Based on deep learning, a 3D dose distribution prediction method for patient specific radiotherapy plan is established in this paper, which can automatically generate the patient specific radiotherapy dose distribution. The prediction results can be used in the evaluation model based on tumor percentage dose coverage, volume dose of normal tissues and organs at risk as parameters to evaluate the advantages and disadvantages of dose distribution.
In the case of given planning image, a method for predicting the optimal dose distribution and segmented anatomy are developed, and the deep learning technology is applied to the previously optimized database and approved IMRT plan. A total of 80 cases of early nasopharyngeal carcinoma (NPC) were included in this study. Randomly selected 70 cases as the training set, the rest as the test set. The input is a structured image, and each target and organ at risk (OAR) is assigned a unique tag. The output is dose map, including coarse dose map and converted fine dose map (FDM) convolution. Two types of structured input images are used in the model construction. One input includes an unprocessed image (with a correlation structure). The second type of input involves modifying the gray label of the image with information from the geometry of the radiation beam. Resnet101 was chosen as the deep learning network of both. The accuracy of predicting dose distribution was evaluated according to the corresponding dose used in clinic.
Input structured images, and each target and organ at risk (OAR) is assigned a unique label. The output dose map includes coarse dose map and converted fine dose map (FDM).
Good data quality
| # | number | name | type |
| 1 | 2017YFC0107500 | Research and implementation of a new conformal intensity modulation technique for multi-particle biology guided by multi-modes | National key R & D plan |
This work is licensed under
CC BY 4.0 (Creative Commons Attribution 4.0 International License).
CSTR:11738.11.ncdc.IMP.db2510.2022
10.12072/ncdc.imp.db2510.2023
