ESTRO 37 Abstract book

S45

ESTRO 37

Conclusion With APM it is possible to quantify and automatically minimize the joint dosimetric impact of biological and physical sources of uncertainty within a probabilistic optimization process. The resulting robust treatment plans feature an automated homogenization of RBE. OC-0089 How reliable are deformable image registration methods for scanned proton 4D dose calculations? C. O. Ribeiro 1 , A. Knopf 1 , J.A. Langendijk 1 , D.C. Weber 2 , A.J. Lomax 2 , Y. Zhang 1 University Medical Center Groningen UMCG, Department of Radiation Oncology, Groningen, The Netherlands 2 Paul Scherrer Institut PSI, Center for Proton Therapy, Villigen-PSI, Switzerland Purpose or Objective Due to the dosimetric sensitivity of pencil beam scanned proton therapy (PBS-PT), respiratory-induced dosimetric impacts need to be evaluated by extensive 4D dose calculations (4DDCs). To achieve a 4DDC, deformable image registration (DIR) is mandatory for estimating deformation vector fields (DVFs) from 4D images, which represent the deformation of the structures/organs. Although geometric uncertainties induced by different DIR algorithms were investigated in the literature, errors in the resulting 4D dose distributions are often not evaluated due to the lack of a dense ground truth (GT) for validation. Therefore, we aim here to quantify the systematic and random errors induced by different DIR methods in their resulting 4DDCs by using GT-DVFs extracted from 4DMRI. Material and Methods Six different DIR methods : ANACONDA (DIR1), Morfeus (DIR2), B-splines (DIR3), Demons (DIR4), C T Deformable (DIR5), and Total Variation (DIR6), were respectively applied to nine 4DCT-MRI data sets. These 4DCT-MRI data sets consist of end-of-exhalation 3DCTs of three liver cancer patients (CTV volumes of 122, 264, and 403 cm 3 ), each modulated by consecutive GT-DVFs from 4DMRI with three different amplitudes (mean of 7.82, 20.61, and 16.88 mm). The GT-DVFs and the derived DVFs from the six DIRs were then used as input for the 4DDC. Different plan configurations (single- or three-field plans) and rescanning scenarios (single scan/rescanning) were investigated. Mean and standard deviation (SD) dose distributions of the six DIR estimated 4D plans (mean(DIR) and SD(DIR)) for each patient geometry, motion amplitude, plan configuration, and rescanning strategy were calculated. DIR induced systematic error was assessed by individually comparing the resultant mean(DIR) 4D dose distributions to those obtained with GT-DVFs. The random error gives the extent of potential variation induced by multiple DIR algorithms and was quantified with the SD(DIR) dose distribution. Results Substantial differences in 4D dose distributions among different DIR algorithms, and compared to the results using GT-DVFs, were observed (Fig. 1). The systematic error of DIR estimated 4D plans, given by the absolute difference of V95(CTV) and D5-D95(CTV) between GT and mean(DIR) dose distribution, were up to 5.01 ± 3.56 % and 5.40 ± 2.62 % respectively (Table 1). Additionally, the random error of DIR 4DDCs, quantified by the mean of the SD(DIR) dose distribution, went up to 2.55 ± 0.77 % in the CTV + 1 cm volume. Multiple-field plans or rescanning helped to decrease the systematic and random errors resulting from DIR for 4D dose distributions.

uncertainties in position and weight enables analytical calculation of the expectation value and standard deviation of RBExD. The covariance matrix allows to model arbitrary physical and biological uncertainty patterns. This in turn allows closed form probabilistic optimization based on the expectation value of the standard piece-wise quadratic objective function. Our approach is evaluated on a 1D artificial case treated with opposing fields. We assume a prescribed dose of 3GyE and a biological system with alpha x =0.1 Gy -1 and beta x =0.05 Gy -2 . Linear quadratic model parameter of carbon ions are obtained using the local effect model IV. We model 3.5% range uncertainty and a depth dependent uncertainty in α c and β c (25% at Bragg peak, less before and after).

Results Figure 2 shows a maximal deviation between APM (solid lines) and 5000 random samples (crosses) of 1.28% for the expectation value and 14.5% for the standard deviation. In comparison to conventional optimization neglecting uncertainties (figure 2 (a)), a probabilistic optimization considering biological and physical uncertainties with APM reduces the mean σ[RBExD] in the target from 0.42 to 0.23 GyE through automatic margin generation and RBE homogenization.