Implications of Including Tumor Presence Probabilities as Weights in Radiotherapy Optimization

Abstract

<h3>Purpose/Objective(s)</h3> Recently, the fixed nature of the clinical target volume has been challenged in several studies that choose to incorporate microscopic tumor infiltration as a source of uncertainty in radiation therapy treatment planning. The approach to handle this uncertainty typically involves voxel-wise objective function weights that are proportional to each voxel's marginal probability of tumor presence. The main argument for doing so is to balance the probability of target coverage against the increased dose to organs at risk (OARs). We study the implications of basing the optimization on such probability weights, given an isotropic tumor infiltration model used for both optimization and evaluation, in which the infiltration distance is taken from a truncated gaussian probability distribution. We hypothesize that the low voxel weights far from the gross tumor volume (GTV) will cause the dose to fall off unevenly between directions that have high and low conflict with OARs, leading to a suboptimal trade-off between the probability of target coverage and OAR sparing. <h3>Materials/Methods</h3> We implement the probability weight method on a phantom geometry where suspected target voxels are weighted with probabilities from the tumor infiltration model, decreasing away from the GTV. This probability weighted plan is then compared against plans that are optimized using target volumes from an enumeration over a discrete set of scenarios from the tumor infiltration model. For both methods, plans are optimized using a range of scaling factors which scale the (uniform dose) objectives on the target in relation to those on OARs (max EUD), allowing comparison of the approximate Pareto fronts of the plans in the enumeration scheme to that of the probability weighted plan. The plans are evaluated on target samples from the tumor infiltration model which are used to compute target dose statistics. For each plan, the probability of satisfying a target coverage criterion (D98 > 0.95*D_pres.) is then plotted against the OAR objective to evaluate whether the probability weight method can produce Pareto optimal plans with respect to these criteria. In evaluation, we also include a tumor control probability (TCP) model to investigate how the methods compare on a smooth target coverage criterion. <h3>Results</h3> For a given scaling factor, the probability weighted plan is usually close to the corresponding set of plans from the enumeration scheme. However, the non-dominated subset of all generated plans consists almost entirely of plans from the enumeration scheme. Although the result holds also for TCP, there is less difference between the non-dominated subset of plans and the set of probability weighted plans. <h3>Conclusion</h3> The results indicate that finding the right combination of scaling factor and target expansion distance gives a better trade-off than the probability weight method, although efficiently doing so remains a future challenge. Meanwhile, using probability weights may serve as a practical alternative.