Neuro-dynamic programming for fractionated radiotherapy planning

Abstract

Summary. We investigate an on-line planning strategy for the fractionated radiotherapy planning problem, which incorporates the effects of day-to-day patient motion. On-line planning demonstrates significant improvement over off-line strategies in terms of reducing registration error, but it requires extra work in the replanning procedures, such as in the CT scans and the re-computation of a deliverable dose profile. We formulate the problem in a dynamic programming framework and solve it based on the approximate policy iteration techniques of neuro-dynamic programming. In initial limited testing the solutions we obtain outperform existing solutions and offer an improved dose profile for each fraction of the treatment.

Adaptive and robust radiation therapy optimization for lung cancer

“…Within the operations research community, several ART methods that can be considered dose-reactive have been developed. As stated above, Ferris and Voelker [2004] and Deng and Ferris [2008] apply neuro-dynamic programming (NDP) to ART, and consider NDP policies constructed from simpler heuristic policies. One such policy is a "reactive" policy that at each stage adjusts the dose to account for how much has been been delivered up to that point; our choice of the name "dose-reactive" stems from these studies.…”

“…For example, Sir et al [2010] showed that an adaptive open-loop feedback control (OLFC) policy achieves better performance than a non-adaptive OLFC policy. Ferris and Voelker [2004] and Deng and Ferris [2008] applied neuro-dynamic programming (NDP) to ART and similarly showed that their NDP policy offers an improvement over a constant, non-adaptive policy.…”

“…The dose-reactive methods we present in Chapter 5 build on the method described in Chapter 4 and thus (1) are specialized for lung cancer and breathing motion uncertainty, (2) do not make any distributional assumptions about the uncertainty in each fraction, and (3) do not assume that the nature of the uncertainty stays the same throughout the treatment. A notable difference between our work and the methods considered in Ferris and Voelker [2004] and Deng and Ferris [2008] is the decision variable in each stage: in the latter two studies, the decision for each fraction is the dose distribution to be delivered, while in our method, the decision for each fraction is the collection of beamlet intensities to use in that fraction (from which the dose distribution is determined via a linear transformation). By focusing on beamlet intensities, we allow only for dose distributions that are physically deliverable.…”

A dynamic programming approach to adaptive fractionation

“…Such adaptation is shown to improve treatment efficacy and to allow for dose escalation to the tumor [21,9]. Another approach focused on using a smaller PTV initially and re-optimizing treatment plans between fractions to compensate for the accumulated dose errors [15,5,6,18,19]. We do note that the practical applicability of this method relies on the ability to accurately determine the delivered dose at each voxel.…”

“…The DP approach is useful for sequential decision making problems, especially when there is a need for balancing the immediate and future costs associated with making a decision in any particular stage [1]. For off-line ART, the DP approach can be used to compensate for past accumulated errors in dose to the tumor [7,6,16]. For on-line ART, an approach for adaptive fractionation based on biological models also makes use of DP [3].…”