaptive radiotherapy (ART).
Methods and Materials: Six patients were included in this study. Three images of each patient were selected: a pre-treatment planning CT, an in-treatment planning CT and a CBCT. The selected in-treatment planning CT and CBCT were acquired on the same day (i.e., treatment day 22nd), and typically within a 1-2 hour time interval. Seven OARs were manually contoured on the pre-treatment and in-treatment CTs. Manual contours on the in-treatment CT were copied onto the CBCT using local rigid-body registration due to the inherent inferior image quality of CBCT. Three segmentation procedures involved in an offline HN ART process were evaluated, including 1) initial auto-segmentation on the pre-treatment planning CT, 2) intra-patient CT-to-CT deformable propagation and 3) intra-patient CT-to-CBCT deformable propagation. The auto-contours were compared with the manual contours on CTs and the copied contours on CBCT to evaluate their accuracies. Dice similarity coefficients (DSC) and mean surface distances (MSD) were calculated for evaluation.
Results: For initial auto-segmentation, the mean DSC and mean MSD of most OARs were well above 0.8 and less than 2 mm, respectively. For intra-patient CT-to-CT deformable propagation, the mean DSC of most OARs were >0.85 except for right submandibular gland. The mean MSD for CT-to-CT propagation ranged from 0.5 mm to 1.3 mm for difference OARs. Compare to CT-to-CT, the mean DSC for intra-patient CT-to-CBCT deformable propagation decreased by ~0.03 (range, 0.01-0.06) and the mean MSD increased by ~0.2 mm (range, 0.1-0.4 mm).
Conclusions: Atlas-based auto-segmentation and intra-patient deformable propagation can generate OAR structures on CT and CBCT with clinically acceptable accuracy.