Room: AAPM ePoster Library
Monte Carlo (MC) algorithms generate accurate modeling of dose calculation by simulating the transport and interactions of many particles through all aspects of radiological physics within 3D heterogeneous media such as the human body. However, given their random nature, an MC calculation for doses in radiotherapy usually use a huge number of simulated particles to achieve acceptable statistical uncertainty (noise), results in significantly greater computation times and impeding the MC methods from wider clinic applications, especially the expansion of treatment planning system based on MC method. Therefore, there is always a trade-off between the computation time and the noise level in MC calculation.
In this work, we propose a deep convolutional neural network (CNN) approach to achieve the fast and fully automated denoising and accelerating of the MC dose calculations in radiotherapy, which is trained to predict the noise-free dose maps of 1*109 photons from the 1*106 noise-much dose maps. Proposed model is trained on MC dose distributions of 50 patients’ different tumor sites (brain, head and neck, liver, lungs, and prostate) acquired from Elekta treatment planning system Monaco. Dilated U-Net is trained by considering a number of adjacent slices, using mean squared error (MSE) as loss function and 3D Gamma Index Passing Rate (GIPR) as the evaluation of the performance of predicted dose maps.
The preliminary results demonstrate that the proposed mode can improve the GIPR of dose maps of 1e6 photons, and provide comparable qualitative and quantitative results as the MC dose distribution simulated with 1e9 photons, yielding over 100 times speed-up in the MC simulations of IMRT (15 s vs 95 min).
Our method shows the feasibility of using a deep learning approach in MC clinic practice of advanced treatment technologies.