Multiobjective decision theory for computational optimization in radiation therapy

Abstract

Machine-guided iterative optimization in radiation oncology requires ordinal or cardinal ranking of competing treatment plans. When the clinical objectives are multifaceted and incommensurable, the ranking formalism must take into account the decision maker's tradeoff strategies in a multidimensional decision space. To capture the decision processes in treatment planning, a multiobjective decision-theoretic scheme is formulated. Ranking among a group of candidate plans is based on a generalized distance metric. A dynamic metric weighting function is defined based on the state energy of the decision system, which is assumed to undergo thermodynamic cooling with iteration time. The decision maker is required to specify a baseline ranking of the objectives, which is taken to be the ground state of the decision system. This decision-theoretic formalism was applied to idealized cases in stereotactic radiosurgery and prostatic implantation, using the genetic algorithm as the optimization engine. The optimization pathways and the outcome at limited horizons indicated that the combined scheme of decision-theoretic steering and iterative optimization was robust and produced treatment plans consistent with the user's expectation. The effect of treatment uncertainties was simulated using imperfect objectives; however, certain recurring plans could be identified as optimized baseline solutions. Overall, the present formalism provides a realistic alternative to complete utility assessment or human-guided exploration of the efficient solution set.