| 王思斌,赵毅,杜鹏,张晶,肖越勇,张肖.基于DeepSurv模型预测Ⅰ期非小细胞肺癌经CT引导下射频消融后预后[J].中国医学影像技术,2026,42(3):336~341 |
| 基于DeepSurv模型预测Ⅰ期非小细胞肺癌经CT引导下射频消融后预后 |
| DeepSurv model for predicting prognosis of stage Ⅰ non-small cell lung cancer after CT-guided radiofrequency ablation |
| 投稿时间：2025-12-30 修订日期：2026-03-14 |
| DOI：10.13929/j.issn.1003-3289.2026.03.004 |
| 中文关键词: 癌,非小细胞肺 射频消融 预后 影像组学 |
| 英文关键词:carcinoma,non-small-cell lung radiofrequency ablation prognosis radiomics |
| 基金项目:军委后勤保障部卫生局技术产品研究项目(24BJZ16)。 |
|
| 摘要点击次数: 107 |
| 全文下载次数: 38 |
| 中文摘要: |
| 目的 观察基于深度生存神经网络(DeepSurv)模型预测Ⅰ期非小细胞肺癌(NSCLC)经CT引导下射频消融(RFA)后预后的价值。方法 回顾性纳入217例接受CT引导下RFA的Ⅰ期NSCLC患者,分为训练集(n=109)、内部验证集(n=46)及外部验证集(n=62)。采用nnU-Net架构TotalSegmentator工具分别于CT图像中自动分割肿瘤区,获得治疗前肿瘤区(GTR)及治疗后消融区(PTZ),经DeedsBCV算法弹性配准后,以GTR外扩5 mm为理想消融靶区计算三维目标覆盖率(TCR);于PTZ中提取并筛选最优影像组学特征,计算影像组学评分(Radscore)。以多因素Cox回归分析筛选局部肿瘤进展(LTP)的独立影响因素,联合Radscore输入DeepSurv模型,以一致性指数(C-index)及Brier评分评估模型效能及校准度。结果 性别、纤维蛋白原(FIB)及三维TCR均为LTP独立影响因素[HR (95%CI)=1.568(1.020,2.410)、1.540(1.170,2.027)、0.110(0.044,0.273),P均＜0.05]。将性别、FIB、三维TCR及Radscore输入DeepSurv模型,其预测训练集、内部验证集及外部验证集Ⅰ期NSCLC经RFA后LTP的C-index分别为0.906、0.809及0.861,Brier评分分别为0.187、0.177及0.182,提示其效能及校准度均良好。结论 基于DeepSurv模型融合肿瘤影像组学特征、三维TCR及患者性别、FIB能准确预测Ⅰ期NSCLC经CT引导下RFA后预后。 |
| 英文摘要: |
| Objective To observe the value of deep survival neural network (DeepSurv) model for predicting prognosis of stage Ⅰ non-small cell lung cancer (NSCLC) after CT-guided radiofrequency ablation (RFA). Methods Totally 217 patients with stage Ⅰ NSCLC who underwent CT-guided RFA were retrospectively enrolled and divided into training set (n=109), internal validation set (n=46) and external validation set (n=62). Pre-treatment gross tumor region (GTR) and post-treatment ablation region (PTZ) were automatically segmented on CT images using TotalSegmentator tool based on nnU-Net architecture. After deformable registration with DeedsBCV algorithm, a 5 mm isotropic expansion of GTR was used to construct the ideal ablation target zone, and three-dimensional target coverage ratio (TCR) was calculated. Optimal radiomics features were extracted from PTZ to derive radiomics score (Radscore). Multivariable Cox regression was performed to identify independent impact factors of local tumor progression (LTP). These factors, together with Radscore, were incorporated into DeepSurv model. Model discrimination and calibration were assessed using concordance index (C-index) and Brier score. Results Gender, fibrinogen (FIB), and three-dimensional TCR were all independent impact factors of LTP (HR[95%CI]=1.568[1.020, 2.410], 1.540[1.170, 2.027], and 0.110[0.044, 0.273], respectively; all P＜0.05) and were incorporated into DeepSurv model. C-index value of DeepSurv model for predicting LTP after CT-guided RFA for stage Ⅰ NSCLC was 0.906, 0.809 and 0.861 in training, internal validation and external validation sets, respectively, the corresponding Brier score was 0.187, 0.177 and 0.182, indicating good predictive performance and calibration. Conclusion DeepSurv model integrating tumor radiomics features, three-dimensional TCR, gender and FIB could be used to accurately predict prognosis of stage Ⅰ NSCLC after CT-guided RFA. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
