CT影像组学模型预测膀胱癌术后1年复发的初步研究

  目的  探究基于CT图像的影像组学模型预测膀胱癌术后1年复发的可行性。  方法  回顾性纳入2014年5月至2018年7月于北京协和医院行手术治疗的膀胱癌患者，并对其进行随访，记录疾病复发状况。收集膀胱癌患者术前CT泌尿系成像实质期图像，经滤波处理后进行影像组学特征提取；采用JMIM特征选择算法识别与膀胱癌术后1年复发相关的最佳影像组学特征，采用随机森林模型、自适应增强模型、梯度提升树模型以及3个模型构成的组合模型构建膀胱癌术后1年复发的预测模型，并基于10次10折交叉验证法对各模型进行验证。采用受试者工作特征曲线对各模型的预测性能进行评定。  结果  共228例符合纳入和排除标准的膀胱癌患者入选本研究。随访1年时51例患者复发，177例患者未复发。经交叉验证，随机森林模型、自适应增强模型、梯度提升树模型和组合模型预测膀胱癌术后1年复发的曲线下面积分别为0.729(95% CI: 0.649~0.809)、0.710(95% CI: 0.627~0.793)、0.709(95% CI: 0.624~0.793)、0.732(95% CI: 0.651~0.812)，准确度分别为76.8%(95% CI: 70.6%~82.0%)、73.7%(95% CI: 67.4%~79.2%)、61.8%(95% CI: 54.7%~67.7%)、75.0%(95% CI: 68.8%~80.4%)，灵敏度分别为52.9%(95% CI: 38.6%~66.8%)、62.7%(95% CI: 48.1%~75.5%)、80.4%(95% CI: 64.3%~88.2%)、58.8%(95% CI: 44.2%~72.1%)，特异度分别为83.6%(95% CI: 77.1%~88.6%)、76.8%(95% CI: 69.8%~82.7%)、56.5%(95% CI: 48.9%~63.9%)、79.7%(95% CI: 72.8%~85.2%)。  结论  有机结合基于CT图像构建的多个影像组学模型可预测膀胱癌术后1年的复发风险。

CT-based Radiomics to Predict Recurrence of Bladder Cancer after Resection in One Year: A Preliminary Study

  Objective  To investigate the feasibility of the CT-based radiomics model to predict the recurrence of bladder cancer in one year postoperatively.  Methods  Patients with bladder cancer that received surgical treatment in Peking Union Medical College Hospital from May 2014 to July 2018 were retrospectively enrolled and followed up the recurrence of the disease. Nephrographic phase images of preoperative CT urography(CTU) performed in our hospital were collected. The images were filtered before radiomic feature extraction, and JMIM was used to identify the best radiomic features related to recurrence of bladder cancer. Random forest, AdaBoost, gradient boosting, and their combined model were used to build the model for predicting recurrence of bladder cancer after resection in one year. We applied 10-fold cross validation to validate each model and performed receiver operator characteristic curves to analyze the performance of each model.  Results  A total of 228 cases were included in this study according to inclusion and exclusion criteria. Fifty-one patients had recurrence and the rest 177 patients had no recurrence in one year during postoperative follow-up. In the cross validation, the random forest model, AdaBoost model, gradient boosting model and combined model predicted the recurrence of bladder cancer with AUC of 0.729(95% CI: 0.649-0.809), 0.710(95% CI: 0.627-0.793), 0.709(95% CI: 0.624-0.793)and 0.732(95% CI: 0.651-0.812), accuracy of 76.8%(95% CI: 70.6%-82.0%), 73.7%(95% CI: 67.4%-79.2%), 61.8%(95% CI: 54.7%-67.7%)and 75.0%(95% CI: 68.8%-80.4%), sensitivity of 52.9%(95% CI: 38.6%-66.8%), 62.7%(95% CI: 48.1%-75.5%), 80.4%(95% CI: 64.3%-88.2%)and 58.8%(95% CI: 44.2%-72.1%), specificity of 83.6%(95% CI: 77.1%-88.6%), 76.8%(95% CI: 69.8%-82.7%), 56.5%(95% CI: 48.9%-63.9%)and 79.7%(95% CI: 72.8%-85.2%), respectively.  Conclusion  Integration of CT-based radiomics prediction models can predict the recurrence risk of bladder cancer in one year postoperatively.