Program Information
Bayesian Optimization for Automatic 4π Planning
A Landers*, D O'Connor , D Ruan , F Scalzo , K Sheng , UCLA School of Medicine, Los Angeles, CA
Presentations
TU-C2-GePD-J(B)-5 (Tuesday, August 1, 2017) 10:00 AM - 10:30 AM Room: Joint Imaging-Therapy ePoster Lounge - B
Purpose: 4π non-coplanar radiotherapy uses integrated beam selection and fluence map optimization, offering high potential to be automated, but the individual organ weightings still need to be manually tuned for clinically optimal results. In this study, Bayesian optimization (BO) is utilized to automate 4π hyperparameter tuning.
Methods: The study is divided into two components. First, “ultimate” 4π plans using all non-colliding non-coplanar beams, typically 300-700, were manually created for 20 prostate patients. Based on these plans, expected doses of each individual patient were predicted using spectral regression (SR) based on OAR voxel features including various components of Euclidean distance from the PTV, OAR and PTV size, and prescription dose. Secondly, Bayesian optimization (BO) used an objective function encouraging the 4π dose to meet or improve the SR expected doses. The BO applies a prior probability distribution over the 4π objective to find a set of hyperparameters with the highest expected improvement of the objective. After evaluating the objective with the chosen set of hyperparameters, BO then uses all data from previous evaluations to select the next set. BO is performed until no further improvement is observed. The final BO dose is compared to the clinical VMAT dose.
Results: 7 BO plans have been evaluated so far, running for an average of 375 iterations. BO plans met all clinical dose constraints. Although the VMAT plan doses were manually created and clinically treated, BO was able to improve the bladder V20 (-15.1%), right femoral head V16 (-61.9%), and rectum V20 (-56.7%), V32 (-32.6%), and V36 (-20.9%) on average.
Conclusion: Based on a set of existing plans for the prostate, BO is able to automatically create new clinically acceptable and high quality plans without manual intervention. The method is highly scalable to include new patients and helps achieve fully automated plan quality improvement.
Funding Support, Disclosures, and Conflict of Interest: DOE DE-SC0017057 NIH R44CA183390 NIH R01CA188300 NIH R43CA183390 NIH U19AI067769
Contact Email: