中华放射医学与防护杂志

王凯玥,薛娴,孙海涛,等.基于深度学习的子宫内膜癌术后近距离放疗危及器官自动勾画研究[J].中华放射医学与防护杂志,2025,45(10):958-965.Wang Kaiyue,Xue Xian,Sun Haitao,et al.Deep learning-based automatic segmentation of organs at risk in postoperative brachytherapy for endometrial carcinoma[J].Chin J Radiol Med Prot,2025,45(10):958-965

基于深度学习的子宫内膜癌术后近距离放疗危及器官自动勾画研究

Deep learning-based automatic segmentation of organs at risk in postoperative brachytherapy for endometrial carcinoma

投稿时间：2024-09-19

DOI：10.3760/cma.j.cn112271-20240919-00362

中文关键词: 深度学习子宫内膜肿瘤近距离放射治疗自动勾画危及器官

英文关键词:Deep learning Endometrial neoplasm Brachytherapy Automatic segmentation Organs at risk(OARs)

基金项目:国家重点研发计划(2022YFC2404606)；北京市自然科学基金(Z210008)；北京大学第三医院院临床重点项目(BYSYZD202011)

作者	单位	E-mail
王凯玥	北京大学第三医院肿瘤放疗科, 北京 100191
薛娴	中国疾病预防控制中心辐射防护与核安全医学所辐射防护与核应急中国疾病预防控制中心重点实验室, 北京 100088
孙海涛	北京大学第三医院肿瘤放疗科, 北京 100191
江萍	北京大学第三医院肿瘤放疗科, 北京 100191
王俊杰	北京大学第三医院肿瘤放疗科, 北京 100191	junjiewang_edu@sina.cn

摘要点击次数: 3548

全文下载次数: 644

中文摘要:

目的开发并评估一种用于子宫内膜癌术后近距离放疗危及器官(OAR)自动勾画的深度学习模型。方法收集2021年11月至2022年10月于北京大学第三医院接受高剂量率¹⁹²Ir源阴道残端腔内近距离放疗的108例子宫内膜癌术后患者的CT图像，并手动勾画直肠、结肠、小肠和膀胱。采用随机数表法划分出90例训练基于3D nnU-Net的分割模型，18例用于测试。几何指标采用戴斯相似系数(DSC)、豪斯多夫距离(HD)和平均表面距离(MSD)，剂量-体积参数(DVP)指标采用危及器官接受最大照射剂量的0.1、1.0和2.0 cm³体积接受的最小剂量(D_{0.1 cm³}、D_{1.0 cm³}和D_{2.0 cm³})，评价自动勾画的精确性及临床可用性。结果 3D nnU-Net分割直肠、结肠、小肠和膀胱的平均DSC分别为0.90，0.85，0.88和0.95，均优于对照的3D U-Net及V-Net模型，三者间差异具有统计学意义(F=21.78、24.33、36.00、20.11，P < 0.001)。3D nnU-Net分割结肠、小肠、膀胱的HD在3种方法间的差异有统计学意义(F=17.33、24.11、6.33，P < 0.05)。3D nnU-Net分割各器官的MSD均小于对照模型，差异具有统计学意义(F=29.78、27.11、27.11、14.78，P < 0.001)。3D nnU-Net勾画与手动勾画器官的DVP差异无统计学意义(P>0.05)。Bland-Altman分析显示了3D nnU-Net分割与手动分割DVP的一致性良好。结论基于3D nnU-Net算法的近距离放疗危及器官自动勾画模型具有较好的几何精度，与手动勾画的剂量一致性高，可用于提升临床工作效率。

英文摘要:

Objective To develop and assess a deep learning-based model for automatic segmentation of organs at risk (OARs) in postoperative brachytherapy for endometrial carcinoma (EC). Methods A retrospective study was conducted on the computed tomography (CT) images of 108 EC patients who received high-dose-rate (HDR) ¹⁹²Ir intracavitary vaginal-cuff brachytherapy (VCB) at the Peking University Third Hospital from November 2021 to October 2022. Then, the rectum, colon, small intestine, and bladder in these images were manually segmented. These patients were randomly divided into two groups using a random number table: 90 cases for training the 3D no-new-U-Net (nnU-Net) segmentation model and 18 cases for model testing. The precision and clinical applicability of the automatic segmentation model were assessed using geometric indexes including Dice similarity coefficient (DSC), Hausdorff distance (HD), and mean surface distance (MSD), as well as dose-volume parameters (DVPs) including the minimum dose to 0.1, 1.0, and 2.0 cm³ of OARs that received the highest irradiation doses (D_{0.1 cm³}, D_{1.0 cm³}, and D_{2.0 cm³}). Results The 3D nnU-Net model yielded mean DSC values of 0.90, 0.85, 0.88, and 0.95, respectively for the segmentations of the rectum, colon, small bowel, and bladder, all of which were better than those of the 3D U-Net and V-Net models. The differences among the three models were statistically significant (F = 21.78, 24.33, 36.00, 20.11, P < 0.001). The 3D nnU-Net exhibited statistically significant differences in HD values for the colon, small intestine, and bladder segmentations among the three method (F = 17.33, 24.11, 6.33, P < 0.05). The 3D nnU-Net model yielded lower MSD values for the segmentations of all organs compared to the control model, with statistically significant differences (F = 29.78, 27.11, 27.11, 14.78, P < 0.001). No statistically significant difference was found in all DVPs between the 3D nnU-Net model-based and manual segmentations (P > 0.05). Bland-Altman analysis demonstrated great consistency between the 3D nnU-Net and manual segmentations. Conclusions The 3D nnU-Net-based model exhibits high geometric accuracy and dosimetric consistency with manual segmentation of OARs in brachytherapy, holding potential to improve clinical efficiency.

HTML 查看全文查看/发表评论下载PDF阅读器

关闭