Room: Karl Dean Ballroom A2
Purpose: To evaluate a 3D voxel-based inverse optimization (IO) pipeline as an improvement over conventional dose mimicking (DM) methods.
Methods: We developed a novel IO pipeline to generate fluence maps using clinical dose distributions as input. The IO approach learns weights for voxel-based objectives from a dose distribution and generates a new plan using inverse planning; thus it performs dose mimicking by intelligently weighting different parts of the dose distribution automatically. Our methodology was benchmarked against a conventional dose mimicking approach that minimizes quadratic penalties in several voxel- and structure-based objective functions. IO and DM plans were constrained to the same fluence heterogeneity. We applied both approaches to 217 clinical head and neck treatment plans, and compared the resulting IO and DM plans using clinical planning criteria. We also computed clinical criteria satisfaction for the clinical plans as a baseline comparison.
Results: Clinical criteria were satisfied more frequently in the IO plans (79%) compared to the DM plans (75%). Both the IO and DM plans improved over the clinical plans (68%). While they performed the same with respect to organ at risk (OAR) criteria (74%), the IO plans dominated in target criteria satisfaction (93% vs. 84%). Every IO plan had improved OAR sparing compared to its corresponding DM plan. For example, 95% of IO plans reduced average OAR dose by at least 2.3Gy. Compared to clinical plans, IO plans delivered 7.5Gy less OAR dose on average.
Conclusion: A direct comparison of plans generated with the IO method and DM method over a large cohort showed that the IO method achieved superior plan quality. This indicates that future development of this method is warranted moving forward.