摘要
目的基于动态对比增强MRI(dynamic contrast-enhanced MRI,DCE-MRI)生境影像组学和临床特征构建可解释的联合模型,探讨模型预测直肠癌淋巴结转移(lymph node metastasis,LNM)状态的应用价值。材料与方法回顾性分析2016年1月至2024年7月甘肃省人民医院收治的148例直肠癌患者的临床-病理-影像学资料,根据术后病理确诊的LNM状态将患者分为有LNM组(61例)和无LNM组(87例)。按7∶3比例随机分为训练集103例和测试集45例。用ITK-SNAP软件勾画感兴趣区,基于Ktrans图提取19项标准化影像组学特征,通过K-means聚类划分肿瘤生境亚区(K=4),分别提取各个肿瘤亚区的影像组学特征和全肿瘤影像组学特征,合并各亚区特征形成一个综合的生境影像组学特征集。采用组内相关系数(intra-class correlation coefficient,ICC)评估肿瘤整体影像组学特征的可重复性。使用Z-score标准化、相关性分析和最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)进行特征筛选,利用极端随机树、逻辑回归、随机森林及支持向量机构建生境及全肿瘤影像组学预测模型。采用logistic回归筛选临床独立预测因子,建立临床模型。最后,联合临床特征(癌胚抗原和MRI检查报告的N分期)和生境影像组学评分构建联合模型。绘制受试者工作特征(receiver operating characteristic,ROC)曲线,以曲线下面积(area under the curve,AUC)评估模型效能,决策曲线评估临床实用性。通过沙普利可加性模型解释算法(Shapley additive explanations,SHAP)分析描述特征重要程度并可视化解释联合模型预测结果。结果单因素分析发现,癌胚抗原和MRI检查报告的N分期是影响训练集直肠癌患者淋巴结状态的影响因素,比值比分别为2.346[95%置信区间(confidence interval,CI):1.052~5.233]和7.727(95%CI:2.273~26.268),P<0.05。基于生境影像组学特征的预测模型(训练集:0.890,测试集:0.801)预测直肠癌LNM状态的AUC优于全肿瘤影像组学预测模型(训练集:0.774,测试集:0.684),联合临床-生境影像组学联合模型的AUC值最高(训练集:0.896,测试集:0.866)。决策曲线显示联合模型显示出更高的临床净收益。联合预测模型SHAP算法可以为模型提供了定量的解释,生境影像组学评分为模型最重要的特征。结论基于术前DCE-MRI生境影像组学特征和临床因素构建的可解释联合模型能够准确预测直肠癌患者的淋巴结状态,并通过SHAP可视化预测过程,为个性化治疗提供科学依据。
Objective:To construct an interpretable integrated model based on dynamic contrast-enhanced MRI(DCE-MRI)habitat imaging radiomics and clinical features,and to assess its utility in predicting lymph node metastasis(LNM)status in rectal cancer.Materials and Methods:A retrospective analysis was conducted on the clinicopathological and imaging data of 148 patients with rectal cancer admitted to Gansu Provincial People's Hospital between January 2016 and July 2024.Patients were stratified into LNM-positive and LNM-negative groups based on postoperative pathological confirmation.They were then randomly divided into a training cohort(n=103)and a test cohort(n=45)in a 7∶3 ratio.The region of interest(ROI)was manually delineated on the DCE-MRI parametric map using ITK-SNAP software.Subsequently,19 standardized radiomics features were extracted from the Ktrans maps.K-means clustering(K=4)was applied to partition the tumor into distinct habitat subregions.Radiomics features were extracted separately from each tumor subregion(habitat-specific features)and from the whole tumor volume(whole-tumor features).The intra-class correlation coefficient(ICC)was calculated to assess the reproducibility of the whole-tumor radiomics feature extraction.Feature selection involved Z-score normalization,correlation analysis,and the least absolute shrinkage and selection operator(LASSO)algorithm.Predictive models for LNM status were developed using four machine learning classifiers:extremely randomized trees,logistic regression,random forest,and support vector machine.These models were built based on habitat-specific radiomics features and whole-tumor radiomics features separately.Logistic regression was also used to identify independent clinical predictors and construct a clinical model.Finally,an integrated model was built by combining significant clinical predictors with the radiomics signature derived from the habitat analysis.Model performance was evaluated using the receiver operating characteristic(ROC)curve and quantified by the area under the curve(AUC).Decision curve analysis(DCA)was performed to assess the clinical utility of the models.The importance of features in the final integrated model was determined,and the model's predictions were explained visually using Shapley additive explanations(SHAP)analysis.Result:Univariate analysis identified carcinoembryonic antigen(CEA)level and MRI-reported N-stage as significant predictors of lymph node status in the training cohort[odds ratios(OR)=2.346 and 7.727,respectively;95% confidence intervals(CI):1.052 to 5.233 and 2.273 to 26.268,respectively;P<0.05].The predictive model based on habitat radiomics features demonstrated superior performance,with AUC values of 0.890(training cohort)and 0.801(test cohort),outperforming the whole-tumor radiomics model(AUC:0.774 training,0.684 test).The integrated model,combining clinical features with the habitat radiomics signature,achieved the highest AUC values:0.896 in the training cohort and 0.866 in the test cohort.DCA indicated that the integrated model provided a higher net clinical benefit across a range of threshold probabilities.SHAP analysis provided quantitative interpretability for the integrated model's predictions,revealing the habitat radiomics score as the most significant predictor.Conclusions:The interpretable integrated model,constructed using preoperative DCE-MRI habitat imaging radiomics features and clinical factors,accurately predicts lymph node status in rectal cancer patients.By providing visual interpretation of individual predictions through SHAP,this model offers a valuable tool to support personalized treatment decision-making.
作者
孙赟
李飞翔
陈学敏
张颖颖
黄刚
SUN Yun;LI Feixiang;CHEN Xuemin;ZHANG Yingying;HUANG Gang(The First Clinical Medical College of Gansu University of Chinese Medicine,Lanzhou,730000,China;Department of Radiology,Gansu Provincial Hospital,Lanzhou,730000,China)
出处
《磁共振成像》
北大核心
2026年第1期69-78,共10页
Chinese Journal of Magnetic Resonance Imaging
基金
甘肃省自然科学基金项目(编号:24JRRA1054)。
关键词
直肠癌
淋巴结转移
磁共振成像
影像组学
生境分析
rectal cancer
lymph node metastasis
magnetic resonance imaging
radiomics
habitat analysis
