Convolutional neural network based dose prediction method for intensity-modulated radiotherapy of cervical cancer
10.3969/j.issn.1005-202X.2025.04.001
- VernacularTitle:基于卷积神经网络的宫颈癌调强放疗剂量预测方法
- Author:
Xiaojuan WU
1
;
Yibao ZHANG
;
Hongru REN
;
Lingjun MENG
Author Information
1. 广东工业大学自动化学院,广东 广州 510000;山东大学齐鲁医院德州医院放射治疗科,山东 德州 253000
- Publication Type:Journal Article
- Keywords:
cervical cancer;
dose prediction;
convolutional neural network;
intensity-modulated radiotherapy
- From:
Chinese Journal of Medical Physics
2025;42(4):421-428
- CountryChina
- Language:Chinese
-
Abstract:
Objective To develop a convolutional neural network based model for predicting the dose distribution of intensity-modulated radiotherapy(IMRT)in cervical cancer,and to evaluate its potential applications in automated treatment planning.Methods The pelvic IMRT plans for 100 female patients were collected,with 80 cases in the training set,10 in the validation set,and 10 in the test set.A dose prediction model was built based on the three-dimensional(3D)residual network for forecasting 3D dose distribution.Masks for organs-at-risk and planning target areas were extracted from CT images and RT Structure files.Density values were assigned to different structures according to a density map,and the resulting CT maps were used as input images for model training.The optimal model was used to predict the 3D dose distribution,and the predicted results were compared with the dose distribution from manual treatment planning in terms of dosimetric parameters.Results The experimental results on the 10-case test set demonstrated that dosimetric parameter differences were insignificant and within clinically acceptable ranges.The mean absolute error,average Dice similarity coefficient,and 95%Hausdorff distance for 10 cases in test set were(0.58±0.16)Gy,0.90±0.03,and(10.61±7.17)mm,respectively.Compared with manual planning,prediction model showed slightly decreased rectal V45,small bowel D2cc,and the V20 of bilateral femoral heads was reduced.The predicted D95 of planning target area was lower than manual planning,but the differences in D90,homogeneity index,and conformity index were trivial.There were minor differences in 3D dose distributions between the two,and the dose distribution generated by prediction model met clinical requirements.Conclusion The convolutional neural network based dose prediction model can accurately forecast the dose distribution for cervical cancer IMRT,exhibiting the potential to be used in automated treatment planning and quality evaluation.
