Premium
SU‐E‐J‐49: Evaluation of Deformable Image Co‐Registration in Adaptive Dose Painting by Numbers for Head and Neck Cancer
Author(s) -
Berwouts D,
Olteanu A,
De Neve W,
Vercauteren T,
De Gersem W,
Madani I
Publication year - 2011
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.3611817
Subject(s) - nuclear medicine , image registration , medicine , head and neck cancer , artificial intelligence , region of interest , hounsfield scale , radiation therapy , voxel , thresholding , computer science , computed tomography , radiology , image (mathematics)
Purpose: To evaluate the accuracy of automated contour deformation for head‐and‐neck cancer in adaptive treatment. Methods: Data from 13 head‐and‐neck patients in a phase I trial for adaptive treatment were used. Adaptation was based on [18F]FDG‐PET‐guided dose painting by numbers (DPBN) plans. Each patient had two DPBN plans based on: (i) a pretreatment PET/CT scan and (ii) a during‐treatment PET/CT scan acquired after 8 fractions. Contours manually drawn on the pretreatment CT scan were deformed using commercial deformable image registration software onto the during‐treatment CT scan. Deformed contours of regions of interest (ROIdef) were visually inspected by an experienced radiation oncologist and, if necessary, adjusted (ROIdef_ad) and both sets of contours were compared to manually redrawn ROIs (ROIm) using Jaccard (JI) and overlap indices (OI). ROI indices and volumes were compared for all contour sets used a paired t‐test and one‐way ANOVA pairwise comparison, respectively. Results: Almost all deformed ROIs in all patients required adjustment after visual inspection. The largest adjustments were made in GTVs when substantial tumor regression occured, e.g., ROIdef=9.2 cm3 vs. ROIdef_ad=2.2 cm3 vs. ROIm=2.1 cm3. The swallowing structures were the most frequently adjusted ROIs. The mandible was the most acurately propagated ROI requiring little or no adaptation: JI=0.7 and OI=0.8. The upper esophageal sphincter was the worst propagated ROI: JI=0.3 and OI=0.3 for the ROIdef, JI=0.5 and OI=0.6 for the ROIdef_ad. Despite the variation in indices, there was no statistically significant difference between ROIdef, ROIdef_ad and ROIm volumes. Generating ROIm took 4–6 hours, generating ROIdef took a few minutes and generating ROIdef_ad took less than 2 hours. Conclusions: Deformable image co‐registration followed by visual inspection does require adjustment of most deformed ROIs. Nevertheless, fast automatic ROI propagation followed by user‐driven adjustments appears to be more efficient than labor intensive de‐novo re‐contouring.