| 刘博文,王霄霄,卢超,王芷旋,张久楼,王泽辉,陆思远,蒋小玥,戚明瑶,潘冬刚,单秀红.CT虚拟单能量图、非线性融合图及混合能量图3D-nnU-Net模型自动分割进展期胃癌效能[J].中国医学影像技术,2025,41(5):753~758 |
| CT虚拟单能量图、非线性融合图及混合能量图3D-nnU-Net模型自动分割进展期胃癌效能 |
| Efficacy of 3D-nnU-Net model of CT virtual monoenergetic images, non-linear blending images and mixed-energy images for automatically segmenting advanced gastric cancer |
| 投稿时间：2024-07-25 修订日期：2025-03-31 |
| DOI：10.13929/j.issn.1003-3289.2025.05.013 |
| 中文关键词: 胃肿瘤 人工智能 体层摄影术,X线计算机 图像分割 |
| 英文关键词:stomach neoplasms artificial intelligence tomography, X-ray computed image segmentation |
| 基金项目:2023年度镇江市健康与生命科学重点实验室重点课题(GZSYS202302)。 |
| 作者 | 单位 | E-mail | | 刘博文 | 江苏大学附属人民医院医学影像科, 江苏 镇江 212002 | | | 王霄霄 | 江苏大学附属人民医院医学影像科, 江苏 镇江 212002 | | | 卢超 | 江苏大学附属人民医院医学影像科, 江苏 镇江 212002 | | | 王芷旋 | 江苏大学附属人民医院医学影像科, 江苏 镇江 212002 | | | 张久楼 | 南京医科大学第一附属医院医学影像科, 江苏 南京 210029 | | | 王泽辉 | 镇江中澳人工智能研究院, 江苏 镇江 212021 | | | 陆思远 | 江苏大学附属人民医院医学影像科, 江苏 镇江 212002 | | | 蒋小玥 | 江苏大学附属人民医院医学影像科, 江苏 镇江 212002 | | | 戚明瑶 | 江苏大学附属人民医院医学影像科, 江苏 镇江 212002 | | | 潘冬刚 | 江苏大学附属人民医院医学影像科, 江苏 镇江 212002 | | | 单秀红 | 江苏大学附属人民医院医学影像科, 江苏 镇江 212002 | 13913433095@163.com |
|
| 摘要点击次数: 65 |
| 全文下载次数: 12 |
| 中文摘要: |
| 目的 基于3D-nnU-Net构建自动分割CT虚拟单能量图(VMI)、非线性融合图(NLBI)及混合能量图(MEI)中的进展期胃癌(AGC)模型,并对比其分割效能。方法 回顾性分析216例AGC,以其中185例构建、训练及验证模型,按5 ∶ 1比例划分训练集(n=154)与测试集(n=31);以另31例为验证集,评估模型泛化性。基于全腹双能量模式静脉期CT重建70 keV能级VMI(VMI70 keV)、NLBI及MEI,以3D-nnU-Net分别构建AGC自动分割模型(VMI70 keV、NLBI及MEI模型)。以手动分割结果为金标准,采用戴斯相似系数(DSC)、交并比(IoU)及平均对称表面距离(ASSD)评估各模型分割测试集及验证集病灶及其中T2期病灶的效能。结果 三种模型分割测试集和验证集所有病灶的DSC均＞0.80,VMI70 keV及NLBI模型的DSC及IoU均高于而ASSD低于MEI模型(P均＜0.05)。对测试集和验证集中10个T2期AGC(n均=5),MEI模型的DSC均低于VMI70 keV及NLBI模型、IoU低于VMI70 keV模型而ASSD高于NLBI模型(P＜0.05)。结论 基于3D-nnU-Net的VMI70 keV、NLBI及MEI模型均能有效分割CT所示AGC,前二者效能更优。 |
| 英文摘要: |
| Objective To compare the segmenting efficacy of automatic segmentation models for advanced gastric cancer (AGC) on CT virtual monoenergetic images (VMI), non-linear blending images (NLBI) and mixed-energy images (MEI) based on 3D-nnU-Net. Methods Totally 216 cases of AGC were retrospectively enrolled, among them 185 cases were used to construct, train and validate models and divided into training set (n=154) and test set (n=31) at the ratio of 5 ∶ 1, while the other 31 cases were used as validation set to evaluate the generalization of the models. The 70 keV energy level VMI (VMI70 keV), NLBI and MEI were reconstructed with whole-abdominal dual-energy mode venous CT, and automatic segmentation models of AGC, including VMI70 keV, NLBI and MEI models were constructed using 3D-nnU-Net, respectively. Taken manually segmented results as golden standards, the efficacy of each model for segmenting all lesions and T2 stage lesions in test set and validation set were evaluated using Dice similarity coefficient (DSC), intersection over union (IoU) and average symmetric surface distance (ASSD). Results For all lesions in test and validation sets, DSC of 3 models were all ＞0.80. DSC and IoU of VMI70 keV and NLBI models were both higher, while their ASSD was lower than those of MEI model (all P＜0.05). For T2 stage AGC in both test set and validation set (each n=5), DSC of MEI model was lower than that of VMI70 keV and NLBI models (both P＜0.05), while IoU of MEI model was lower than that of VMI70 keV model (P＜0.05), and its ASSD was higher than that of NLBI model (P＜0.05). Conclusion All 3D-nnU-Net-based VMI70 keV, NLBI and MEI models could effectively segment AGC on dual-energy CT images, and the segmentation efficacy of the former two were better. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
