Room: Track 3
Purpose: introduce a novel and efficient dose calculation algorithm --- Deep DoseNet (DDN), based on the deep learning strategy which use raytracing as the first order approximation and learns dose deposition from conventional dose calculation algorithm with high accuracy.
Methods: uses raytracing and CT as the input, and dose calculated with the conventional algorithm as the training data. Ten patient CT datasets of different disease sites, including brain, thorax, abdomen and pelvis, are selected for this study. For each patient, 100 beam dose distributions are computed with random beam iso-center locations, XY jaw positions and gantry angles using AXB algorithm at 2.5 mm grid size for a Truebeam machine with an energy of 10fff. With the same beam parameters, we performed ray tracing utilizing the Siddon algorithm and calculated approximate dose depositions with the PDD and beam profiles from the golden beam data. DDN is adapted from the state-of-the-art network structures: RESNET and DENSENET. The optimization uses the ADAM algorithm with a learning rate of 10^-4 and takes ~ 500 Epochs of training to converge with a batch of 32 on a Nvidia DGX system. Each epoch takes ~20 minutes.
Results: compared the raytracing and DDN outputs with the AXB dose. For the evaluation set, the average mean-square-error decreased from 2.9x10^-3 to 1.1x10^-4 and the average Gamma index passing rate at 3mm3% improved from 53.3% to 92.7%. For a single beam, the average calculation time is less than 1 milliseconds for one slice.
Conclusion: trained with a large amount of dosimetry data, can potentially be employed as a general-purpose dose calculation engine to speed up clinical workflow. It also brought a new way of performing 2nd dosimetry check.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by NIH (1R01 CA176553 and R01CA227713), Varian Medical Systems, and a Faculty Research Award from Google Inc.