Click here to


Are you sure ?

Yes, do it No, cancel

A Method Based On Deep Convolution Network for Predicting Three-Dimensional Dose Distribution in Small Datasets of IMRT Nasopharyngeal Carcinoma

F Guo*, L Dingjie, He Nan Cancer Hospital, Zhengzhou, 41AF,


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

To establish a 3D dose prediction model, which is capable of automatically learning the relationship between patients’ geometric anatomical structure and corresponding dose distribution of small number of IMRT Nasopharyngeal carcinoma (NPC) patients.


Collect clinical radiotherapy plans for NPC and extract relevant information needed for dose prediction from the radiotherapy plan. In the establishment of the model, the required information is the section slice of contour structure, which mainly includes the geometric structure distribution of ROIs, and the corresponding dose distribution matrix of this layer. The model architecture is based on the U-net and Densenet. The network has five symmetrical layers, each layer is made up of denseblocks to achieve feature reuse. In order to improve the accuracy of the model on small data sets, influence of different section slice at different angles was considered to training different model.

The total of 40 cases of NPC were collected from the clinical plan in this study, of which 32 cases were training data and 7 cases were validation data, And model performance is validated by cross validation. The experimental results show that the model has satisfactory predictability and different cross section angles will greatly affect the model fitting effect. It is easy to see that the accuracy of coronal section is higher than horizontal section, where voxel-based error of coronal section was maintained at 2.5% of the prescription dose.
First, a deep convolutional network was proposed to predict the 3D dose distribution of radiotherapy in NPC. By comparison DVH and three-dimensional dose distribution, the predicted dose is close to the planned dose. At the same time, by analyzing the structure distribution of the patient, finding the section that is more suitable to expressing the organs distribution characteristics can better improve the training effect in the small dataset.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email