| 李红伟,韩鸣,史以龙,姚晖,孟歌.基于深度学习的乳腺癌保乳术后调强放疗剂量分布预测[J].中华放射医学与防护杂志,2023,43(10):779-783 |
| 基于深度学习的乳腺癌保乳术后调强放疗剂量分布预测 |
| Dose distribution prediction of breast-conserving postoperative intensity-modulated radiotherapy for breast cancer based on deep learning |
| 投稿时间：2023-03-21 |
| DOI：10.3760/cma.j.cn112271-20230321-00088 |
| 中文关键词: 深度学习 卷积神经网络 剂量预测 调强放疗 |
| 英文关键词:Deep learning Convolutional neural networks Dose prediction Intensity-modulated radiotherapy |
| 基金项目: |
|
| 摘要点击次数: 950 |
| 全文下载次数: 336 |
| 中文摘要: |
| 目的 研究基于深度学习的方法预测乳腺癌保乳术后调强放疗(IMRT)剂量分布,并评估其预测精度。方法 回顾性分析2018年1月至2023年3月在上海国际医学中心接受IMRT的110例左侧乳腺癌保乳术后患者的调强放疗数据,随机固定选择80例作为训练集,随机固定10例作为验证集,剩余20例作为测试集。首先将患者的计算机体层成像(CT)图像、感兴趣区、体素与靶区距离和对应的剂量分布四通道特征作为输入数据,然后使用U-net网络进行训练得到预测模型,利用该模型对测试集进行剂量预测,验证体素与靶区距离特征在剂量预测中的影响,并将剂量预测结果与实际手动计划剂量进行比较。结果 加入体素与靶区距离特征的模型使预测精度更高,测试集中20例患者的剂量评分和剂量体积直方图(DVH)评分分别为2.10±0.18和2.28±0.08,与手动计划剂量分布更加接近(t=2.52、2.40,P<0.05)。靶区和危及器官(OAR)的剂量预测结果与手动计划剂量的偏差在4%以内,健侧乳腺平均剂量增加了13 cGy,均在临床可接受范围内。除PTV60的D2、D98(Di为i%的PTV体积接受的剂量)、Dmean(平均剂量)和患侧肺的Vsub>5(Vi为接受i Gy剂量的OAR体积百分比)、Dmean差异有统计学意义外(t=3.74、2.91、2.99、3.47、2.29,P<0.05),其他差异无统计学意义(P>0.05)。结论 基于深度学习的方法可以精准预测乳腺癌保乳术后调强放疗剂量分布,并通过实验证明加入体素与靶区距离特征可以有效提升预测精度,有助于物理师提高计划设计的优质性和一致性。 |
| 英文摘要: |
| Objective To develop the method based on deep learning to predict the dose distribution of breast-conserving postoperative intensity-modulated radiotherapy(IMRT) for breast cancer,and to evaluate accuracy of the prediction model. Methods The data of 110 left-sided breast-conserving postoperative IMRT for breast cancer patients were reviewed, among them, 80 cases were randomly selected for training set, 10 cases for validation set and the remaining 20 cases were used as test set.Firstly, the four-channel characteristics of the patients' computed tomography (CT) images, regions of interest, distances between voxel and planning target volume(PTV), and corresponding dose distributions were taken as input data.The established U-Net was used for training and obtaining prediction model which was utilized to perform dose prediction on the test set,in order to verify the influence of the features of distance between voxel and PTV in dose prediction,and to compare the dose prediction result with the actual manual planned dose. Results By incorporating the features of distance between voxel and PTV, the model achieved higher accuracy in predicting the dose distribution.The dose scores and dose volume histogram(DVH) scores of the testing set, consisting of 20 patients, were 2.10±0.18 and 2.28±0.08, respectively, and the predicted dose distribution was closer to the manually planned distribution(t=2.52, 2.40,P<0.05).The deviation between the predicted doses of the PTV and the organ at risk (OAR) and the manually planned doses were within 4%, the average dose to the contralateral breast was increased by 13 cGy, all of them within the clinically acceptable range. Except for the statistically significant differences in D2,D98(Di represents the dose received by i% of the PTV volume), Dmean(mean dose) of PTV60 and V5(Vi was the volume percentage of OAR receiving i Gy dose.), Dmean of the ipsilateral lung (t=3.74, 2.91, 2.99, 3.47, 2.29, P < 0.05), there were no statistically significant differences in other parameters. Conclusions The deep learning-based method can accurately predict the dose distribution of breast-conserving postoperative IMRT for breast cancer,and it has been proven through experiments that by incorporating the features of distance between voxel and PTV can effectively improve the prediction accuracy,which helps physicists to improve the quality and consistency of treatment planning. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|