Room: AAPM ePoster Library
Purpose: There is a strong clinical need to accelerate multi-criterial optimization (MCO) algorithms to streamline the treatment planning workflow and improve patient quality of care. This study aims to accelerate the MCO workflow using a dose prediction model which produces several Pareto-optimal plans.
Methods: Twenty SBRT treatment plans were created to represent the trade-off possibilities between PTV coverage and small bowel sparing with different optimization objectives. Twenty pancreas cancer patients were retrospectively included for model training and testing. The model first initializes an output dose as a fitted inverse decay from the PTV as a function of distance. This dose is iteratively updated using a neural network based on the previous iteration, the objective priorities used during optimization, and transformations of the anatomical contours. The model then minimizes the root-mean-square error (RMSE) between the predicted output and real dose distributions. The quality of the Pareto surface predictions was determined using the average nearest-point distance (ANPD), a recently-developed Pareto surface similarity metric. Predictions were made using an NVIDIA Quadro M4000 GPU, which is several orders of magnitude slower than typical TPS hardware.
Results: The dose distribution RMSE was 3.7% ± 0.8% and 5.3% ± 1.6% of prescription dose for the training and testing datasets, respectively, while the Pareto surface ANPD was 2.5% ± 0.8% and 3.7% ± 1.2% of prescription dose. Dose prediction took 0.22 seconds/plan, which is significantly faster than inverse optimization using a standard TPS.
Conclusion: Preliminary results suggest that this dose prediction model achieves good prediction speed and accuracy. The ANPDs suggest that the dose distribution errors are more prevalent outside of the critical structures, diminishing their effect on Pareto surface prediction quality. This model could accelerate the radiation therapy planning process by rapidly estimating the range of potential dosimetric trade-offs without compromising clinical feasibility.
Funding Support, Disclosures, and Conflict of Interest: This work is partially supported by NIH grant R01CA201212.