Room: Stars at Night Ballroom 1
Purpose: Creating clinically-viable radiation therapy plans can be time-consuming without prior knowledge of the dose objectivesâ€™ feasibility. The radiation therapy planning process could be significantly accelerated by utilizing a voxel-wise dose prediction model to estimate feasible doses. This study aims to generate such a machine-learning model that takes into account dose optimization objectives.
Methods: Fifteen VMAT plans were automatically created with different optimization objectives for each of ten prostate cancer patients retrospectively included for model training and evaluation. The designed model uses the patientâ€™s organs-at-risk (OARs), planning target volume (PTV), and the priorities used during plan optimization as inputs, initializes an output dose as a fitted exponential decay from the PTV as a function of distance, and optimizes fitting parameters at the beginning of training. This dose is iteratively updated using a neural network based on the previous iteration, optimization priorities, and transformations of the anatomical contours. The model optimizes root-mean-square error (RMSE) between the predicted output and real dose maps, restricted to slices containing OARs or the PTV. 10-fold cross-validation was used across the 10 patients, first with 10% as training data and 90% as testing data, and then with 90% as training data and 10% as testing data. Model performance metrics were root-mean-square error of dose and mean dose-volume histogram (DVH) error.
Results: The average RMSE was less than 3.9% of prescription dose for both ratios of training data to testing data. The average mean DVH errors for the PTV, bladder, and rectum for both ratios of training data to testing data were less than 2.8%, 2.4%, and 4.3% of prescription dose, respectively.
Conclusion: Preliminary results suggest that this voxel-wise dose prediction model can learn from small numbers of patients while achieving good prediction accuracy. This model could accelerate the radiation therapy planning process without compromising clinical feasibility.