摘要
目的对比基于UNet与ResUNet++模型的宫颈癌放射治疗危及器官自动分割效果。方法在PyTorch平台搭建UNet与ResUNet++模型。将2023年6月至2024年2月于医院行放射治疗的232例宫颈癌患者的治疗计划作为研究对象,其中194例计划用于模型的训练和验证,38例计划用于测试。危及器官包括肝、膀胱、直肠、脊髓、肾、股骨、股骨头。使用3D-戴斯相似性系数(3D-DSC)及95%豪斯多夫距离(HD95%)评估2种模型的分割结果。结果UNet模型分割结果显示,直肠的3D-DSC较低,为0.847(0.809,0.868),其他危及器官的3D-DSC均较高,位于0.938(0.929,0.945)至0.978(0.975,0.979)范围内;肝和膀胱的HD95%较高,分别为11.449(8.822,13.740)和13.038(11.365,15.699),其他危及器官的HD95%均位于2.638(2.341,2.812)至6.424(5.502,8.071)范围内。ResUNet++模型分割结果显示,直肠的3D-DSC较低,为0.792(0.707,0.855),其他危及器官的3D-DSC均位于0.929(0.876,0.950)至0.977(0.976,0.979)范围内;肝和膀胱的HD95%较高,分别为10.954(8.552,13.460)和13.114(11.066,16.664),其他危及器官的HD95%均位于2.640(2.161,3.029)至6.824(6.050,8.066)范围内。2种模型分割的肝、右肾3D-DSC比较,差异无统计学意义(P>0.05);2种模型分割的其他器官的3D-DSC比较,差异均有统计学意义(P<0.05)。UNet模型分割的左股骨头HD95%低于ResUNet++模型,差异有统计学意义(P<0.05);其余器官的HD95%比较,差异均无统计学意义(P>0.05)。结论UNet与ResUNet++模型均可进行宫颈癌放射治疗危及器官的自动分割,且UNet模型的整体分割效果好于ResUNet++模型。
Objective To compare the automatic segmentation effects for organs at risk in cervical cancer radiotherapy based on UNet and ResUNet++models.Methods UNet and ResUNet++models are built on the PyTorch platform.With the treatment plan of 232 cervical cancer patients underwent radiation therapy in hospitals from June 2023 to February 2024,among them,194 cases were planned for model training and validation,and 38 cases were planned for testing.Organs at risk include the liver,bladder,rectum,spinal cord,kidney,femur,and femoral head.The 3 Dimensions dice similarity coefficient(3D-DSC)and hausdorff distance 95th percentile(HD95%)were used to evaluate the segmentation results of the two models.Results According to the segmentation results of UNet model,the 3D-DSC of rectum was relatively low,which was 0.847(0.809,0.868).The 3D-DSC values of other organs at risk were higher,ranging from 0.938(0.929,0.945)to 0.978(0.975,0.979).The HD95%of liver and bladder were higher,which were 11.449(8.822,13.740)and 13.038(11.365,15.699),respectively.The HD95%of other organs at risk were within the range of 2.638(2.341,2.812)to 6.424(5.502,8.071).The segmentation results of the ResUNet++model showed that the 3D-DSC of the rectum was relatively low,which was 0.792(0.707,0.855).The 3D-DSC of other organs at risk is relatively high,ranging from 0.938(0.929,0.945)to 0.978(0.975,0.979);The HD95%of the liver and bladder were relatively high,which were 10.954(8.552,13.460)and 13.114(11.066,16.664),respectively.The HD95%of other organs at risk ranged from 2.640(2.161,3.029)to 6.824(6.050,8.066).There were no statistically significant differences in the 3D-DSC of the liver and right kidney segmented by the two models(P>0.05);The differences in 3D-DSC of other organs segmented by the two models were statistically significant(P<0.05).The HD95%of the left femoral head segmented by the UNet model was lower than that of the ResUNet++model,and the difference was statistically significant(P<0.05);There was no significant difference in HD95%of other organs(P>0.05).Conclusion Both UNet and ResUNet++models can automatically segment organs at risk in cervical cancer radiotherapy,and the overall segmentation effect of the UNet model is better than that of the ResUNet++model from the data evaluation.
作者
柏朋刚
王国华
陈榕钦
陈济鸿
陈文娟
林家帆
欧阳敏
Bai Penggang;Wang Guohua;Chen Rongqin;Chen Jihong;Chen Wenjuan;Lin Jiafan;Ouyang Min(Clinical Oncology School of Fujian Medical University,Fujian Cancer Hospital,Fuzhou Fujian 350014,China;School of Nuclear Science and Technology,University of South China,Hengyang Hunan 421001,China)
出处
《医疗装备》
2024年第13期1-6,共6页
Medical Equipment
基金
福建省科技计划项目(2022Y0056,2020J011122,2021Y0052)
福建省卫生健康科技项目(2021CXB013,2018-ZQN-19)
福建省科技联合创新项目(2021Y9190)
福建省消化、呼吸及泌尿生殖系统恶性肿瘤放射与治疗临床医学研究中心(2021Y2014)。
