Evaluation of Three Deformable Image Registration Techniques Between CT and CBCT in Prostate Cancer Radiation Therapy

Abstract

Daily CBCT images can be used to track dose to critical structures in order to identify deviations from the planned treatment, however, this requires accurate contour delineation on daily images and accumulation of the daily dose. Deformable image registration (DIR) between the planning CT (pCT) and CBCT can be used to generate daily contours on the CBCT and to transfer the daily dose for dose accumulation. In this study two novel methods for DIR are compared to a previously described normalized intensity-based DIR. CT and corresponding CBCT images (totaling 124 dose fractions) from 16 subjects treated with radiation therapy for prostate cancer were analyzed. Contours for the bladder, rectum, and prostate were manually defined on each CBCT by a radiation oncologist. For each patient, all contours were deformed from the pCT to each daily CBCT using the following three DIR methods: normalized intensity based (NID), contour surface based (CSD), and multi-modality (MD). The CSD method minimizes surface differences between corresponding contours on two images. The MD method utilizes a diffusion regularized, free-form deformation algorithm to maximize the correspondence of high-dimensional feature descriptors computed for each image voxel. The deformed contours (dPC) were compared to manual reference contours on each input image by calculating dice similarity coefficients (DSC). Paired t-test p-values were used to evaluate statistical significance of the results. Average DSC across all scans and patients for all structures improved from 0.66 for NID to 0.87 for CSD and 0.74 for MD, with an overall improvement of 60.4% and 23.1%, respectively. The CSD dPC were significantly more accurate for prostate, bladder, and rectum over the other two methods (p<1.0*10-15). Though CSD performed best overall, the MD method outperformed NID, improving all three structures by an average of 23%.Abstract TU_39_3279; Table 1Average DSC (Range)NID DICECSD DICEMD DICEBladder0.70 ± 0.14 (0.17-0.93)0.88 ± 0.09 (0.47-0.97)0.78 ± 0.15 (0.28-0.94)Prostate0.67 ± 0.13 (0.28-0.90)0.87 ± 0.05 (0.71-0.96)0.72 ± 0.11 (0.35-0.93)Rectum0.61 ± 0.11 (0.19-0.80)0.84 ± 0.05 (0.72-0.94)0.71 ± 0.08 (0.45-0.88) Open table in a new tab Out of the three methods tested, the CSD method was the most accurate, with average dice falling between 0.84 and 0.88 for all three analyzed structures. However, this method requires contours to be present on both images, meaning it cannot be used for the purpose of adaptive recontouring. It can, however, be used as an adjunct to other methods. The MD method, which does not require two sets of contours, also performed well, with average dice falling between 0.71 and 0.78 for all structures. These two methods could be used in combination to achieve successful and accurate adaptive recontouring and dose transfer for accumulation and dose tracking. We plan to continue to investigate the potential of these novel DIR methods for the purpose of dose-guided adaptive therapy.