两种自动勾画方法对T3分期鼻咽癌靶区勾画结果对比分析

目的：比较和分析基于常规图谱集(Atlas)自动勾画方法和深度学习自动勾画方法对T₃分期鼻咽癌靶区的勾画结果。方法：回顾性选取本院T₃分期鼻咽癌患者138例，由一名高年资医生在CT上勾画GTV和CTV，勾画结果经另外两名高年资医生审核。建立3D-Unet模型，随机选取110例病例作为训练集，28例病例作为测试集。比较3D-Unet模型与Atlas模型的优劣性。结果：与医生勾画结果相比，3D-Unet模型的GTV和CTV平均勾画结果如下所示，平均表面距离：3.01和1.84 mm,95%豪斯多夫距离：16.05和7.70 mm,Dice相似性系数：0.71和0.83,Jaccard相似性系数：0.56和0.71，精确率：0.70和0.85，召回率：0.76和0.81，各项参数均显著优于Atlas模型(P<0.05)。结论：与Atlas自动勾画技术相比，基于3D-Unet模型的自动勾画方法对鼻咽癌靶区的勾画准确率明显提升。

Comparison of target segmentation results of two auto-segmentation methods for T3 nasopharyngeal carcinoma

Abstract: Objective To compare and analyze the accuracy of tumor segmentation using Atlas-based auto-segmentation method and deep learning method for T3 nasopharyngeal carcinoma (NPC). Methods A total of 138 cases of T3 NPC were selected retrospectively. GTV and CTV were outlined in CT by one senior doctor, and the delineation results were reviewed by two other senior doctors. A 3D-Unet model was established, with 110 cases as the training set and 28 cases as the test set. The performances of 3D-Unet model and Atlas models were compared. Results The ASD, 95%HD, DSC, JSC, precision, and recall rate of GTV and CTV segmentations using 3D-Unet model were 3.01 and 1.84 mm, 16.05 and 7.70 mm, 0.71 and 0.83, 0.56 and 0.71, 0.70 and 0.85, 0.76 and 0.81, respectively. The comparison of the above mentioned parameters revealed that 3D-Unet model was superior to Atlas models (P<0.05). Conclusion Compared with Atlas models, the 3D-Unet based auto-segmentation method significantly improves the accuracy of target segmentation in NPC.