| 冯学振,王明清,左国平,等.基于射束路径有符号距离图的肺癌调强放疗多处方剂量预测[J].中华放射医学与防护杂志,2026,46(9):836-844.Feng Xuezhen,Wang Mingqing,Zuo Guoping,et al.Multi-prescription dose prediction for intensity-modulated radiotherapy in lung cancer based on beam path signed distance maps[J].Chin J Radiol Med Prot,2026,46(9):836-844 |
| 基于射束路径有符号距离图的肺癌调强放疗多处方剂量预测 |
| Multi-prescription dose prediction for intensity-modulated radiotherapy in lung cancer based on beam path signed distance maps |
| 投稿时间：2025-07-07 |
| DOI：10.3760/cma.j.cn112271-20250707-00226 |
| 中文关键词: 肺肿瘤 调强放射治疗 剂量预测 有符号距离图 级联卷积神经网络 |
| 英文关键词:Lung neoplasm Intensity-modulated radiotherapy Dose prediction Signed distance map Cascaded convolutional neural network |
| 基金项目:国家重点研发计划项目（2024YFA1014104）；北京市科委项目(Z221100003522028) |
|
| 摘要点击次数: 237 |
| 全文下载次数: 2 |
| 中文摘要: |
| 目的 通过引入射束路径有符号距离图，开发一种用于肺癌多处方调强放疗（IMRT）剂量分布预测的级联卷积神经网络CascU-Net-BDM。方法 回顾性分析2016年1月至2024年5月在北京大学第三医院接受治疗的190例肺癌患者IMRT混合处方的放疗数据。首次基于射束路径生成有符号距离图，并将其与CT影像、射束方向及解剖结构图像一起作为多通道输入，训练级联模型CascU-Net-BDM。分别构建并比较了三通道（CT、解剖结构、射束方向）的CascU-Net、四通道（CT、解剖结构、射束方向、计划靶体积距离图）的CascU-Net-PDM与本研究模型的性能，采用体素平均绝对误差（MAE）及关键剂量学参数进行评估。结果 CascU-Net-BDM模型的体素MAE为（1.19±0.41）Gy，显著低于CascU-Net的（2.51±0.51）Gy和CascU-Net-PDM的（1.42±0.42）Gy，差异有统计学意义（t=1.83、2.35，P＜0.05）。在剂量学指标方面，除PTV50的D 95%、D mean、适形指数、均匀性指数及PTV60的D 95%、D 98%、D mean外，其余指标在3个模型间差异均无统计学意义（P> 0.05）。除双肺V 20 Gy外，CascU-Net-BDM在所有OARs剂量指标中均表现出最低MAE。各结构体素MAE方面，CascU-Net-BDM较其他两模型均有显著降低，差异有统计学意义（t=1.98、2.11，P＜0.05）。结论 引入射束路径有符号距离图的CascU-Net-BDM在肺癌多处方IMRT剂量分布预测中优于CascU-Net及CascU-Net-PDM，尤其在射束边界外区域表现突出，能够实现准确、稳定的单模型多处方放疗剂量预测。 |
| 英文摘要: |
| Objective To develop a cascaded convolutional neural network, CascU-Net-BDM, for predicting dose distributions in multi-prescription intensity-modulated radiotherapy (IMRT) for lung cancer by incorporating beam path signed distance maps.Methods A retrospective analysis was performed on radiotherapy data of 190 patients with lung cancer who received mixed-prescription IMRT at Peking University Third Hospital between January 2016 and May 2024. For the first time, beam path signed distance maps were generated and, together with computed tomography (CT) images, beam directions, and anatomical structure images, used as multi-channel inputs to train the cascaded model CascU-Net-BDM. The performance of a three-channel CascU-Net (CT, anatomical structures, beam directions), a four-channel CascU-Net-PDM (CT, anatomical structures, beam directions, planning target volume distance map), and the proposed model were constructed and compared. These models were evaluated using voxel-wise mean absolute error (MAE) and key dosimetric parameters.Results The voxel-wise MAE of the CascU-Net-BDM model was (1.19±0.41) Gy, significantly lower than (2.51±0.51) Gy of CascU-Net and (1.42±0.42) Gy of CascU-Net-PDM (t=1.83, 2.35, P<0.05). In terms of dosimetric parameters, except for the minimum dose received by 95% of the target volume, mean dose within the target volume, conformity index, and homogeneity index of planning target volume with a prescribed dose of 50 Gy, as well as the minimum dose received by 95% of the target volume, minimum dose received by 98% of the target volume, and mean dose within the target volume of planning target volume with a prescribed dose of 60 Gy, no other parameters differed significantly among the three models (P>0.05). Aside from the bilateral lung volume receiving ≥20 Gy, CascU-Net-BDM demonstrated the lowest MAE across all organ-at-risk dose parameters. Regarding the voxel-wise MAE across the different structures, CascU-Net-BDM achieved a significantly lower value compared with the other two models (t=1.98, 2.11, P<0.05).Conclusions CascU-Net-BDM, which introduces the beam path signed distance maps, outperforms CascU-Net and CascU-Net-PDM in the prediction of multi-prescription IMRT dose distribution for lung cancer, especially in the regions outside the beam boundaries. It can achieve accurate and stable prediction of multi-prescription radiotherapy dose distribution in a single model. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|