基于磁共振成像检查的集成深度学习模型预测中低位直肠癌切除术中直线切割闭合器使用次数的临床价值

Clinical value of magnetic resonance imaging based integrated deep learning model for predic-ting the times of linear staplers used in middle-low rectal cancer resection

目的探讨基于磁共振成像(MRI)检查的集成深度学习模型预测中低位直肠癌切除术应用直肠双吻合技术(DST)直线切割闭合器使用次数的临床价值。方法采用回顾性队列研究方法。收集2018年1月至2022年12月上海交通大学医学院附属瑞金医院收治的263例行低位前切除术(LAR)中低位直肠癌患者的临床病理资料(训练集);男183例,女80例;年龄为63(55,68)岁。另收集128例中低位直肠癌患者的临床病理资料(验证集);男83例,女45例;年龄为65(57,70)岁。训练集用于构建预测模型,验证集用于验证预测模型。观察指标:(1)训练集患者的临床病理特征。(2)影响训练集患者术中直线切割闭合器使用次数≥3次的因素分析。(3)预测模型的构建。(4)预测模型的效能评价。(5)预测模型的验证。偏态分布的计量资料以M(Q_(1),Q_(3))表示,组间比较采用Mann-Whitney U检验。计数资料以绝对数表示,组间比较采用χ2检验。单因素分析采用Logistic回归模型,多因素分析采用Logistic逐步回归模型。绘制受试者工作特征(ROC)曲线并计算曲线下面积(AUC)。AUC>0.75表示模型可接受。AUC比较采用Delong检验。结果(1)训练集患者的临床病理特征。263例患者中,48例术中直线切割闭合器使用次数≥3次,215例≤2次。48例术中直线切割闭合器使用次数≥3次和215例≤2次患者术前血清癌胚抗原>5μg/L、吻合口漏、肿瘤长径≥5 cm的例数分别为20、12、13例和56、26、21例,两者比较,差异均有统计学意义(χ^(2)=4.66,5.29,10.45,P<0.05)。(2)影响训练集患者术中直线切割闭合器使用次数≥3次的因素分析。多因素分析结果显示:术前血清癌胚抗原>5μg/L、肿瘤长径≥5 cm是影响术中直线切割闭合器使用次数≥3次的独立危险因素[优势比=2.26,3.39,95%可信区间(CI)为1.15~4.43,1.50~7.65,P<0.05]。(3)预测模型的构建。根据多因素分析结果,纳入术前血清癌胚抗原和肿瘤长径建立临床预测模型Logit(P)=-2.018+0.814×术前血清癌胚抗原(>5μg/L取1,≤5μg/L取0)+1.222×肿瘤长径(≥5 cm取1,<5 cm取0)。将基于掩模区域的卷积神经网络(Mask R-CNN)分割的影像资料输入三维卷积网络(C3D),通过训练,完成影像预测模型的构建。将基于Mask R-CNN分割的影像资料及临床独立危险因素输入C3D神经网络,通过训练,完成综合预测模型的构建。(4)预测模型的效能评价。临床预测模型的灵敏度、特异度、准确度分别为70.0%、81.0%、79.4%,约登指数为0.51。影像预测模型的灵敏度、特异度、准确度分别为50.0%、98.3%、91.2%,约登指数为0.48。综合预测模型的灵敏度、特异度、准确度分别为70.0%、98.3%、94.1%,约登指数为0.68。临床预测模型、影像预测模型、综合预测模型的AUC分别为0.72(95%CI为0.61~0.83)、0.81(95%CI为0.71~0.91)、0.88(95%CI为0.81~0.95)。综合预测模型的效能分别与影像预测模型和临床预测模型比较,差异均有统计学意义(Z=2.98,2.48,P<0.05)。(5)预测模型的验证。通过验证集对3个预测模型进行外部验证。临床预测模型的灵敏度、特异度、准确度分别为62.5%、66.1%、65.6%,约登指数为0.29。影像预测模型的灵敏度、特异度、准确度分别为58.8%、95.5%、92.1%,约登指数为0.64。综合预测模型的灵敏度、特异度、准确度分别为68.8%、97.3%、93.8%,约登指数为0.66。临床预测模型、影像预测模型、综合预测模型的AUC分别为0.65(95%CI为0.55~0.75)、0.75(95%CI为0.66~0.84)、0.84(95%CI为0.74~0.93)。综合预测模型的效能与临床预测模型比较,差异有统计学意义(Z=3.24,P<0.05)。结论基于MRI检查的集成深度学习模型可以预测中低位直肠癌切除术DST直线切割闭合器使用次数≥3次的高危人群。

