(Wednesday, 7/17/2019) 1:45 PM - 3:45 PM
Room: 301
Purpose: Fluence map optimization is the most time-consuming and computation-intensive step in the inverse planning of IMRT and VMAT. In order to bypass the optimization problem, deep learning is used to predict fluence maps from the projections of PTV and OARs directly.
Methods: 65 prostate cancer IMRT plans with seven 10MV beams are collected from our institution to build a deep learning model for fluence map prediction. All patients have similar beam angle distribution which makes the learning possible and circumvents the beam angle selection problem. Model structure is 3D U-net whose input has three channels for PTV, rectum and bladder respectively and output has one channel for fluence map. Each input channel is a stack of seven Beam Eye View projections(BEV) of organ masks while output channel is a stack of seven fluence maps.
Results: Dataset is split into training set (41), validation set (11) and test set (13). Corresponding mean square errors are 0.0011±0.0004, 0.0024±0.0009 and 0.0035±0.0024. Further, the model is able to predict fully-deliverable dose simply by calculating the dose from predicted fluence map. Here CCCS algorithm is used for the calculation of predicted dose. Then it is compared to TPS dose by gamma analysis, which is a clinical-meaning way of evaluating the quality of fluence map prediction. Gamma passing rates with 3%/3mm criteria for training set, validation set and test set are 0.94±0.05, 0.87±0.06 and 0.83±0.08.
Conclusion: The feasibility of using deep learning to predict fluence map is validated by the scenario of seven beam IMRT prostate cancer plans. Direct fluence map prediction not only reduce the workload by skipping optimization, but also allows for showing physicians a deliverable dose, which effectively shortens planning time. In the future the same idea will be tested on other clinical scenario like H&N VMAT.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
