Room: AAPM ePoster Library
To introduce a mathematical framework for robust adaptive radiation therapy (ART) using (i) scenario-reduction to adress computational tractability, and (ii) Bayesian inference to handle adaptation cost and interfractional geometric variations following a probability distribution deviating from the prior.
In this framework, the initial and adapted plans are generated from robust optimization problems based on expected-value (EV) or worst-case (WC) optimization. Robust optimization problems can be intractable since the number of scenarios grows exponentially with the number of fractions. Scenario-reduction is used to reduce the size of the original problem by trading-off scenario-probabilities and distance of the scenario values. Thus, robust plans can be optimized in an accurate and fast manner. To evaluate the distribution of the variations for every individual case and decide whether adaptation is triggered, Bayesian inference is used to introduce mathematical-based decision-criteria, and compute the posterior distribution which is included in the robust problem to generate the adapted plans. The performance of the proposed framework is evaluated on a one-dimensional phantom for a series of simulated cases. For comparison, the simulations are also performed with strategies employing the non-adaptive approach, high adaptation frequency and/or no Bayesian inference.
The simulations indicate the following. Scenario-reduction can provide accurate robust plans. WC-optimization appears to be most compatible with Bayesian inference to create individualized plans that improve target coverage and spare the OAR. The decision-making-process can be supported by mathematical-based criteria. The methods with which variations are handled appear to have a greater impact on treatment outcome than increased adaptation frequency.
According to this proof-of-concept, scenario-reduction may be worth exploring further in both ART and robust planning. Overall, the mathematical properties of the robust optimization problem and methods with which the measured variations are processed seem to be essential, which may inform future development of ART frameworks.
Not Applicable / None Entered.