MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Deep Learning MC: Fast CNN-Based Prediction of Monte Carlo Dose for MR-Guided Treatment Planning

R Neph*, Y Huang , Y Yang , K Sheng , UCLA Radiation Oncology, Los Angeles, CA

Presentations

(Tuesday, 7/16/2019) 1:45 PM - 3:45 PM

Room: Stars at Night Ballroom 2-3

Purpose: Acceleration of planning dose calculation in settings where Monte Carlo (MC) methods are essential, such as magnetic resonance guided radiotherapy (MRgRT), would make online re-planning using automatic selection of optimal beam orientations or full-arcs practical. Slower MC is deemed necessary for MRgRT due to the complex and altered dose deposition patterns induced by electron return effects that cannot be accurately accounted for using the faster analytical methods. To significantly accelerate MC planning dose calculation for more effective MRgRT, we propose a deep learning MC (DL-MC) method using deep convolutional neural networks (CNN) that can be used in addition to traditional MC acceleration techniques.

Methods: We have developed and trained a deep CNN dose prediction model on paired high- and low-variance MC-simulated photon dose beamlets for thousands of randomly selected orientations in patient geometry, possible of any coplanar MR-guided photon treatment. The accuracy of DL-MC was compared to that of direct high-variance MC simulation on a separate testing set by calculating average normalized dose error and gamma score maps with MC-simulated low-variance dose as ground truth.

Results: DL-MC reduced the average dose error by more than 1.6 orders of magnitude, from 23.9% to 0.7%, over direct high-variance MC-simulation while reducing computation time by more than 2.7 orders of magnitude over low-variance simulation without using conventional speedup techniques. Additionally, DL-MC shows improved dose accuracy in areas where EREs induced high dose concentrations, like tissue-air interfaces (e.g. skin, trachea, and oral cavity). DL-MC also efficiently filters out low probability events commonly observed in high-variance simulation such as scattering electron trails and dose loops that must typically be averaged out by much slower low-variance simulation.

Conclusion: We have developed an efficient CNN-based DL-MC method that is capable of accurately generating low-variance photon beamlet dose for use in accelerated MR-guided photon treatment planning.

Funding Support, Disclosures, and Conflict of Interest: This investigation was supported by the following NIH grants: NIH R01CA230278, NIH R44CA183390, NIH R01CA188300

Keywords

Dose, MR, Treatment Planning

Taxonomy

TH- External beam- photons: dose computation engines- Monte Carlo

Contact Email