Quantitative deformable image registration uncertainty framework for proton head and neck treatment
PD-0400
Abstract
Quantitative deformable image registration uncertainty framework for proton head and neck treatment
Authors: Florian Amstutz1,2, Peter G D'Almeida1,3, Xin Wu1,4, Francesca Albertini1, Barbara Bachtiary1, Damien C Weber1,5,6, Jan Unkelbach5, Antony J Lomax1,7, Ye Zhang1
1Paul Scherrer Institut, Center for Proton Therapy, Villigen, Switzerland; 2ETH Zurich, Department of Physics, Zurich, Switzerland; 3ETH Zurich, Department of Information Technology & Electrical Engineering, Zurich, Switzerland; 4ETH Zurich, Department of Information Technology & Electrical Engineering , Zurich, Switzerland; 5University Hospital Zurich, Department of Radiation Oncology, Zurich, Switzerland; 6University Hospital Bern, Department of Radiation Oncology, Bern, Switzerland; 7ETH Zurich, Department of Physics, Villigen, Switzerland
Show Affiliations
Hide Affiliations
Purpose or Objective
Deformable image registration (DIR) is one key component for adaptive radiotherapy treatments (e.g. dose accumulation), but discrepancies among different DIRs are inescapable [Nenoff et al. 2020]. To overcome one barrier for its broader clinical implementation, we developed a quantitative uncertainty evaluation and estimation framework at both geometric and dosimetric levels and validated its effectiveness and accuracy for pencil beam scanning proton therapy (PT) for head and neck (HN) cancer.
Material and Methods
For five HN cancer patients with pronounced anatomy variations during PT, five different DIR algorithms (namely Elastix, Plastimatch Bspline & Demon, Velocity, NiftyReg) were used to establish geometric correspondences (represented by deformable vector fields (DVF)) between the planning CT (pCT) and each of 4-8 CTs acquired during the treatment course (repCT). A geometric quality assurance (gQA) of DVFs for each DIR was achieved by calculating the Dice coefficient between the binary image of the patient surface from repCT, with respect to propagated surfaces of the pCT using the derived DVFs (Fig. 1a). Accumulated dose distributions were computed by averaging over the warped dose distributions recalculated on each repCT based on their individual DVFs. Dosimetric uncertainties were quantified as voxel-wise dose differences resulting from the accumulated dose distributions between the different DIRs (Fig. 1b). Moreover, we also compared actual dosimetric uncertainties to model-based estimates using a previously developed approach [Amstutz et al. 2021]. This validation was performed for both, first series of sequential boost and simultaneous integrated boost (SIB) plans (Fig. 1c).
Results
The gQA results are summarized as boxplots in Fig. 2a. Four out of five DIRs gave acceptable Dice scores, while one (NiftyReg) performing even worse than rigid registration. This algorithm was therefore eliminated from the subsequent dosimetric analysis, shown in Fig. 2b, where substantial differences in the accumulated dose DVHs can be seen for the patient with the largest anatomical changes. Additionally, dose indexes for V95-PTV and left parotid D20 for all patients show pronounced differences, highlighting the importance of dose variations in accumulated dose due to choice of DIR. Finally, our model-based estimation of these uncertainties resulted in >90% voxel with a prediction error < 5%, which was comparable (91.6% vs. 92.4%) for both sequential and SIB plans (Fig. 2c).
Conclusion
A comprehensive assessment framework for geometric and dosimetric DIR uncertainties was developed and validated for HN patients. DIR induced dosimetric uncertainties for dose accumulation of PT for HN cases are substantial and potentially of clinical relevance. However, the model-based estimation provides a good estimate of these uncertainties, and more accurately than previous validations for the lung case [Amstutz et al. 2021].