(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: The Monte Carlo (MC) algorithm has long been the gold standard for dose calculation in radiation therapy, especially in proton therapy. However due to long computational times, MC cannot be used for online dose verification. In this work, we implement a fast pseudo-MC calculation using a dual-pyramid deep learning (DL) framework.
Methods: 41 prostate patient treatment plans using two IMPT fields each were used to generate training and testing datasets. The clinical ion beam plan was calculated using the Pencil Beam (PB) algorithm in Raystation 9A with a 3 mm³ dose grid, and then re-calculated with MC for a 1 mm³ dose grid. The PB dose and corresponding CT were used as the source data and the MC dose was used as the target. The dual-pyramid network is designed for mapping two inputs to one output, one pyramid extracts features from the PB dose while the other pyramid extracts features from the CT image. After independent feature extraction, the networks are combined and refined via deconvolutional layers and attention gates, and these combined features are used to predict the pseudo-MC dose distribution. A total of 28 patients (56 beams) were used for training and 13 patients were used for testing.
Results: DL-generated and PB dose were compared to the MC calculation. Mean absolute error for DL and PB across all beams was 0.46 ± 0.05% and 0.49 ± 0.02%, respectively. Mean peak signal-to-noise ratios for DL and PB methods were 30.64 ± 1.32 and 27.39 ± 1.08. Calculation of a pseudo-MC distribution takes an average of 30 seconds, while the 1 mm³ MC calculation takes roughly 10 minutes.
Conclusion: this work, we show that a deep learning framework can be used to generate fine-resolution dose grids with near comparable accuracy to MC algorithms in only a few seconds.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
