MO‐C‐17A‐13: Uncertainty Evaluation of CT Image Deformable Registration for H&N Cancer Adaptive Radiotherapy

Abstract

Purpose:To evaluate uncertainties of organ specific Deformable Image Registration (DIR) for H&N cancer Adaptive Radiation Therapy (ART).Methods:A commercial DIR evaluation tool, which includes a digital phantom library of 8 patients, and the corresponding “Ground truth Deformable Vector Field” (GT‐DVF), was used in the study. Each patient in the phantom library includes the GT‐DVF created from a pair of CT images acquired prior to and at the end of the treatment course. Five DIR tools, including 2 commercial tools (CMT1, CMT2), 2 in‐house (IH_FFD1, IH_FFD2), and a classic DEMON algorithms, were applied on the patient images. The resulting DVF was compared to the GT‐DVF voxel by voxel. Organ specific DVF uncertainty was calculated for 10 ROIs: Whole Body, Brain, Brain Stem, Cord, Lips, Mandible, Parotid, Esophagus and Submandibular Gland. Registration error‐volume histogram was constructed for comparison.Results:The uncertainty is relatively small for brain stem, cord and lips, while large in parotid and submandibular gland. CMT1 achieved best overall accuracy (on whole body, mean vector error of 8 patients: 0.98±0.29 mm). For brain, mandible, parotid right, parotid left and submandibular glad, the classic Demon algorithm got the lowest uncertainty (0.49±0.09, 0.51±0.16, 0.46±0.11, 0.50±0.11 and 0.69±0.47 mm respectively). For brain stem, cord and lips, the DVF from CMT1 has the best accuracy (0.28±0.07, 0.22±0.08 and 0.27±0.12 mm respectively). All algorithms have largest right parotid uncertainty on patient#7, which has image artifact caused by tooth implantation.Conclusion:Uncertainty of deformable CT image registration highly depends on the registration algorithm, and organ specific. Large uncertainty most likely appears at the location of soft‐tissue organs far from the bony structures. Among all 5 DIR methods, the classic DEMON and CMT1 seem to be the best to limit the uncertainty within 2mm for all OARs.Partially supported by research grant from Elekta.