Study on radiation dose distribution based on generative adversarial network / 中华放射肿瘤学杂志

ABSTRACT

Objective:To investigate whether the combination of the advantages of deep learining in image processing and radiotherapy will make the radiotherapy process more intelligent.Methods:The generative adversarial network (GAN) is a generation model using neural network. High-quality dose distribution images can be generated by inputting relevant features. Firstly, random unconditional GAN was utilized to verify the ideal data, then conditional GAN (cGAN) was employed to train DICOMRT data of tumor patients, and the target contour information was used to directly generate dose distribution images.Results:For the verification of ideal data, the generation of ideal distribution yielded good effect. By extracting target contour and real dose slice data and using cGAN training, the dose distribution maps of planning target volume (PTV) and organs at risk (OAR) of tumor patients could be obtained. The absolute error evaluation of the maximum and average values between the predicted value and the real dose was shown as[3.57%, 3.37%](PTV), [2.63%, 2.87%](brain), [1.50%, 2.70%](CTV), [3.87%, 1.79%](GTV), [3.60%, 3.23%](OAR1) and[4.40%, 3.13%](OAR2), respectively.Conclusions:GAN model can be used to generate ideal dose distribution data, and cGAN model with prior knowledge can be employed to establish the relationship between target information and dose distribution. Directly generating the corresponding dose distribution image by inputting the target contour information is an effective attempt for dose prediction.