ESTRO 2024 - Abstract Book

S3712

Physics - Dose prediction, optimisation and applications of photon and electron planning

ESTRO 2024

Material/Methods:

The developed robust optimization algorithm is based on the stochastic approach to robust planning. Stochastic planning aims at minimizing the expectation value of the cost function over the set of error scenarios.

This definition of the cost function can be rephrased as the sum of two terms [2]. The first term consists in the evaluation of the chosen cost function over the expectation value of the dose distribution. The second term accumulates instead the variance of the dose distribution over all voxels within the structure. The key advantage of such approach is the possibility to pre-compute both the expected dose-influence matrix and the total variance-influence matrix. This way, there is no need to hold in memory the large dose-influence matrices for each of the error scenarios. Additionally, evaluation of the cost function at each iteration step during optimization is achieved via a limited amount of matrix products. This separation of the cost function allows also for an independent leverage of the optimization outcome. The relative weight of the expectation and variance components can indeed be adjusted to enhance the reduction of uncertainty within selected structures. The memory saving capabilities of this approach are particularly interesting when robust optimization is applied to a 4DCT data-sets. In this case a stochastic optimization approach can easily become unfeasible due to the high number of error scenarios. This approach was implemented within the treatment planning toolkit "matRad" [3]. A 4DCT lung cancer patient data-set was selected and consists of 10 different planning CTs, each representing a different phase of the breathing cycle. A CTV was delineated for every CT phase, and allowed for the definition of the ITV. A dose prescription of 1 Gy/fraction was assigned to the ITV and maximum a dose reference of 0.6 Gy/fraction was set for the organs at risk (OAR). Margin-based optimization was carried out with two different target definitions. First a 2 mm margin was added to the ITV on the nominal CT. As a second approach, each CT phase was included with the same margin for the CTV. Squared-deviation and squared-overdosing cost functions were applied respectively to the target and the OARs. Finally, 10 random setup error scenarios (2mm σ) were sampled for each of the CT phases, resulting in a total amount of 100 scenarios. This allowed for the evaluation of different DVH metrics as D95, D50 and D5 together with their standard deviation.

Both the new approach and a stochastic algorithm were then applied to optimize CTV coverage. The DVH analysis was repeated for both robust plans and compared to the margin-based optimizations.

The average time-per-iteration required for solving the optimization problem was also measured. Calculations were performed on an AMD Ryzen 9 7950X with 64GB RAM.

Results:

Figure 1 reports a comparison between the obtained DVHs for the target and OARs. Both the robust optimization approaches lead to comparable results and sufficient target coverage. With respect to the margin based approaches, significant sparing of the OARs is achieved.