【摘要】 目的:构建基于深度学习(deep learning, DL)的卷积神经网络模型，实现宫颈癌患者放射治疗计划的临床靶区体积(clinical target volume, CTV)和危及器官(organ at risks, OARs)自动勾画。方法:回顾性分析在福建省肿瘤医院行放射治疗的宫颈癌患者99例。对患者CT图像进行预处理，作为模型输入。设计一种基于DL的自动勾画模型，使用组合损失函数训练该模型。以医师手动勾画为度量标准，计算DL自动勾画模型下CTV靶区和膀胱、直肠、乙状结肠、左右骨髓、左右股骨头的准确率，并与基于图谱的自动勾画方法(atlas-based auto segmentation, ABAS)相比较。结果:DL模型在CTV靶区和7种危及器官(膀胱、直肠、乙状结肠、左右骨髓、左右股骨头)的戴斯系数分别为(0.85±0.02、0.94±0.04、0.87±0.03、0.67±0.14、0.85±0.03、0.87±0.03、0.87±0.06和0.87±0.06),95%豪斯多夫距离(mm)分别为(3.22±0.56、1.37±0.37、1.41±0.34、27.39±35.63、1.40±0.17、1.36±0.22、6.78±7.89和6.45±7.44),平均表面距离(mm)分别为(0.25±0.05、0.12±0.06、0.19±0.05、2.29±2.71、0.16±0.04、0.15±0.03、0.36±0.33和0.38±0.37)。DL勾画模型的戴斯系数均高于ABAS勾画模型。除乙状结肠外，DL勾画模型的95%豪斯多夫距离和平均表面距离均小于ABAS勾画模型。结论:提出的DL模型能较好地实现宫颈癌放疗临床靶区和危及器官的自动勾画，可为临床医师勾画提供初步参考，节省临床靶区和危及器官勾画的时间。更多还原

【Abstract】 Objective:A convolutional neural network model based on deep learning(DL) was constructed to realize the automatic contouring of clinical target volume(CTV) and organs at risk(OARs) of radiotherapy plan for cervical cancer patients.Methods:An analysis of 99 patients with cervical cancer who underwent radiotherapy in Fujian Cancer Hospital was conducted retrospectively.An automatic contouring model based on DL was designed and trained by combinatorial loss function.The accuracy of CTV target, bladder, rectum, sigmoid colon, left and right marrow, and left and right femoral head under the DL automatic contour model was calculated and compared with the accuracy of atlas-based auto segmentation(ABAS) method.Results:The Dice coefficient of DL model in CTV and senven OARs(bladder, rectum, sigmoid colon, left and right marrow, left and right femoral head) were(0.85±0.02,0.94±0.04,0.87±0.03,0.67±0.14,0.85±0.03,0.87±0.03,0.87±0.06 and 0.87±0.06) respectively, 95 pesrcent Hausdorf distances(mm) were(3.22 ±0.56,1.37±0.37,1.41±0.34,27.39±35.63,1.40±0.17,1.36±0.22,6.78±7.89 and 6.45±7.44) respectively and the mean surface distances(mm) were(0.25±0.05,0.12±0.06,0.19±0.05,2.29±2.71,0.16±0.04,0.15±0.03,0.36±0.33 and 0.38±0.37) respectively.The Dice coefficient of DL automatic contour model was higher than of ABAS.Except for sigmoid colon, 95 percent Hausdorf distance and Mean surface distance of DL automatic contour model were smaller than of ABAS.Conclusion:The proposed DL model can realize the automatic contouring of CTV and OARs in radiotherapy for cervical cancer, which can provide preliminary reference for clinicians and save the time of delineation os CTV ans OARs.更多还原