基于增强CT栖息地的影像组学预测肝细胞癌微血管侵犯的研究  

Radiomics-Based Prediction of Microvascular Invasion in Hepatocellular Carcinoma Using Contrast-Enhanced Computed Tomography Habitat

作　　者：赵凯 李聚贤 乔雪 孙中琪 邹治炫 姜慧杰[1] ZHAO Kai;LI Juxian;QIAO Xue(Department of Radiology,The 2nd Affiliated Hospital of Harbin Medical University,Harbin,Heilongjiang Province 150001,P.R.China)

机构地区：[1]哈尔滨医科大学附属第二医院放射科,150001 [2]广州中医药大学顺德医院放射科,528300

出　　处：《临床放射学杂志》2025年第4期664-669,共6页Journal of Clinical Radiology

摘　　要：目的 探讨基于增强CT(CECT)栖息地的影像组学技术预测肝细胞癌(HCC)患者微血管侵犯(MVI)的价值,并通过Shapley算法对模型进行可视化呈现。方法 回顾性搜集130例经病理证实HCC患者的CECT门静脉期图像,随机分为MVI组(n=91)和非微血管侵犯(non-MVI)组(n=39)。采用K均值聚类算法对肿瘤区域进行分区,生成3~10个不同聚类亚区,使用卡林斯基-哈拉巴斯分数(CH score)评估不同聚类数的聚类水平,选择最优聚类数。基于每个亚区提取影像组学特征,采用Spearman相关性分析和最小绝对收缩和选择算子(LASSO)算法进行特征筛选,并使用随机森林算法构建预测MVI的不同栖息地模型。使用受试者工作特征曲线(ROC)的曲线下面积(AUC)评估模型性能,用校准曲线验证模型的校准能力,用决策曲线分析和比较模型的临床实用,并通过Shapley算法可视化呈现特征的重要性。结果 根据聚类分析,最终选择了3个栖息地亚区,每个亚区提取1834个特征。所有模型中,Combined Habitat模型性能最优,训练集和测试集中的AUC分别为0.836(95%CI:0.749~0.918)和0.791(95%CI:0.641~0.923),高于不同亚区单独构建的模型。决策曲线提示Combined Habitat模型较其他模型具有较好的临床实用价值,校准曲线显示模型拟合良好。Shapley算法展示了Combined Habitat中不同亚区的影像组学特征重要性,其中来自Habitat3的“exponential_firstorder_Maximum_h3”特征对模型预测具有最强的正向影响。结论 基于CECT栖息地构建的影像组学模型对预测HCC MVI具有重要价值。通过对肿瘤不同异质性亚区进行精细划分和特征提取,能够更好地捕捉肿瘤异质性。Objective To explore the value of radiomics technology based on contrast-enhanced computed tomography(CECT) habitat for predicting microvascular invasion(MVI) in hepatocellular carcinoma(HCC),and to visualize the model using Shapley Additive exPlanations algorithm.Methods A retrospective analysis of 130 HCC patients with pathologically confirmed MVI was performed,with the cohort divided randomly into the MVI group(n=91) and non-MVI group(n=39).K-means clustering algorithm was used to segment tumor regions,generating 3-10 different clustering subregions.The clustering level was evaluated using the Calinski-Harabasz score(CH score),and the optimal number of clusters was selected.Radiomics features were extracted from each subregion,followed by Spearman correlation analysis and feature selection using least absolute shrinkage and selection operator(LASSO).Random forest algorithm was used to construct different habitat models for MVI prediction.The model performance was assessed using the area under the receiver operating characteristic curve(AUC),using calibration curves for model calibration ability,and using decision curves for clinical applicability.Feature importance was visualized through the SHAP algorithm.Results Based on the clustering analysis,3 habitat subregions were selected,with a total of 1834 features extracted from each subregion.Among all models,the combined habitat model exhibited the best performance,with AUC values of 0.836(0.749-0.918) and 0.791(0.641-0.923) for the training and testing sets,respectively,outperforming models constructed with individual subregions.Decision curve analysis indicated that the combined habitat model had superior clinical utility compared to the other models,and the calibration curve demonstrated good model fit.The SHAP algorithm visualized the importance of radiomics features from different subregions in the combined habitat model,with the feature “exponential_firstorder_Maximum_h3” from Habitat 3 having the strongest positive impact on model prediction.Conclusio

关 键 词：肝细胞癌 微血管侵犯 影像组学 机器学习 体层摄影术 X线计算机 

分 类 号：R73[医药卫生—肿瘤]