Simultaneous beam geometry and intensity map optimization in intensity-modulated radiation therapy.

PURPOSE

In current intensity-modulated radiation therapy (IMRT) plan optimization, the focus is on either finding optimal beam angles (or other beam delivery parameters such as field segments, couch angles, gantry angles) or optimal beam intensities. In this article we offer a mixed integer programming (MIP) approach for simultaneously determining an optimal intensity map and optimal beam angles for IMRT delivery. Using this approach, we pursue an experimental study designed to (a) gauge differences in plan quality metrics with respect to different tumor sites and different MIP treatment planning models, and (b) test the concept of critical-normal-tissue-ring--a tissue ring of 5 mm thickness drawn around the planning target volume (PTV)--and its use for designing conformal plans.

METHODS AND MATERIALS

Our treatment planning models use two classes of decision variables to capture the beam configuration and intensities simultaneously. Binary (0/1) variables are used to capture "on" or "off" or "yes" or "no" decisions for each field, and nonnegative continuous variables are used to represent intensities of beamlets. Binary and continuous variables are also used for each voxel to capture dose level and dose deviation from target bounds. Treatment planning models were designed to explicitly incorporate the following planning constraints: (a) upper/lower/mean dose-based constraints, (b) dose-volume and equivalent-uniform-dose (EUD) constraints for critical structures, (c) homogeneity constraints (underdose/overdose) for PTV, (d) coverage constraints for PTV, and (e) maximum number of beams allowed. Within this constrained solution space, five optimization strategies involving clinical objectives were analyzed: optimize total intensity to PTV, optimize total intensity and then optimize conformity, optimize total intensity and then optimize homogeneity, minimize total dose to critical structures, minimize total dose to critical structures and optimize conformity simultaneously. We emphasize that the objectives that include optimizing conformity make use of the critical-normal-tissue-ring. Three tumor sites: head-and-neck, pediatric brain, and prostate are used for comparison.

RESULTS

The critical-normal-tissue-ring acts as a good device for enforcing conformity. Trends in the characteristics and quality of plans resulting from each model were observed. Attempts to reduce dose to critical structures tend to worsen PTV conformity (1.542 to 3.092) and homogeneity (1.223 to 1.984), depending on the relative size and spatial distance of the critical structures to the PTV. When the critical structures are relatively small compared with the PTV (as in the case for the pediatric brain tumor, where each is less than 15% in volume), dose reduction to critical structures is accompanied by much worse scores in conformity (2.482) and homogeneity (1.984). When the critical structures are larger, as in the case of head-and-neck (approximately 50%), the conformity and homogeneity deterioration is less significant (1.542 and 1.239, respectively). There is a clear tradeoff between homogeneity, conformity, and minimum dose to organs at risk (OARs). For head-and-neck and pediatric brain tumor, the model that minimizes total dose to critical structures and optimizes conformity simultaneously offers a compromise among these factors, resulting in reduced critical structure dose with conformal and homogeneous plans. In the prostate case, the tumor is smaller than the two large nearby critical structures, and all models provide very homogeneous PTV dose distribution. However, minimizing dose to critical structures worsens conformity, as it spreads the radiation to the area surrounding the PTV. The maximum dose to the critical structures also increases slightly. Compared with plans used in the clinic which generally have uniformly spaced beam angles, the optimal clinically acceptable plans obtained via the methods herein do not have equispaced beams. The optimal beam angles returned appear to be nonintuitive, and depend on PTV size and geometry and the spatial relationship between the tumor and critical structures.

CONCLUSIONS

The MIP model described allows simultaneous optimization over the space of beamlet fluence weights and beam and couch angles. Based on experiments with tumor data, this approach can return good plans that are clinically acceptable and practical. This work distinguishes itself from recent IMRT research in several ways. First, in previous methods beam angles are selected before intensity map optimization. Herein, we employ 0/1 variables to model the set of candidate beams, and thereby allow the optimization process itself to select optimal beams. Second, instead of incorporating dose-volume criteria within the objective function as in previous work, herein, a combination of discrete and continuous variables associated with each voxel provides a mechanism to strictly enforce dose-volume criteria within the constraints. Third, using the construct of critical-normal-tissue-ring within the objective function can enhance the achievement of conformal plans. Based on the three tumor sites considered, it appears that volume and spatial geometry with respect to the PTV are important factors to consider when selecting objectives to optimize, and in estimating how well suited a particular model is for achieving a specified goal.