金属伪影对鼻咽癌放疗危及器官自动勾画的影响

目的:评估CT金属伪影对鼻咽癌放疗危及器官（OAR）自动勾画的影响。方法:选取有无牙齿修复物的鼻咽癌患者各16例，由放疗医师和深度学习自动勾画平台AccuContour分别勾画26种OAR轮廓。比较有无金属伪影患者不同OAR轮廓三维相似性系数（DSC）和Hausdorff距离（HD）以及有无金属伪影横断面内口腔和下颌骨的二维DSC和HD。同时记录人工勾画和自动勾画全部OAR的时间。结果:所有OAR的三维DSC和HD在有无金属伪影患者组间均无显著差异（P>0.05）。无金属伪影横断面内口腔的二维DSC和HD优于有伪影横断面（P<0.01），且伪影越严重，自动勾画的口腔轮廓局部偏离基准值越明显。自动勾画效率（<2 min）显著优于人工勾画效率（>70 min）。结论:牙齿修复物伪影对基于深度学习的鼻咽癌放疗OAR自动勾画的准确性和工作效率影响有限，较人工勾画方法仍然具备明显优势。

Effects of metal artifacts on automatic segmentation of organs-at-risk in patients receiving radiotherapy for nasopharyngeal carcinoma

Abstract: Objective To evaluate the effects of CT metal artifacts on the automatic segmentation of organs-at-risk (OAR) in nasopharyngeal carcinoma (NPC) patients treated with radiotherapy. Methods Two groups of NPC patients with or without dental restorations were enrolled, with 16 patients in each group, and a total of 26 kinds of OAR were segmented by an experienced radiation oncologist and a deep learning-based automatic segmentation platform (AccuContour), separately. The three-dimensional Dice similarity coefficient (DSC) and Hausdorff distance (HD) of different OAR in patients with or without metal artifacts were compared, and moreover, the two-dimensional DSC and HD of mandibles and oral cavity on axial slices with or without metal artifacts were also compared. Meanwhile, the time taken for manual and automatic segmentations was recorded. Results There were no significant differences in three-dimensional DSC and HD for all OAR between patients with metal artifacts and those without metal artifacts (P>0.05). The two-dimensional DSC and HD of the oral cavity on axial slices without metal artifacts were better than those on axial slices with metal artifacts (P<0.01). The more severe the artifact was, the greater the deviation of automatically segmented contours of oral cavity from the ground truth was. The efficiency of automatic segmentation was significantly higher than manual segmentation (<2 min vs >70 min). Conclusion Dental restoration artifacts has limited effects on the accuracy and efficiency of the deep learning-based automatic segmentation of OAR in radiotherapy for NPC, and automatic segmentation still has distinct advantages over manual segmentation.