摘要目的:基于机器学习提出可应用于低图像质量、多叶准直器(MLC)遮挡和非刚性变形兆伏级(MV)图像的无标记射束方向观(BEV)肿瘤放疗跟踪算法。方法:采用窗口模板匹配法和Voxelmorph端到端无监督网络，处理MV图像中的配准问题。使用动态胸部模体，验证肿瘤跟踪算法的准确性。将模体质量保证(QA)计划在加速器上手动设置治疗偏移后执行，收集治疗过程中的682幅电子射野影像系统(EPID)图像作为固定图像；同时采集计划系统中对应射野角度的数字影像重建(DRR)图作为浮动图像，进行靶区跟踪研究。收集21例肺部肿瘤放疗的533对EPID和DRR图像进行肿瘤跟踪研究，提供治疗过程中肿瘤位置变化定量结果。图像相似度用于算法的第三方验证。结果:算法可应对不同程度(10%~80%)的图像缺失，且对数据缺失图像的非刚性配准表现较好。模体验证中86.8%的跟踪误差<3 mm，<2 mm的比例约80%作用。配准后标准化互信息(NMI)由1.18±0.02提高到1.20±0.02( t＝－6.78， P＝0.001)。临床病例肿瘤运动以平移为主，平均位移3.78 mm，最大位移可达7.46 mm。配准结果显示存在非刚性形变，配准后NMI由1.21±0.03增至到1.22±0.03( t＝－2.91， P＝0.001)。 结论:肿瘤跟踪算法跟踪精度可靠且鲁棒性好，可用于无创、实时、无额外设备和辐射剂量的肿瘤跟踪。

abstractsObjective:To propose a machine learning-based markerless beam′s eye view (BEV) tumor tracking algorithm that can be applied to low-quality megavolt (MV) images with multileaf collimator (MLC)-induced occlusion and non-rigid deformation.Methods:This study processed the registration of MV images using the window template matching method and end-to-end unsupervised network Voxelmorph and verified the accuracy of the tumor tracking algorithm using dynamic chest models. Phantom QA plans were executed after the treatment offset was manually set on the accelerator, and 682 electronic portal imaging device (EPID) images obtained during the treatment were collected as fixed images. Moreover, the digitally reconstructed radiography (DRR) images corresponding to the portal angles in the planning system were collected as floating images for the study of target volume tracking. In addition, 533 pairs of EPID and DRR images of 21 lung tumor patients treated with radiotherapy were collected to conduct the study of tumor tracking and provide quantitative result of changes in tumor locations during the treatment. Image similarity was used for third-party validation of the algorithm.Results:The algorithm could process images with different degrees (10%-80%) of data missing and performed well in non-rigid registration of images with data missing. As shown by the phantom verification, 86.8% and 80% of the tracking errors were less than 3 mm and less than 2 mm, respectively, and the normalized mutual information (NMI) varied from 1.18 ± 0.02 to 1.20 ± 0.02 after registration ( t = -6.78, P = 0.001). The tumor motion of the clinical cases was dominated by translation, with an average displacement of 3.78 mm and a maximum displacement of 7.46 mm. The registration result of the cases showed the presence of non-rigid deformations, and the corresponding NMI varied from 1.21 ± 0.03 before registration to 1.22 ± 0.03 after registration ( t = -2.91, P = 0.001). Conclusions:The tumor tracking algorithm proposed in this study has reliable tracking accuracy and high robustness and can be used for non-invasive and real-time tumor tracking requiring no additional equipment and radiation dose.