PV QA 3 - Poster Viewing Q&A 3
TU_14_3249 - Clinical Application of a Novel Voxel- and Machine Learning-Based Automated Planning Method for Prostate Volumetric Arc Radiation Therapy
Tuesday, October 23
1:00 PM - 2:30 PM
Location: Innovation Hub, Exhibit Hall 3
Clinical Application of a Novel Voxel- and Machine Learning-Based Automated Planning Method for Prostate Volumetric Arc Radiation Therapy
A. Berlin1,2, L. Conroy2,3, M. C. Tjong1,2, T. Craig2,3, P. Chung1,2, C. McIntosh3,4, and T. G. Purdie2,5; 1Radiation Medicine Program, Princess Margaret Cancer Centre, University Health Network, Toronto, ON, Canada, 2Department of Radiation Oncology, University of Toronto, Toronto, ON, Canada, 3Medical Physics, Princess Margaret Cancer Centre, University Health Network, Toronto, ON, Canada, 4Techna Institute, University Health Network, Toronto, ON, Canada, 5Radiation Medicine Program, University Health Network and Princess Margaret Cancer Centre, Toronto, ON, Canada
Purpose/Objective(s): We have developed a novel voxel- and machine learning-based automated planning method that requires the planning image dataset, target contours, relevant organ at risk (OAR) volumes as inputs for dose distribution prediction, and a voxel dose dose-mimicking step that generates a complete treatment plan in 20-25 minutes. As the method has automated determination of appropriate clinical trade-offs, user input during the optimization process is not required. Herein, we validate our method for clinical application in prostate volumetric arc radiotherapy (RT).
Materials/Methods: A dose prediction algorithm was trained and tested using 116 consecutive clinically used plans for prostate cancer radical RT. An independent dataset of 20 patients with localized prostate cancer was randomly selected and the dose prediction algorithm was used to generate RT plans. These were evaluated alongside the original clinical plans by three expert reviewers (2 GU radiation oncologists and 1 medical physicist) blinded to the plan origin. Three plans were excluded prior to any review as they contained hip prosthesis. Evaluation criteria included acceptability and preference with regards to target coverage, dose conformity, dose gradient at rectum/prostate interface, lateral symmetry, OAR sparing, and overall approval.
Results: Of the 17 patients reviewed, automated plans were deemed superior in target coverage and OAR sparing, similar in gradient at rectum/prostate interface, and inferior in high-dose conformity and lateral dose symmetry. Both automated and clinical plans were deemed clinically acceptable in 49 of 51 instances across the three reviewers. One automated plan was deemed clinically unacceptable by two reviewers, and two different clinical plans were deemed clinically unacceptable by one reviewer. When compared head-to-head by the three reviewers, automated plans were deemed superior in 38 (74.5%) and equivalent to clinical plans in 3 (5.9%) reviews. Overall, in 15 of 17 (88.2%) patients, automated plans were deemed superior or equivalent to the delivered clinical plans.
Conclusion: Our machine learning-based automated planning framework has the potential for clinical application in prostate RT, providing and opportunity to improve efficiency and consistency in the planning process. Furthermore, without the need for user interaction during plan optimization process, can enable increased throughput and liberate human resources for being allocated elsewhere within the RT process. Integration into our clinical practice including prospective performance evaluation of the proposed method is ongoing.
Author Disclosure: A. Berlin: Research Grant; Astellas, AbbVie. Board of Directors; Vaccinex. L. Conroy: None. M.C. Tjong: None. T. Craig: None. P. Chung: Research Grant; Sanofi. C. McIntosh: Patent/License Fees/Copyright; RaySearch Laboratories.