Purpose: We have proposed a Deep Learning-based Monte Carlo (MC) dose denoising framework, which innovatively uses MC statistical uncertainty (SU) distribution in combination with patient CT images for denoising to obtain accurate dose maps.
Methods: First, we generate a database of MC dose maps and corresponding SU distributions for low- to high-histories using MC calculation engine. Then train the model to map the low-histories dose to high-histories dose maps. At the core of this framework is a multi-channel denoising network, which has multi-channel inputs with CT images and low-histories MC dose and its corresponding SU distributions, and outputs a denoised dose map. Finally, we evaluated the feasibility of the model to obtain high precision dose map through evaluation metrics MAE, ME, GPR, DVH, and dose maps.
Results: The dose maps show the denoised dose is more consistent with the structural distribution shown in CT images, especially in the high-dose non-uniform areas. DVH curves provide more accurate target coverage and OARs distribution information. The predicted results effectively eliminate singularities in the input dose, and match better with the ground truth in both OARs and PTVs areas. The global prediction performance is also confirmed in the gamma maps. In the quantitative comparison results, the predicted dose demonstrated substantially improved match to the ground truth, with average GPR 99.38%±3.9%, the global ME and MAE were ₋0.25%±0.12% and 0.83%±0.33% respectively. The accuracy improvement between the input dose and the predicted dose can be up to tenfold.
Conclusion: By using the SU distribution with dose difference information and CT images with electron density and anatomical structure information, the model better learned the potential inaccuracies in the input, further improving the accuracy of dose denoising. Our framework more efficiently obtain accurate dose distribution, to help physicists make more informed decisions when planning and evaluating treatment plans.
Funding Support, Disclosures, and Conflict of Interest: (1)Natural Science Foundation of Guangdong Province, China (No. 2018A0303100020), (2)National Key R&D Program of China (Grant NO. 2017YFC0113203) (3)National Natural Science Foundation of China (Grant NO. 11805292, 81601577and 81571771).
Not Applicable / None Entered.