算法和匹配数目对宫颈癌危及器官自动勾画的影响

目的:探讨基于Atlas实施宫颈癌危及器官自动勾画时勾画算法及匹配数目对自动勾画结果的影响。方法:基于MIM-Maestro软件建立宫颈癌Atlas模板库，入库病例数目为60例。随机选择Atlas库外10例宫颈癌患者，由临床医生手动勾画危及器官（膀胱、直肠和双侧股骨头），并定义为参考勾画（Vref）。应用多数投票算法和STAPLE算法，匹配数分别选择1、3、5、7、9进行自动勾画。采用勾画时间（T）、相似性系数（DSC）、敏感性指数（SI）、质心偏差（DC）、Jaccard系数（JAC）、Hausdorff距离（HD）评价勾画结果，并进行单因素方差分析和配对样本t检验。结果:勾画时间随匹配数目增大呈线性增加，与勾画算法无关。多数投票算法和STAPLE算法勾画结果均显示，匹配数为1时膀胱的SI和左股骨头的DSC、HD、JAC与匹配数为3、5、7、9时有统计学差异。STAPLE算法中，直肠和双侧股骨头的SI均显示匹配数目为1、3与5、7、9有统计学差异。两种勾画算法的比较结果显示，仅双侧股骨头的SI有统计学差异。结论:基于Atlas实施危及器官自动勾画时，勾画算法对结果基本无影响，所需时间与匹配数呈正比，综合勾画结果建议匹配数目选择3。

Effects of algorithm and matching number on the auto-segmentation of organs-at-risk in patients with cervical cancer

To evaluate the effects of segmentation algorithms and matching numbers on the Atlas-based auto-segmentation of organs-at-risk (OAR) in patients with cervical cancer. Methods The Atlas database of cervical cancer which contains 60 cases was established with MIM-Maestro software. Another 10 patients with cervical cancer were randomly selected out of the Atlas database. The manual segmentations of OAR, namely bladder, rectum and bilateral femoral head, by an experienced radiation oncologist, were defined as the reference volume (Vref). Subsequently, the OAR were automatically segmented by majority vote algorithm versus STAPLE algorithm, with the matching number of 1, 3, 5, 7 and 9. The segmentation results were assessed using time for segmentation (T), Dice similarity coefficient (DSC), sensitivity index (SI), deviation of centroid (DC), Jaccard coefficient (JAC) and Hausdorff distance (HD). The results were also analyzed with one-way analysis of variance and paired sample t test. Results The time for segmentation was independent of the algorithm and increased linearly with the increasing of matching number. The results obtained by majority vote algorithm and STAPLE algorithm showed that there were statistical differences in the SI of bladder and the DSC, HD, JAC of left femoral head between the matching number of 1 and 3, 5, 7, 9. For STAPLE algorithm, the SI of bladder and bilateral femoral head showed statistical differences between the matching number of 1, 3 and 5, 7, 9. Moreover, significant difference was only found in the SI of bilateral femoral head between two algorithms. Conclusion The effect of segmentation algorithms on the Atlas-based OAR auto-segmentation is trivial. The time for segmentation was positively correlated with matching number. With the comprehensive consideration of segmentation results, the matching number of 3 is recommended for clinical application.