(Tuesday, 7/14/2020) 3:30 PM - 5:30 PM [Eastern Time (GMT-4)]
Room: Track 3
Purpose: To design a novel optimization-free interactive planning framework for volumetric modulated arc therapy (VMAT), in which prior physics knowledge acts in concert with deep learning to bypass iterative optimizations for real-time plan tuning.
Methods: The proposed framework consists of three modules: 1) a dose initialization module; 2) an interactive tuning module; and 3) a fluence map (FM) prediction module. Dose initialization module generates population-based initial dose distribution automatically. Interactive tuning module has a front-end interface for interactive dose volume histogram (DVH) curves and iso-dose lines modification, by which users explore possible trade-offs in real time till they obtain the desired dose. In backend it’s achieved by iterations between a dose massager which enhances desirability and an auto-encoder which guarantees smoothness for high deliverability. FM prediction module takes projections of desired dose as input and outputs VMAT FMs. It firstly uses a U-Net, acting like inverse of dose influence matrix, to map projections of dose to FMs in phantom geometry. Then it uses a plan scaling technique to scale FMs from phantom geometry to patient geometry.
Results: Framework is trained on 341 and tested on 102 clinical Head and Neck VMAT plans. Initialization module runs off-line before planning. Interactive tuning module takes less than 10 iterations (<1sec/iter) to obtain desired dose. FM prediction module takes < 1 second to predict FMs. We calculated dose from predicted FMs and compared it with clinical planned dose on testing plans. Gamma passing rate is 98.06±2.64% under 3%/3mm criteria. Mean dose differences in all ROIs are less than 2%.
Conclusion: We developed a novel optimization-free interactive planning framework. Without involving iterative optimization, the proposed framework can obtain desired dose distribution and corresponding FMs in real time, which could be a game changer for original and adaptive VMAT planning.
Funding Support, Disclosures, and Conflict of Interest: This work is in part supported by NIH grants (R01 CA235723, R01 CA218402).
Keywords
Taxonomy
Not Applicable / None Entered.
Contact Email
