Room: Exhibit Hall | Forum 7
Purpose: To alleviate the tedious manual beam selection during treatment planning, many optimization algorithms have been designed to optimize beam orientations, but suffer from slow calculation of the dose influence matrices of all candidate beams. We propose a fast beam orientation selection method, based on deep neural networks, capable of developing a plan that is non-inferior to a state-of-the-art optimization method called column generation(CG).
Methods: A supervised deep learning neural network (DNN) is trained to mimic (CG), which iteratively chooses beam orientations one-by-one, by calculating beam fitness values based on Karush-Kush-Tucker optimality conditions at each iteration. Specifically, the DNN learns to predict these beam fitness values. 70 prostate cancer patients, split into 50 training, 7 validation, and 13 test, with 6 contours, PTV, body, bladder, rectum, left and right femoral heads were used to develop and test the model. CG was implemented with a GPU-based Chambolle-Pock algorithm, a primal-dual proximal-class algorithm, to create a total of 6000 plans. The DNN trained over 400 epochs each with 2500 steps, using the Adam optimizer at a learning rate of 10e-5.
Results: The average and standard deviation of training, validation and test loss functions among 6-fold were 0.62Â±0.09%, 1.04Â±0.06%, and 1.44Â±0.11% respectively. By using CG and DNN, two sets of plans for each scenario in test set were generated. The DNN took 1 second to select a set of orientations and 320 seconds to solve the FMO, while CG needed at least 648360 seconds, including the calculation of the dose influence matrix. Bladder had the minimum average differences in dose received by OARs(0.956Â±1.184) then rectum(2.44Â±2.11%) and Left-Femoral Head(6.03Â±5.86%) while Right-Femoral head had the maximum(5.885Â±5.515).
Conclusion: We developed a fast beam orientation selection method based on deep DNN that finds beam selections in orders of seconds, and is therefore is suitable for clinical routines.
Not Applicable / None Entered.