Room: Track 3
Purpose: To investigate if machine learning auto planning can manage the large variation of tumor location in lung SBRT and identify challenging locations for manual planning.
Methods: A machine learning optimization (MLO) model was trained in RayStation 9B by 99 VMAT plans with multi-criteria optimization (MCO), each treating one lesion to 50 Gy in 5 fractions with dose escalation of 55 to 65 Gy to ITV-4mm when possible. MLO uses random forest to predict voxel dose and a collection of dose-volume rules (called strategy) to refine 3D dose distribution. The model had five strategies to manage different dose escalations. It was validated on 14 representative patients and tested on 22 consecutive new patients. The auto plans were scored as clinically acceptable, acceptable after minimal post-processing or no efficiency gained. The dosimetric quality for accepted auto plans was compared to the manual plan. Conformality index (CI) using RTOG 0833 criteria and robustness evaluation (5-mm setup error) were assessed.
Results: Eleven auto plans were accepted directly and five after simple processing. The six plans requiring more processing included two in right apex (due to competing goals on chest wall and mediastinum) and one in anterior left lung (due to coronary artery) and three in posterior left lung abutting mediastinum (due to various vessels). The 16 accepted plans offered almost identical target coverage and organ sparing to the manual plans. The auto plans were more conformal (all passes vs. 14/16 passes), a bit more robust (+0.9% on ITV coverage in worst case scenario) but used 39% more MU.
Conclusion: The MLO model produced acceptable auto plan for 16 of the 22 testing patients (73% of the time) and identified challenging cases that may need manual planning. It can provide multiple auto plans, facilitating the clinicians to choose the most desirable dose escalation level.
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation