MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Development and Evaluation of a Machine Learning Optimization Model for Automatic Planning in Liver SBRT Using a Multicriteria Optimization Knowledge Base in RayStation

H Prichard*, J Wo, Y Wang, Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA

Presentations

(Wednesday, 7/17/2019) 9:30 AM - 10:00 AM

Room: Exhibit Hall | Forum 2

Purpose: To develop and evaluate a machine learning optimization (MLO) model for auto-planning in liver SBRT, using a knowledge base of IMRT and VMAT plans with multicriteria optimization (MCO).

Methods: An MLO model was built in RayStation 8B using 45 liver SBRT patients treated with MCO-IMRT (17) or MCO-VMAT (28). All patients received 50 Gy in five fractions, with no substantial compromise on CTV coverage. The model generates VMAT auto-plans using three arcs and a 6-MV FFF beam (referred to as pre-processed plan), which could be post-processed in standard optimization (post-processed plan). The model was tested on ten new patients (7 IMRT and 3 VMAT). The pre- and post-processed MLO plans were compared to clinical MCO plans. All MLO plans were reviewed by a radiation oncologist.

Results: It takes MLO ~4 to 10 minutes to generate a fully-deliverable pre-processed plan, whereas it takes MCO ~20 to 35 minutes to create Pareto plans for navigation. The pre- and post-processed MLO plans provided higher CTV coverage and lower liver dose than the MCO plans – CTV V100% up by 1.39% (p = 0.017) and 0.52% (p=0.071), and liver-GTV mean dose down by 0.21 Gy (p = 0.526) and 0.68 Gy (p = 0.003), respectively. Dose to other organs were comparable. The MLO model was only trained on organs, but not planning organ at risk volumes (PRVs). All pre-processed plans were clinically acceptable, with two on borderline due to dose-limiting PRVs. After post-processing, all ten plans were further improved and all PRV goals were met.

Conclusion: An MLO auto-planning model was created using an MCO knowledge base for liver SBRT and tested on ten new patients. The model produced clinically acceptable auto-plans with reduced calculation time. MLO may also improve plan quality by reducing planner variation and providing favorable starting point for post-processing.

Keywords

Treatment Planning

Taxonomy

TH- External beam- photons: treatment planning/virtual clinical studies

Contact Email