(Monday, 7/13/2020) 3:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Room: Track 3
Purpose: Magnetic Resonance image-guided radiotherapy (MRgRT) is developed to manage the daily changes of patient setup and internal anatomy for more effective treatment of cancer. The new treatment modality creates a demand for rapid online calculation of physically accurate dose. However, fast analytical methods cannot accurately handle electron return effects (EREs), and Monte Carlo (MC), even with GPU acceleration, is unacceptably slow for beamlet dose calculation. To significantly accelerate MC planning dose calculation for more effective MRgRT, we propose a MC post-processing method, termed DeepMCDose, that employs deep convolutional neural networks (CNNs) as a complement to traditional MC acceleration techniques.
Methods: Our novel approach combines a 3D CNN with traditional MC simulation to predict noise-free dose from paired noisy MC dose and CT intensity volumes. Existing CNN-based denoisers only address simpler full-beam prediction without considering magnetic field-induced EREs. Our CNN architecture combines distinctive features of both dose deposition and anatomy for more accurate MR-guided dose prediction. Predictive accuracy is compared to noisy MC using normalized mean absolute error (NMAE) on separate Head-and-Neck (HN) testing data and by optimized plan quality with noise-free 1.5T MC dose as ground truth.
Results: Predicted dose has NMAE of 5.687e-5 compared to 1.014e-4 for noisy dose, enabling an acceleration factor of 48.9 with only 5.7ms of additional processing per beamlet. Noisy dose is too imprecise for planning, overestimating the planning Dmax for PTV (28.2%; 22.55Gy) and left parotid (84.3%; 26.96Gy), and overestimating the truly delivered PTV Dmax (Dmean) by 13.8% (2.9%) compared to 7.7% (1.2%) for predicted dose.
Conclusion: DeepMCDose is a novel deep learning-based method offering substantial acceleration of MRgRT planning beamlet dose calculation with miniscule additional cost, alongside traditional MC simulation and acceleration techniques, demonstrating clinically feasible acceleration for solving large online-adaptive planning problems including VMAT and IMRT with automatic beam-angle selection
Funding Support, Disclosures, and Conflict of Interest: This work is supported by the following sources of funding: R44CA183390, R43CA183390, R01CA188300 and R01CA230278
Keywords
Dose, Treatment Planning, Monte Carlo
Taxonomy
TH- External Beam- Photons: Computational dosimetry engines- Monte Carlo
Contact Email
