| 贾建业,谢洪燕,王佳豪,等.基于CT弹性成像的深度学习模型预测食管鳞癌新辅助放化疗后病理完全缓解状态[J].中华放射医学与防护杂志,2026,46(9):878-886.Jia Jianye,Xie Hongyan,Wang Jiahao,et al.A deep learning model based on computed tomography elastography for predicting pathological complete response after neoadjuvant chemoradiotherapy in esophageal squamous cell carcinoma[J].Chin J Radiol Med Prot,2026,46(9):878-886 |
| 基于CT弹性成像的深度学习模型预测食管鳞癌新辅助放化疗后病理完全缓解状态 |
| A deep learning model based on computed tomography elastography for predicting pathological complete response after neoadjuvant chemoradiotherapy in esophageal squamous cell carcinoma |
| 投稿时间：2025-11-12 |
| DOI：10.3760/cma.j.cn112271-20251112-00398 |
| 中文关键词: 食管鳞癌 新辅助放化疗 病理完全缓解 CT弹性成像 深度学习 |
| 英文关键词:Esophageal squamous cell carcinoma Neoadjuvant chemoradiotherapy Pathological complete response Computed tomography elastography Deep learning |
| 基金项目: |
|
| 摘要点击次数: 226 |
| 全文下载次数: 2 |
| 中文摘要: |
| 目的 构建并验证一种融合CT弹性成像的深度学习模型，用于术前预测食管鳞癌（ESCC）患者新辅助放化疗（nCRT）后的病理完全缓解（pCR）状态。方法 回顾性纳入2020年1月至2025年1月首都医科大学附属北京友谊医院和南京医科大学附属淮安第一医院的482例ESCC患者，术前均行增强CT扫描。基于差分仿射不变性原理，结合单调信号抑制映射对CT图像进行归一化与增强处理，生成结构化弹性特征图像，并提取肿瘤感兴趣（ROI）区域内的弹力均值 （E_mean）作为定量参数。随后，将增强CT图像及弹性特征图输入SimpViT模型中提取深度特征，并基于Transformer注意力机制特征融合模块捕获多图像空间交互关系，构建多参数融合模型。通过曲线下面积（AUC）、校准曲线和临床决策曲线（DCA）对模型性能进行综合评价。梯度加权类激活映射（Grad-CAM）用于模型的可视化分析。结果 相较于单一参数图像模型，多参数融合模型表现出了更佳的预测效能，在内部和外部验证队列中的AUC值达到了0.888（95%CI: 0.794 ~ 0.982）及0.863（95%CI: 0.781 ~ 0.945）。单因素与多因素逻辑回归筛选出T分期、E_mean为临床独立危险因素。最后，在此基础上构建的联合模型表现出更佳的效能，其内部和外部验证队列的AUC值分别为0.906（95%CI: 0.838 ~ 0.973）及0.879（95%CI: 0.803 ~ 0.956）。DCA曲线显示，在合理阈值范围内，联合模型临床净获益明显优于其他模型。Grad-CAM可视化显示，模型主要关注于肿瘤边界及周缘区域，与实际病理响应区域具备一致性。结论 基于CT弹性成像的深度学习模型可在术前非侵入性地预测ESCC患者nCRT后的pCR状态，表现出良好的预测准确性和临床净获益潜力，有望为术前评估与个体化治疗策略的制定提供有力影像支持。 |
| 英文摘要: |
| Objective To develop and validate a deep learning model integrating computed tomography (CT) elastography for preoperative prediction of pathological complete response status in patients with esophageal squamous cell carcinoma (ESCC) following neoadjuvant chemoradiotherapy.Methods A total of 482 patients with ESCC from Beijing Friendship Hospital, Capital Medical University and The Affiliated Huai′an No. 1 People′s Hospital of Nanjing Medical University who underwent preoperative contrast-enhanced CT scans were retrospectively enrolled. CT images were normalized and enhanced using a combination of the differential affine invariance principle and monotonic signal suppression mapping to generate structured elastography feature maps. The mean elasticity within the tumor region of interest was extracted as a quantitative parameter. Subsequently, the contrast-enhanced CT images and the corresponding elastography feature maps were input into the SimpViT model to extract deep features. A multi-parameter fusion model was constructed by capturing spatial interactions across multiple images using a Transformer attention-based feature fusion module. Model performance was comprehensively evaluated using the area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve analysis. Gradient-weighted Class Activation Mapping was employed for visual analysis of the model.Results Compared with the single-parameter image model, the multi-parameter fusion model exhibited better predictive performance, achieving AUC values of 0.888 (95%CI: 0.794-0.982) and 0.863 (95%CI: 0.781-0.945) in the internal and external validation cohorts, respectively. Univariable and multivariable logistic regression analyses identified T stage and mean elasticity as independent risk factors. Finally, the combined model constructed on this basis demonstrated better efficacy, with AUC values of 0.906 (95%CI: 0.838-0.973) and 0.879 (95%CI: 0.803-0.956) in the internal and external validation cohorts, respectively. Decision curve analysis showed that within a reasonable threshold range, the clinical net benefit of the combined model was clearly superior to other models. Gradient-weighted Class Activation Mapping visualization demonstrated that the model primarily focused on the tumor margins and peritumoral regions, aligning well with the actual pathological response zones.Conclusions The deep learning model based on CT elastography enables non-invasive, preoperative prediction of pathological complete response status in patients with ESCC after neoadjuvant chemoradiotherapy, demonstrating good predictive accuracy and potential clinical net benefit. This approach may offer strong imaging support for preoperative evaluation and the development of personalized treatment strategies. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|