Bayesian Optimization in Treatment Planning of High Dose Rate Brachytherapy
H Jafarzadeh1*, X Mao 2, S Abbasinejad Enger1,3, (1) Medical Physics Unit, McGill University, Montreal, QC, CA (2) Mila - Quebec AI Institute, University of Montreal, Montreal, QC, CA (3) Lady Davis Institute for Medical Research, Jewish General Hospital, Montreal, QC, CA
Presentations
SU-H400-IePD-F7-4 (Sunday, 7/10/2022) 4:00 PM - 4:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 7
Purpose: Selection of organ penalty weights to obtain clinically acceptable dose volume histogram (DVH) metrics in inverse treatment planning process of high dose rate (HDR) brachytherapy is a time-consuming task and plan quality depends on planner’s experience. We investigated multi objective Bayesian optimization (MOBO), which is a reinforcement learning based approach, to automatically find the penalty weights for HDR prostate cancer plans.
Methods: The treatment planning problem was formulated as a function of weights to the DVH metrics (f(w)). The penalty weights were used in the linear objective function of the Fast Mixed Integer Optimiser (FMIO), which solves the dwell time optimization problem with weights for the planning target volume (PTV), urethra, rectum, and bladder. The DVH metrics are D90(PTV), D0.1cc(urethra), D1cc(rectum), and D1cc(bladder). The weights range from 0.001 to 1.0. MOBO aims to find the Pareto surface with as few calls to f(w) as possible by using a Gaussian surrogate function to approximate f(w), and an acquisition function that calculates the weight vector that is most likely to be near the true Pareto surface. We used q-Noisy Expected Hypervolume Improvement (qNEHVI) as our acquisition function. MOBO was tested on a treatment plan with around 8000 voxels for 30, 50, 100, 150 and 200 iterations. Execution time and fraction of clinically acceptable plans (success rate) were recorded for each iteration number.
Results: MOBO was able to approximate the Pareto surface within 30 iterations, taking 67 seconds and resulted in 13% success rate. As the number of iterations increased to 100 and 200, the time complexity grew super-linearly (552 s and 2793 s), while the success rate improved logarithmically (35% and 45%).
Conclusion: MOBO is a promising penalty weight automation algorithm for HDR brachytherapy. Next steps are to test it on multiple patients and other cancer types.
Funding Support, Disclosures, and Conflict of Interest: This project was funded by CIHR grant number 103548 and Canada Research Chairs Program (grant #252135).
Keywords
Bayesian Statistics, Mixed Integer Programming
Taxonomy
TH- Brachytherapy: Treatment planning using machine learning/automation
Contact Email