基于临床特征与CT深度学习的膀胱癌根治性 膀胱切除术后总生存预测研究

A Study on Predicting Overall Survival after Radical Cystectomy for Bladder Cancer Using Clinical Characteristics and CT-Based Deep Learning

摘要: 目的：探讨临床特征、术前增强CT深度影像特征及其联合模型在膀胱癌根治性膀胱切除术后总体生存期(OS)预测中的应用价值。方法：回顾性纳入2017年7月至2023年2月于青岛大学附属医院接受腹腔镜根治性膀胱切除术且术后病理证实为膀胱癌的304例患者，收集其临床资料及术前增强CT影像。采用单因素及多因素Cox回归筛选临床预后因素并构建临床模型；对CT图像经标准化、超分辨率重建及多层ROI整合后，利用预训练ResNet-50提取深度影像特征，并结合LASSO-Cox构建影像模型；进一步融合临床与影像风险信息构建联合模型。采用一致性指数(C-index)和Kaplan-Meier曲线评价模型性能。结果：全队列中位随访时间为50个月。临床模型训练集和测试集C-index分别为0.789和0.716，CT深度影像模型分别为0.756和0.657，联合模型分别为0.828和0.732。结论：临床特征与术前CT深度影像特征均具有膀胱癌术后OS预测价值，二者融合后可进一步提升模型区分能力，为患者风险分层和个体化随访管理提供参考。

Abstract: Objective: To investigate the application value of clinical characteristics, deep learning features derived from preoperative contrast-enhanced CT images, and their combined model in predicting overall survival (OS) following radical cystectomy for bladder cancer. Methods: This retrospective study included 304 patients who underwent laparoscopic radical cystectomy at Qingdao University Affiliated Hospital between July 2017 and February 2023 and were pathologically confirmed to have bladder cancer. Clinical data and preoperative contrast-enhanced CT images were collected. Univariate and multivariate Cox regression analyses were performed to identify clinical prognostic factors and construct a clinical model. After standardization, super-resolution reconstruction, and multi-layer region-of-interest (ROI) integration of the CT images, deep image features were extracted using a pre-trained ResNet-50 model, and an image model was constructed using LASSO-Cox regression. A combined model was further developed by integrating clinical and imaging risk information. Model performance was evaluated using the C-index and Kaplan-Meier curves. Results: The median follow-up duration for the entire cohort was 50 months. The C-index for the clinical model was 0.789 in the training set and 0.716 in the validation set; for the CT deep image model, it was 0.756 and 0.657, respectively; and for the combined model, it was 0.828 and 0.732, respectively. Conclusion: Both clinical features and preoperative CT deep learning features possess predictive value for postoperative OS in bladder cancer. The integration of these two approaches further enhances the model’s discriminatory ability, providing a reference for patient risk stratification and personalized follow-up management.