| 赵博,张晗,朱海涛,孙瑞佳,史燕杰,孙应实.基于3D nnU-Net联合增强CT影像组学预测食管鳞状细胞癌T分期[J].中国医学影像技术,2025,41(12):2011~2015 |
| 基于3D nnU-Net联合增强CT影像组学预测食管鳞状细胞癌T分期 |
| 3D nnU-Net combined with contrast enhanced CT radiomics for predicting T stage of esophageal squamous cell carcinoma |
| 投稿时间：2025-09-09 修订日期：2025-12-02 |
| DOI：10.13929/j.issn.1003-3289.2025.12.016 |
| 中文关键词: 食管肿瘤 癌,鳞状细胞 肿瘤分期 体层摄影术,X线计算机 影像组学 深度学习 |
| 英文关键词:esophageal neoplasms carcinoma, squamous cell neoplasm staging tomography, X-ray computed radiomics deep learning |
| 基金项目:北京市自然科学基金(7244511)、临床医学发展专项"扬帆"计划(ZLRK202522)。 |
| 作者 | 单位 | E-mail | | 赵博 | 北京大学肿瘤医院暨北京市肿瘤防治研究所医学影像科, 恶性肿瘤发病机制及转化研究教育部重点实验室, 北京 100142 | | | 张晗 | 北京大学肿瘤医院暨北京市肿瘤防治研究所医学影像科, 恶性肿瘤发病机制及转化研究教育部重点实验室, 北京 100142 | | | 朱海涛 | 北京大学肿瘤医院暨北京市肿瘤防治研究所医学影像科, 恶性肿瘤发病机制及转化研究教育部重点实验室, 北京 100142 | | | 孙瑞佳 | 北京大学肿瘤医院暨北京市肿瘤防治研究所医学影像科, 恶性肿瘤发病机制及转化研究教育部重点实验室, 北京 100142 | | | 史燕杰 | 北京大学肿瘤医院暨北京市肿瘤防治研究所医学影像科, 恶性肿瘤发病机制及转化研究教育部重点实验室, 北京 100142 | | | 孙应实 | 北京大学肿瘤医院暨北京市肿瘤防治研究所医学影像科, 恶性肿瘤发病机制及转化研究教育部重点实验室, 北京 100142 | sys27@163.com |
|
| 摘要点击次数: 111 |
| 全文下载次数: 18 |
| 中文摘要: |
| 目的 观察基于3D nnU-Net联合增强CT影像组学预测食管鳞状细胞癌(ESCC)T分期的价值。方法 回顾性分析429例ESCC的胸部增强静脉期CT,包括训练集(n=236,含T1~T2期119例、T3期117例)、验证集(n=102,含T1~T2期53例、T3期49例)及测试集(n=91,含T1~T2期53例、T3期38例)。于训练集图像中手动勾画肿瘤ROI用于训练3D nnU-Net,采用戴斯相似系数(DSC)、平均表面距离(MSD)、95%豪斯多夫距离(HD95)、精确率和召回率评估其在验证集中的分割效能。于3D nnU-Net在CT图像中自动分割的肿瘤区域提取、筛选影像组学特征并建立影像组学模型,以受试者工作特征曲线下面积(AUC)评估其预测ESCC T分期的效能。结果 基于3D nnU-Net自动分割验证集ESCC的DSC、MSD、HD95、精确率和召回率分别为0.74、4.30 mm、10.98 mm、79.14%及74.65%,形态、范围均与人工勾画的ROI较为一致。影像组学模型评估训练集、验证集和测试集ESCC T分期的AUC分别为0.845、0.837及0.832,拟合度及临床净收益均良好。结论 基于3D nnU-Net联合增强CT影像组学可有效预测ESCC T分期。 |
| 英文摘要: |
| Objective To observe the value of 3D nnU-Net combined with contrast enhanced CT radiomics for predicting T stage of esophageal squamous cell carcinoma (ESCC). Methods Venous phase chest contrast enhanced CT images of 429 patients with ESCC were retrospectively analyzed. The patients were divided into training set (n=236, including 119 cases of T1—T2 stage and 117 cases of T3 stage), validation set (n=102, including 53 cases of T1—T2 stage and 49 cases of T3 stage) and test set (n=91, including 53 cases of T1—T2 stage and 38 cases of T3 stage). 3D nnU-Net model was trained using ROI of tumors manually delineated in images in training set, then its segmentation performance in validation set was evaluated using Dice similarity coefficient (DSC), mean surface distance (MSD), 95% Hausdorff distance (HD95), precision and recall. Radiomics features were extracted and selected from tumor regions automatically segmented with 3D nnU-Net in CT images, then a radiomics model was constructed, and its performance for assessing ESCC T stage was evaluated using the area under the receiver operating characteristic curve (AUC). Results DSC, MSD, HD95, precision and recall of 3D nnU-Net for automatic segmentation of ESCC in validation set was 0.74, 4.30 mm, 10.98 mm, 79.14% and 74.65%, respectively, both the morphology and scope were highly consistent to manually delineated ROI. AUC of the radiomics model for predicting ESCC T stage was 0.845, 0.837 and 0.832 in training, validation and test sets, respectively, with good fitting degree and clinical net benefit. Conclusion 3D nnU-Net combined with contrast enhanced CT radiomics could effectively predict T stage of ESCC. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
