Purpose: We propose a second dosimetry check framework based on data collected from model-based calculations or measured beam doses, from which a convolution neural network is trained to preform dose calculation with high accuracy and speed.
Methods: We first estimate the dose from primary photon transport using the cross-axis profile of field size 40 cm and the PDD for 100 cm SSD and 10 cm beam size corrected by Mayneord factor using radiological distances calculated from raytracing at the 5 mm voxel size. The network, named as Deep DoseNet (DDN), takes the estimated 3D doses and corresponding CT scans as the input and outputs up sampled final doses at 2.5 mm. DDN uses Residual in Residual Dense Block as the backbone where local features learned at each layer are sent to all subsequent layers to promote feature reuse, which is widely used and the state of art in the image super resolution field. For the training, we generated 2000 beams at 6 MV in Eclipse TPS on 20 patient CT scans. The beam isocenter, gantry angle, collimator angle, jaw position and MLC position are all randomly generated. The doses are calculated with the Acuros algorithm at 2.5x2.5x2.5 mm. The mean square error of the DDN converges after training for ~20 epochs using the ADAM optimization algorithm with an initial learning rate of 10^-4.
Results: For 11 randomly selected vmat arc fields on patient CTs which the network hasn’t seen before, the average 3D gamma passing rate for 1%1mm is 74.8% (58.6%-88.3%), 2%2mm is 96.5% (93.9%-99.1%) and 3%3mm is 99.5% (98.5%-99.8%). The average calculation time is ~30 seconds with one Nvidia V100 GPU card.
Conclusion: DDN successfully generalizes from limited training dataset. It can serve as an independent, fast and accurate second dosimetry check tool for VMAT treatments.