基于Transformer注意力机制的肺癌调强放射治疗三维剂量预测  

Transformer attention mechanism based three-dimensional dose prediction for lung cancer intensity-modulated radiotherapy

作　　者：陈旸汀 杨鑫 靳富 冯彬 罗文[1] Chen Yangting;Yang Xin;Jin Fu;Feng Bin;Luo Wen(School of Nuclear Science and Technology,University of South China,Hengyang 421001,China;Department of Radiotherapy,Chongqing University Cancer Hospital,Chongqing 400030,China)

机构地区：[1]南华大学核科学技术学院,衡阳421001 [2]重庆大学附属肿瘤医院放疗科,重庆400030

出　　处：《中华放射肿瘤学杂志》2024年第6期532-539,共8页Chinese Journal of Radiation Oncology

基　　金：沙坪坝区2023年科卫联合医学科研项目(2023SQKWLH006)。

摘　　要：目的:开发一种基于三维Transformer的深度学习架构,用于肺癌调强放疗(IMRT)计划的剂量分布预测。方法:回顾性分析重庆大学附属肿瘤医院2020年1月—2022年12月间174例肺癌行IMRT患者的资料,设置训练、验证、测试集分别为116、29、29例。通过Swin Unet Transformer(Swin Unetr)模型进行三维剂量分布预测训练,该模型使用带位移窗口的Transformer模块编码器,输入包括CT图像、计划靶区(PTV)和危及器官轮廓图像、射束信息图像和靶区距离图像。使用平均绝对误差(MAE)、戴斯相似性系数(DSC)以及剂量体积直方图(DVH)剂量学参数来评估模型的性能,并与其他3种深度学习模型CGAN、ResSEUnet、ResUnet进行比较。结果:Swin Unetr预测与原临床计划的剂量分布的MAE为0.0143±0.0055,CGAN、ResSEUnet和ResUnet分别为0.0162±0.0055、0.0167±0.0063和0.0164±0.0057。Swin Unetr在各等剂量值均取得了最高的DSC值(>0.85)。剂量学参数方面,除PTV的D 2%和心脏的D mean以外,其余Swin Unetr预测与原临床计划的剂量学参数差异均无统计学意义(P>0.05),且66.67%的总体剂量学参数和75%的PTV剂量学参数评估结果表现最佳。结论:在多个剂量学评估指标上,Swin Unetr取得了最佳评分,在各等剂量值上取得最高DSC。Swin Unetr在肺癌IMRT三维剂量预测方面的准确性较以往模型有显著改进。Objective To develop a deep learning architecture based on 3D Transformers to predict dose distribution within intensity modulated radiation therapy(IMRT)plans for lung cancer.Methods Clinical data of 174 lung cancer patients treated with IMRT in Chongqing University Cancer Hospital between January 2020 and December 2022 were retrospectively analyzed.All patients were divided into the training(n=116),validation(n=29),and test(n=29)sets.We employed the Swin Unet Transformer(Swin Unetr)model to predict the three-dimensional dose distribution.The model was trained using computed tomography(CT)images,planning target volume(PTV)images,organs at risk(OAR)images,beam configuration information images,and distance images.We used various evaluation metrics such as mean absolute errors(MAE),Dice similarity coefficients(DSC),and dose volume histogram(DVH)dosimetric parameters to assess the performance of Swin Unetr and compared it with three mainstream deep learning models:CGAN,ResSEUnet,and ResUnet.Results The MAE of the dose distribution prediction by Swin Unetr was recorded at 0.0143±0.0055.Conversely,the values of CGAN,ResSEUnet,and ResUnet were 0.0162±0.0055,0.0167±0.0063,and 0.0164±0.0057,respectively.Furthermore,Swin Unetr achieved the highest DSC values(>0.85)across all isodose volumes.Regarding DVH dosimetric parameters,excluding D2%of PTV and Dmean of the heart,Swin Unetr exhibited no statistically significant differences in the remaining DVH dosimetric parameters(all P>0.05),demonstrating the best evaluation results in 66.67%of the overall dosimetric parameters and 75%of the PTV dosimetric parameters.Conclusions Swin Unetr achieves the best score in multiple dosimetric evaluation indicators,and the highest DSC across all isodose volumes.Swin Unetr has significantly improved the accuracy of three-dimensional dose prediction during IMRT for lung cancer.

关 键 词：深度学习 剂量预测 肺肿瘤 调强放射疗法 

分 类 号：R734.2[医药卫生—肿瘤]