[1]孔繁图,麦燕华,亓孟科,等.基于神经网络学习方法的放疗计划三维剂量分布预测[J].南方医科大学学报,2018,(06):683.
点击复制

基于神经网络学习方法的放疗计划三维剂量分布预测()
分享到:

《南方医科大学学报》[ISSN:/CN:]

卷:
期数:
2018年06期
页码:
683
栏目:
出版日期:
2018-06-30

文章信息/Info

Title:
Prediction of three-dimensional dose distribution in intensity-modulated radiation therapy based on neural network learning
作者:
孔繁图麦燕华亓孟科吴艾茜郭芙彤贾启源李永宝宋婷周凌宏
关键词:
调强放射治疗三维剂量分布预测神经网络学习
Keywords:
intensity-modulated radiation therapy three-dimensional dose distribution prediction neural network model
摘要:
目的研究利用神经网络建立患者几何解剖结构特性与对应放疗计划三维剂量分布的关联模型,用以新患者的计划剂量 学特性预测。方法收集并应用25例固定13野临床前列腺癌IMRT/SBRT计划,处方剂量为50 Gy,以体素为研究对象,提取器 官体积、射线角度、器官位置关系等信息,采用神经网络构建剂量预测模型。随机选取20例用于训练,5例用于验证,以验证预 测方法的可行性与有效性。结果实验表明,模型训练误差较小,验证样本剂量差异小,预测结果准确。模型训练误差较小,点 对点平均剂量差异不高于0.0919±3.6726 Gy,DVH差异不超过1.7%。5例验证样本,剂量差异为0.1634±10.5246 Gy,百分剂量 差异在2.5%以内,DVH差异在3%以内。三维剂量分布显示剂量差异较小,预测剂量分布合理。模型对膀胱和直肠预测性能表 现较好,股骨头次之。结论本文使用神经网络学习方法成功建立了同肿瘤类型患者几何解剖结构特性和对应计划三维剂量分 布之间的关联,不仅可以为计划提供质量标准,还为实现自动计划控制和计划设计奠定基础。
Abstract:
Objective To establish the association between the geometric anatomical characteristics of the patients and the corresponding three-dimensional (3D) dose distribution of radiotherapy plan via feed-forward back-propagation neural network for clinical prediction of the plan dosimetric features. Methods A total of 25 fixed 13-field clinical prostate cancer intensity-modulated radiation therapy (IMRT)/stereotactic body radiation therapy (SBRT) plans were collected with a prescribed dose of 50 Gy. With the distance from each voxel to the planned target volume (PTV) boundary, the distance from each voxel to each organ-at-risk (OAR), and the volume of PTV as the geometric anatomical characteristics of the patients, the voxel deposition dose was used as the plan dosimetric feature. A neural network was used to construct the correlation model between the selected input features and output dose distribution, and the model was trained with 20 randomly selected cases and verified in 5 cases. Results The constructed model showed a small model training error, small dose differences among the verification samples, and produced accurate prediction results. In the model training, the point-to-point mean dose difference (hereinafter dose difference) of the 3D dose distribution was no greater than 0.0919±3.6726 Gy, and the average of the relative volume values corresponding to the fixed dose sequence in the DVH (hereinafter DVH difference) did not exceed 1.7%. The dose differences among the 5 samples for validation was 0.1634 ± 10.5246 Gy with percent dose differences within 2.5% and DVH differences within 3%. The 3D dose distribution showed that the dose difference was small with reasonable predicted dose distribution. This model showed better performances for dose distribution prediction for bladder and rectum than for the femoral heads. Conclusion We established the relationships between the geometric anatomical characteristics of the patients and the corresponding planning 3D dose distribution via feed-forward back-propagation neural network in patients receiving IMRT/SBRT for the same tumor site. The proposed model provides individualized quality standards for automatic plan quality control.
更新日期/Last Update: 1900-01-01