Simulate baseline shift uncertainties to improve robustness of proton therapy treatments

Souris, Kevin [UCL] Barragan Montero, Ana Maria [UCL] Di Perri, Dario [UCL] Geets, Xavier [UCL] Sterpin, Edmond [UCL] Lee, John Aldo [UCL]

Purpose: Several studies reported both systematic and random variations of the mean position of mobile tumors from fraction to fraction. This so-called baseline shift is a major source of uncertainties for mobile targets and can jeopardize treatment quality. Unlike conventional photon therapy, the inclusion of this error in a PTV margin is inadequate in proton therapy because of the range uncertainties. Accounting for this uncertainty in a robust optimizer is much more appropriate, using for instance population-based estimations of the shifts. We developed a baseline-shift model able to automatically generate modified 4D-CT series used as uncertainty scenarios in the TPS. Methods: An average CT scan and a Mid-Position CT scan (MidPCT) of the patient at planning time are generated from a 4D-CT data. The GTV contour in the MidPCT represents the mean position of the tumor along the breathing cycle. Our model can simulate a baseline shift by generating a local deformation field that moves the tumor on all phases of the 4D-CT, without creating any non-physical artifact. The deformation field is comprised of normal and tangential components with respect to the lung wall, in order to allow the tumor to slide within the lung instead of deforming the lung surface. The deformation field is eventually smoothed in order to enforce continuity. Two 4D-CT series acquired at 1 week of interval were used to validate the model. Results: After rigid registration, a baseline shift of 9.5 mm is measured between the first- and second-week 4D-CT sets (W1-CT and W2-CT). In order to validate our model, a third 4D-CT series (BS-W1-CT) was generated from W1-CT to reproduce the measured shift (Figure 1). Water equivalent thickness (WET) has been computed for each voxel of the 3 MidPCTs and revealed that the baseline shift between W1-CT and W2-CT led to a root mean square error (RMSE) of 0.52 mm in the GTV. This WET RMSE was reduced to 0.18 mm between W2-CT and the simulated BS-W1-CT. In addition, a proton therapy plan was optimized on the average W1-CT scan and recomputed on the average W2-CT and BS-W1-CT scans. Figure 2 compares the resulting DVH for all dose distributions. The dose distribution computed on BS-W1-CT reproduces the dose degradation observed on W2-CT. The degradations for D95 are 10.6% and 11.8% for W2-CT and BS-W1-CT, respectively. Similarly, the D90 degradations are 3.2% and 4%. Conclusion: Our model was validated by comparing the WET and dosimetric deviations between a simulated scenario and the real data set. As a future work, we will use our model to automatically generate uncertainty scenarios to feed a TPS for robustness evaluation and optimization of proton therapy plans. For instance, the new 4D robust optimizer of the RayStation can easily consider multiple 4D-CT series during the optimization process.