Objective To investigate the clinical value of magnetic resonance imaging(MRI)based integrated deep learning model for predicting the times of linear staplers used in double stapling technique for middle-low rectal cancer resection.Methods The retrospective cohort study was conducted.The clinicopathological data of 263 patients who underwent low anterior resection(LAR)for middle-low rectal cancer in Ruijin Hospital of Shanghai Jiaotong University School of Medicine from January 2018 to December 2022 were collected as training dataset.There were 183 males and 80 females,aged 63(55,68)years.The clinicopathological data of 128 patients with middle-low rectal cancer were collected as validation dataset,including 83 males and 45 females,with age as 65(57,70)years.The training dataset was used to construct the prediction model,and the validation dataset was used to validate the prediction model.Observation indicators:(1)clinicopathological features of patients in the training dataset;(2)influencing factors for≥3 times using of linear staplers in the operation;(3)prediction model construction;(4)efficiency evaluation of prediction model;(5)validation of prediction model.Measurement data with skewed distribution were represented as M(Q_(1),Q_(3)),and Mann-Whitney U test was used for comparison between groups.Count data were expressed as absolute numbers,and comparison between groups was conducted using the chi-square test.Wilcoxon rank sum test was used for non-parametric data analysis.Univariate analysis was conducted using the Logistic regression model,and multivariate analysis was conducted using the Logistic stepwise regression model.The receiver operating characteristic(ROC)curve was draw and the area under the curve(AUC)was calculated.The AUC of the ROC curve>0.75 indicated the prediction model as acceptable.Comparison of AUC was conducted using the Delong test.Results(1)Clinicopathological features of patients in the training dataset.Of the 263 patients,there were 48 cases with linear staplers used in the operation≥3 times and 215 cases with linear staplers used in the operation≤2 times.Cases with preoperative serum carcinoembryonic antigen(CEA)>5μg/L,cases with anastomotic leakage,cases with tumor diameter≥5 cm were 20,12,13 in the 48 cases with linear staplers used≥3 times in the operation,versus 56,26,21 in the 215 cases with linear staplers used≤2 times in the operation,showing significant differences in the above indicators between them(χ^(2)=4.66,5.29,10.45,P<0.05).(2)Influencing factors for≥3 times using of linear staplers in the operation.Results of multivariate analysis showed that preoperative serum CEA>5μg/L and tumor diameter≥5 cm were independent risk factors for≥3 times using of linear staplers in the operation(odds ratio=2.26,3.39,95%confidence interval as 1.15-4.43,1.50-7.65,P<0.05).(3)Prediction model construction.According to the results of multivariate analysis,the clinical prediction model was established as Logit(P)=-2.018+0.814×preoperative serum CEA(>5μg/L as 1,≤5μg/L as 0)+1.222×tumor diameter(≥5 cm as 1,<5 cm as 0).The image data segmented by the Mask region convolutional neural network(MASK R-CNN)was input into the three-dimensional convolutional neural network(C3D),and the image prediction model was constructed by training.The image data segmented by the MASK R-CNN and the clinical independent risk factors were input into the C3D,and the integrated prediction model was constructed by training.(4)Efficiency evaluation of prediction model.The sensitivity,specificity and accuracy of the clinical prediction model was 70.0%,81.0%and 79.4%,respectively,with the Yoden index as 0.51.The sensitivity,specificity and accuracy of the image prediction model was 50.0%,98.3%and 91.2%,respectively,with the Yoden index as 0.48.The sensitivity,specificity and accuracy of the integrated prediction model was 70.0%,98.3%and 94.1%,respectively,with the Yoden index as 0.68.The AUC of clinical prediction model,image prediction model and integrated prediction model was 0.72(95%confidence interval as 0.61-0.83),0.81(95%confidence interval as 0.71-0.91)and 0.88(95%confidence interval as 0.81-0.95),respectively.There were significant differences in the efficacy between the integrated prediction model and the image prediction model or the clinical prediction model(Z=2.98,2.48,P<0.05).(5)Validation of prediction model.The three prediction models were externally validated by validation dataset.The sensitivity,specificity and accuracy of the clinical prediction model was 62.5%,66.1%and 65.6%,respectively,with the Yoden index as 0.29.The sensitivity,specificity and accuracy of the image prediction model was 58.8%,95.5%and 92.1%,respectively,with the Yoden index as 0.64.The sensitivity,specificity and accuracy of the integrated prediction model was 68.8%,97.3%and 93.8%,respectively,with the Yoden index as 0.66.The AUC of clinical prediction model,image prediction model and integrated prediction model was 0.65(95%confidence interval as 0.55-0.75),0.75(95%confidence interval as 0.66-0.84)and 0.84(95%confidence interval as 0.74-0.93),respec-tively.There was significant differences in the efficacy between the clinical prediction model and the integrated prediction model(Z=3.24,P<0.05).Conclusion The MRI-based deep-learning model can help predicting the high-risk population with≥3 times using of linear staplers in resection of middle-low rectal cancer with double stapling technique.