Room: Room 209
Purpose: To develop a clinically implementable workflow for our robust dose mimicking pipeline that automatically generates treatment plans using knowledge-based planning (KBP).
Methods: First, we used a random forest KBP model to predict 3D dose distributions for each patient in a large dataset of 217 oropharynx patients using leave-one-out cross-validation. Second, the predicted dose distributions were input into a novel robust dose mimicking (RDM) method to generate plans (RDM-P plans). To facilitate a baseline comparison we also generate â€œRDM-Câ€? plans using clinical dose distributions as a surrogate for perfect KBP predictions. The RDM-P plans were constrained to the same fluence heterogeneity as RDM-C plans. We compared the predictions and RDM-P plans against the RDM-C plans using the gamma metric (3%/3mm) and clinical planning criteria.
Results: RDM-P plans were generally non-inferior to RDM-C plans, and between the plans 86% of voxels passed the gamma criteria. On average, the RDM-P plans improved the clinical criteria by 0.8Gy over the RDM-C plans; average dose to the OARs was also reduced by 1.2Gy in the RDM-P plans. Overall clinical criteria satisfaction was comparable between RDM-P and RDM-C plans (77% vs. 79%). Additionally, RDM-P plans were somewhat desensitized to prediction errors from the KBP model. For example, the 62% of voxels in the predicted dose distributions passed the gamma criteria for the RDM-C plans, but once those predictions were input to the RDM model they generated plans where 86% of the voxels passed the criteria.
Conclusion: Clinical plans represent the gold-standard for input to robust dose mimicking, but we show that KBP predictions are a viable (and clinically more reasonable) alternative. We illustrate that using either KBP predictions or clinical plans as input to RDM can generate similar high quality plans.