基于增强CT的深度学习模型预测胃肠道间质瘤Ki-67表达的双中心研究

A two-center study of a deep learning model based on enhanced CT to predict Ki-67 expression in gastrointestinal stromal tumors

目的探讨基于增强CT的深度学习模型术前无创预测胃肠道间质瘤(GIST)Ki-67表达的价值。方法回顾性收集2所医院经手术病理证实且行免疫病理学染色的262例GIST病人的临床及影像学资料,其中男140例、女122例,平均年龄(55.82±10.13)岁。将一所医院收集的190例病人使用随机分层抽样法以7∶3的比例分为训练组(133例)和内部验证组(57例);将另一所医院收集的病人作为外部验证组(72例)。应用Resnet34、Resnet50、Densenet、Efficientnet和Efficientnetv2等5种基础网络进行模型训练和优化。采用受试者操作特征曲线下面积(AUC)、准确度、特异度、敏感度评估模型的预测效能。使用DeLong检验比较各模型间AUC值差异,得到最佳网络模型。应用梯度类加权激活映射(Grad-CAM)可视化的方法在原始CT图像上生成注意力热图。结果当学习率(Lr)为0.0005时,5种模型在训练组、内部验证组、外部验证组中预测Ki-67表达的AUC值均高于Lr=0.0001时。Densenet(Lr=0.0005)模型在训练组、内部验证组、外部验证组中对Ki-67表达的预测效能、准确度均最佳,AUC分别为0.983、0.930、0.925,预测准确度分别为92.77%、88.14%、87.77%。Efficientnet模型的预测敏感度最佳,Efficientnetv2模型的预测特异度最佳,但两者准确度均低于其他模型。注意力热图显示模型可以从矩形兴趣区(ROI)中正确识别肿瘤区域,合理解释模型的决策逻辑。结论基于增强CT的深度学习模型具有良好的稳定性和诊断效能,是一种无创预测GIST Ki-67表达的潜在方法。

Objective To investigate the value of a deep learning model based on enhanced CT in preoperative noninvasive prediction of Ki-67 expression in gastrointestinal stromal tumor(GIST).Methods The clinical and imaging data of 262 patients(140 males and 122 females,mean age 55.82±10.13 years)with GIST confirmed by surgery and pathology were retrospectively collected.They underwent postoperative immunopathological staining in two hospitals.A total of 190 patients collected from one hospital were randomly divided into training group(133 cases)and internal validation group(57 cases)using a 7∶3 ratio,while 72 patients collected from another hospital were used as external validation group.Five basic networks(Resnet34,Resnet50,Densenet,Efficientnet,and Efficientnetv2)were employed to train and optimize the model.Area under the receiver operating characteristics curves(AUC),accuracy,specificity,and sensitivity were used to assess the predictive power of the model.The differences in AUC values between models were compared using the DeLong test to obtain the best network model.A gradient-class weighted activation mapping(Grad-CAM)visualization method was applied to generate heatmap of attention on original CT images.Results All models,using a learning rate(Lr)of 0.0005,outperformed Lr=0.0001,exhibiting better AUC values in the training,internal validation,and external validation groups.The Densenet(Lr=0.0005)model demonstrated superior predictive efficacy and accuracy for Ki-67 expression in the training,internal validation,and external validation groups,with AUCs of 0.983,0.930,and 0.925,and predictive accuracies of 92.77%,88.14%,and 87.77%,respectively.The Efficientnet model had the best predictive sensitivity and the Efficientnetv2 model had the highest predictive specificity but lower accurate than the other models.The attention heatmap showed that the model can accurately identify the tumor region within the rectangular region of interest(ROI)and provided a reasonable explanation of decision logic.Conclusions The deep learning model based on enhanced CT has good stability and diagnostic efficacy,and is a potential method for noninvasive prediction of Ki-67 expression in GIST.