基于增强CT的临床–深度学习模型预测非肌层浸润性膀胱癌复发

A Clinical-Deep Learning Model Based on Contrast-Enhanced CT for Predicting Recurrence of Non-Muscle-Invasive Bladder Cancer

摘要: 目的：探讨基于多期增强CT图像构建的临床–深度学习模型在预测非肌层浸润性膀胱癌患者术后复发风险中的价值。方法：回顾性纳入126例经手术病理证实的非肌层浸润性膀胱癌患者(青岛大学附属医院)，按照院区不同分为训练集(85例)和验证集(41例)。提取三期增强CT (皮质期、髓质期、排泄期)图像，构建基于Transformer的端到端深度学习模型。通过单因素及多因素Logistic回归分析筛选独立危险因素，构建临床模型；并将深度学习影像特征与临床独立危险因素相结合，构建临床–深度学习模型。利用曲线下的面积(AUC)评估不同模型的区分度，校准曲线评估拟合优度，决策曲线分析(DCA)评估临床净获益。结果：临床–深度学习模型的训练集AUC为0.997 (95%CI: 0.991~1.000)，验证集AUC为0.787 (95%CI: 0.607~0.968)，优于临床模型和深度学习模型，表明模型具有优异的稳定性和区分能力。校准曲线显示模型预测概率与实际发生概率具有良好的一致性。DCA曲线表明该模型在广泛的阈值概率范围内均能为临床决策带来显著的净获益。结论：基于多期增强CT的临床–深度学习模型能有效预测非肌层浸润性膀胱癌患者的术后复发风险，具有较高的临床应用价值，可辅助医生制定个性化的随访和治疗策略。

Abstract: Objective: To investigate the value of a clinical-deep learning model based on multi-phase contrast-enhanced CT images in predicting the postoperative recurrence risk of patients with non-muscle-invasive bladder cancer (NMIBC). Methods: A total of 126 patients with pathologically confirmed non-muscle-invasive bladder cancer from the Affiliated Hospital of Qingdao University were retrospectively enrolled. They were divided into a training cohort (n = 85) and a validation cohort (n = 41) based on the hospital campus. Three-phase contrast-enhanced CT images (corticomedullary, nephrographic, and excretory phases) were extracted to construct an end-to-end deep learning model based on Transformer. Univariate and multivariate logistic regression analyses were performed to screen independent risk factors for constructing the clinical model. Subsequently, the deep learning imaging features were combined with the independent clinical risk factors to establish the clinical-deep learning model. The area under the curve (AUC) was used to evaluate the discrimination of the models, calibration curves were used to assess the goodness of fit, and decision curve analysis (DCA) was applied to evaluate the clinical net benefit. Results: The clinical-deep learning model yielded an AUC of 0.997 (95% CI: 0.991~1.000) in the training cohort and 0.787 (95% CI: 0.607~0.968) in the validation cohort, outperforming both the clinical and deep learning models, which indicated excellent stability and discrimination capability. The calibration curves demonstrated good consistency between the predicted probabilities and the actual observed probabilities. The DCA curves indicated that this model could provide significant net clinical benefit for decision-making across a wide range of threshold probabilities. Conclusion: The clinical-deep learning model based on multi-phase contrast-enhanced CT can effectively predict the postoperative recurrence risk in patients with non-muscle-invasive bladder cancer. It possesses high clinical application value and can assist physicians in formulating personalized follow-up and treatment strategies.