Radiotherapy planning dose prediction using convolutional neural network: a comparison of two kinds of decoders(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2020年第2期
- Page:
- 153-158
- Research Field:
- 医学放射物理
- Publishing date:
Info
- Title:
- Radiotherapy planning dose prediction using convolutional neural network: a comparison of two kinds of decoders
- Author(s):
- CHEN Xinyuan; YI Junlin; DAI Jianrong
- National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing 100021, China
- Keywords:
- Keywords: convolutional neural network; decoder; radiotherapy; dose prediction
- PACS:
- R318；R815
- DOI:
- DOI:10.3969/j.issn.1005-202X.2020.02.004
- Abstract:
- Abstract: Objective To establish a three-dimensional radiotherapy planning dose distribution prediction model using convolutional neural network, and to compare the effects of two kinds of decoders on network performance. Methods The radiotherapy plans of 80 cases of early nasopharyngeal carcinoma were included in the study, and then were randomly divided into a training set (70 cases) and a test set (10 cases). Based on VGG16 convolutional neural network, interpolation VGG16 network (IVGG16) and deconvolution VGG16 network (DVGG16) were established by two kinds of decodes, namely interpolation and deconvolution. The established networks were used for end-to-end radiotherapy planning dose prediction. The main indicators for the evaluation of model accuracy included the mean absolute errors (MAE) of the outline, tumor target areas and organs-at-risk (OAR). Moreover, the training time and prediction time of two kinds of networks were also recorded. Results The three-dimensional dose distribution could be predicted accurately using both two kinds of decoders. The MAE of the outline obtained by IVGG16 and DVGG16 were (5.48±0.46) % and (5.42±0.34)%, respectively, without statistical significance. The doses of the target areas were accurately predicted by two kinds of decodes, with a MAE lower than 2.63%, and there was no statistical difference between two kinds of decodes. The dose distributions of OAR were accurately predicted by two kinds of networks. However, compared with those obtained by IVGG16, the MAE of spinal cord PRV and thyroid obtained by DVGG16 was decreased by 11.8% and 15.6%, respectively (P=0.029, 0.034), and there was no statistical differences in the MAE of the remaining OAR. The training time of IVGG16 and DVGG16 models were 14.8 h and 24.6 h, respectively, and the average prediction time for each case were (6.6±1.0) s and (28.7±3.9) s, respectively. Conclusion The radiotherapy dose prediction using interpolation or deconvolutional decoders has a satisfactory effect. The deconvolutional decoder has a slightly higher prediction accuracy for some OAR dose distribution, but the efficiency of model training and prediction need to be further improved.
Last Update: 2020-03-03