(Saturday, 4/4/2020) 10:30 AM - 12:30 PM [Mountain Time (GMT-6)]
Purpose: To develop a machine learning model for V12Gy/V60% prediction in LINAC-based Treatment Single-Iso-Multiple-Targets (SIMT) radiosurgery planning. This model can assist in planning decision-making and improve dosimetric quality consistency.
Methods: The developed model interprets a SIMT case using the following statistics: 1) prescription level; 2) target number and volumes; 3) modified center-of-mass distances; 4) equivalent spherical surface areas. These statistics were utilized by a gradient boosted trees (GBT) regression for V12Gy/V60% prediction. Hyper-parameters based on the extracted statistics in the GBT were automatically fine-tuned by a random search optimizer.Two models were built in this work. Model A was based on 71 SIMT plans (2~25 lesions, average=6) planned by multiple planners. Model B was based on 27 plans (2~23 lesions, average=5) planned by a single planner. In each model training, 20% cases were used for validation. For independent model tests, another 12 cases (2~11 lesions, average=5), including 5 from Model B’s planner, were studied. In test cases, the predicted V12Gy/V60% results with uncertainty ranges were compared with ground truth values. An API-based one-click GUI was designed to implement the developed models in the clinical treatment planning system (TPS).
Results: V12Gy/V60% in 10 of 12 test plans were accurately predicted by Model A. The mean absolute difference (MAD) of prediction was 1.62cc, and the mean prediction uncertainty (MPU) was 3.65cc. For Model B results, all 5 test plans were accurately predicted, and MAD/MPU results were 0.53cc/2.47cc. Each prediction took less than 1 seconds when being implemented in clinical TPS. Using model predictions as guidelines, 2 cases in which V12Gy/V60% results were higher than the predicted ranges were improved during replanning.
Conclusion: A machine learning model for SIMT SRS V12Gy/V60% prediction was developed. Feasibility for its clinical application in planning decision-making and dosimetric quality consistency improvement was successfully demonstrated.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by NIH R01CA201212.
Contact Email
