(Tuesday, 7/16/2019) 1:45 PM - 3:45 PM
Room: Stars at Night Ballroom 1
Purpose: To improve the quality of knowledge-based automated planning (KBAP) by developing an optimization method that generates a single plan from multiple knowledge-based planning (KBP) predictions.
Methods: We first trained two KBP methods based on generative adversarial networks (GANs) on a dataset of 130 clinical oropharyngeal treatment plans. The first, 3D-GAN, predicts the full three-dimensional dose distribution in one step, and the second, 2D-GAN, predicts the dose to each axial slice independently. Both models were used independently to predict the full 3D dose distribution for 87 out-of-sample patients. Those two predictions were then input into two different optimization models. The first model, single-prediction (SP) optimization, is a conventional approach that uses a single dose prediction to generate the final fluence-based plan. The second optimization model, multi-prediction (MP) optimization, leverages an ensemble of dose predictions to generate a single fluence-based plan. Altogether, we constructed three KBAP plans (3D-GAN+SP; 2D-GAN+SP; 3D-GAN and 2D-GAN+MP) that were benchmarked against the corresponding clinical plans using clinical planning criteria.
Results: The MP optimized plans satisfied clinical criteria more frequently (82%) than the SP optimized plans based on 3D-GAN (78%) and 2D-GAN (76%). On average, the MP optimized plans improved upon the clinical plans by 1.0 Gy over the dose-volume criteria assessed; the SP optimized plans only improved upon the criteria by 0.5 Gy and 0.6 Gy for 3D-GAN and 2D-GAN, respectively. Only the MP plans satisfied all target criteria, and on average they also improved upon the dose-volume values achieved clinically for the parotids and mandible by 1.7 Gy, which is 1.3 Gy and 1.1 Gy better than the two SP plans, respectively.
Conclusion: Our MP optimization model can generate a single plan from multiple dose predictions that is superior to the plans generated by a conventional optimization model that uses the same predictions in isolation.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
