| Most existing intensity-modulated radiation therapy (IMRT) optimization methods assume that the shape and location of the patient anatomy do not change throughout the treatment course and typically produce solutions containing sharp dose gradients between a tumor and its neighboring healthy tissue. Uncertainties caused by the daily setup procedures and organ motion, however, can lead to significant differences between the dose distribution as calculated by a treatment plan and the actual dose distribution delivered to a patient.;Using weighted power loss functions as a measure of performance for a treatment plan, a simple method of calculating the ideal spatial dose distribution when there is positional uncertainty is presented. Immediate results are qualitative insights into the suitability of using a margin to compensate for uncertainties at all, and (if one is to be used) how to select a "good" margin size. Moreover, it is shown that the common practice of raising power parameters in the treatment loss function, in order to enforce target dose requirements, is counter-productive. These insights into desirable dose distributions could be used, in conjunction with well-established inverse radiotherapy planning techniques, to produce dose distributions that are more adept at dealing with uncertainties.;Furthermore, techniques that exploit the dynamic nature of radiotherapy and information gathering by adapting the treatment plan to variations measured during the treatment course of an individual patient are developed. This problem, called off-line adaptive treatment planning, is structured within a dynamic programming (DP) framework. Several (suboptimal) control policies, which are obtained using an approximation to the underlying DP and computationally feasible to implement, are presented and compared. The common properties shared by these policies are that they all (1) perform a re-optimization of beamlet intensities before each fraction, (2) use Bayesian updating of the individual patient's setup variation parameters, (3) use "Multiple Instance of Geometry Approximation" as the model of uncertainty.;These policies only differ in their re-optimization schemes, of which there are three major variants: certainty equivalent control, open-loop feedback control, and cost-to-go approximation via Lagrangian relaxation.;Computational experiments show that these individualized adaptive radiotherapy plans promise considerable improvements over non-adaptive treatment plans. |
