Purpose: The requirement for more accurate dose calculation algorithm and/or higher spatial resolution is growing rapidly for advanced radiation therapy modalities such as SBRT/SRS. We introduce a novel dose calculation method based on deep neural network, which is not only capable of up-sampling the dose, but also transforming the dose calculated in one algorithm to another with high accuracy and speed.
Methods: 22,000 slices of dose distribution are calculated using Eclipse V13.7.14(Varian TPS) on patient CT with a grid size of 1.25x1.25x1.25mm using both AAA(5mm) and Acuros(1.25mm) algorithms. The AAA dose slices and the corresponding down sampled CT slices are combined to form a tensor with a size of 2x64x64, working as the input to the deep learning-based dose calculation network (DoseNet), which output the calculated Acuros dose with a size of 256x256. The DoseNet has 37layers in total and the first 30 layers serve as the feature extraction component and the last 7 layers work mainly for upscaling. We trained the DoseNet to calculate the difference between the AAA and Acuros, cutting the training time by ~10 times. The DoseNet converges after ~100 epochs in the Nvidia DGX system with a learning rate of 10 and a gradient clipping of 10^-7, using stochastic gradient descent.
Results: We compared upsampled AAA dose and DoseNet output with Acuros. For the evaluation set, the average mean-square-error decreased from 3.08x10^-4 between AAA and Acuros to 1.63x10^-5 between DoseNet and Acuros. The average Gamma index passing rate at 3mm3% improved from 80.9% between AAA and Acuros to 97.1% between DoseNet and Acuros. The average calculation time is less than 1 millisecond for a single slice.
Conclusion: DoseNet, trained with a large amount of dosimetric data, can be employed as a general-purpose dose calculation engine across various dose calculation algorithms and for treatment plan optimization.