Gd-EOB-DTPA增强MRI影像组学联合临床特征模型对恶性肝硬化结节的预测价值

Value of joint model based on Gd-EOB-DTPA-enhanced MRI radiomics and clinical characteristics to the prediction of malignant cirrhotic nodules

目的 分析肝硬化结节患者的临床和Gd-EOB-DTPA增强MRI扫描影像组学特征,构建恶性肝硬化结节的联合预测模型,探讨其预测价值。方法 2021年4月—2023年4月新乡市中心医院行Gd-EOB-DTPA增强MRI扫描的肝硬化结节患者140例,采用随机数法以7∶3比例分为训练集98例和测试集42例。140例均行手术或组织穿刺活检病理检查明确结节良恶性质,测试集肝硬化结节恶性15例、良性27例,训练集结节恶性33例为恶性结节组、结节良性65例为良性结节组。比较训练集与测试集、恶性结节组与良性结节组年龄≥60岁、男性、吸烟史、饮酒史、肝癌家族史、脂肪肝、糖尿病比率,肝脏硬度值,体质量指数,组织病理检查前1周检测血小板计数、血白蛋白、凝血酶原时间及血清谷草转氨酶、总胆红素、谷丙转氨酶、甲胎蛋白水平和乙肝病毒脱氧核糖核酸(HBV-DNA)阳性情况;多因素logistic回归分析肝硬化结节为恶性的临床影响因素,并构建临床预测模型。应用Pyradiomics软件对训练集患者肝胆期MRI图像定量特征进行提取,提取每例患者的1 423个影像组学特征,先通过可重复性分析剔除其中823个特征,再对剩余的600个影像特征采用10倍交叉验证lasso回归算法筛选特征变量;多因素logistic回归分析构建肝硬化结节为恶性的影像组学预测模型,计算影像组学评分。将MRI影像组学评分和临床特征纳入多因素logistic回归分析构建联合预测模型。绘制ROC曲线评估临床预测模型、影像组学预测模型、联合预测模型在测试集中预测肝硬化结节为恶性的效能;采用校准曲线、临床决策曲线评价3种模型在测试集中预测肝硬化结节为恶性的价值。结果 (1)训练集与测试集年龄≥60岁、男性、吸烟史、饮酒史、肝癌家族史、脂肪肝、糖尿病比率,HBV-DNA阳性率,肝脏硬度值,体质量指数,血小板计数,血白蛋白,凝血酶原时间及血清谷草转氨酶、总胆红素、谷丙转氨酶、甲胎蛋白水平比较差异均无统计学意义(P>0.05)。(2)恶性结节组年龄≥60岁、男性、饮酒史、肝癌家族史比率,HBV-DNA阳性率,肝脏硬度值及血白蛋白、甲胎蛋白水平均高于良性结节组(P<0.05),吸烟史、脂肪肝、糖尿病比率及体质量指数、总胆红素、谷丙转氨酶、谷草转氨酶、血小板计数、凝血酶原时间与良性结节组比较差异均无统计学意义(P>0.05)。年龄(OR=2.993,95%CI:1.681~3.341,P<0.001)、性别(OR=2.223,95%CI:1.569~3.867,P=0.002)、饮酒史(OR=1.298,95%CI:1.005~1.977,P=0.016)、肝癌家族史(OR=1.236,95%CI:1.005~2.112,P=0.021)、HBV-DNA(OR=3.032,95%CI:1.005~4.968,P<0.001)是肝硬化结节为恶性的影响因素。临床预测模型=1.256+1.096×年龄+0.799×性别+0.261×饮酒史+0.212×肝癌家族史+1.109×HBV-DNA。(3)lasso回归筛选结果显示,最优λ为0.056时,orig_shape_MAL、wave_LHH_GLSZM_GLNU、wave_HHH_GLDM_DE、wave_HLL_FO_Minimum、wave_HHH_FO_TE为最具泛化能力的5个影像组学特征变量。orig_shape_MAL(OR=4.101,95%CI:2.321~6.297,P=0.023)、wave_LHH_GLSZM_GLNU(OR=3.568,95%CI:1.863~4.448,P=0.001)、wave_HHH_GLDM_DE(OR=2.512,95%CI:1.278~4.006,P=0.014)、wave_HLL_FO_Minimum(OR=2.115,95%CI:1.119~3.238,P=0.018)、wave_HHH_FO_TE(OR=3.205,95%CI:2.009~4.317,P=0.025)是肝硬化结节为恶性的影响因素。影像组学评分=1.360+1.411×orig_shape_MAL+1.272×wave_LHH_GLSZM_GLNU+0.921×wave_HHH_GLDM_DE+0.749×wave_HLL_FO_Minimum+1.165×wave_HHH_FO_TE。(4)联合预测模型=1.943+1.016×年龄+0.993×性别+0.310×饮酒史+0.257×肝癌家族史+1.244×HBV-DNA+1.280×影像组学评分。(5)临床预测模型、影像组学预测模型、联合预测模型最佳截断值分别为5.689、4.339、4.727时,预测测试集患者肝硬化结节为恶性的AUC分别为0.711(95%CI:0.532~0.866,P<0.001)、0.836(95%CI:0.665~0.956,P<0.001)、0.937(95%CI:0.851~0.993,P<0.001)。联合预测模型预测测试集患者肝硬化结节为恶性的AUC大于影像组学预测模型、临床预测模型(Z=2.374,P<0.001;Z=2.124,P<0.001),影像组学预测模型大于临床预测模型(Z=1.964,P<0.001)。校准曲线显示,测试集中临床预测模型、影像组学预测模型、联合预测模型的校准曲线与理想曲线一致性良好。决策曲线显示,阈值概率>20%时,联合预测模型的净收益高于影像组学预测模型、临床预测模型,影像组学预测模型高于临床预测模型。结论 Gd-EOB-DTPA增强MRI影像组学评分联合临床特征构建的预测模型对恶性肝硬化结节的预测价值较高。

Objective To analyze the clinical characteristics and radiomics features of Gd-EOB-DTPA-enhanced MRI scanning in patients with cirrhotic nodules,and to construct a joint predictive model for the malignant cirrhotic nodules in order to investigate its predictive value.Methods From April 2021to April 2023,140patients with cirrhotic nodules underwent Gd-EOB-DTPA-enhanced MRI scanning in Xinxiang Central Hospital,and were randomly divided into training set(n=98)and testing set(n=42)in a 7:3ratio.All patients underwent surgical tissue or biopsy tissue pathological examinations to determine the benign and malignant natures of the nodules.There were 15patients with malignant nodules and 27patients with benign nodules in the testing set.The training set was divided into malignant nodules group(n=33)and benign nodules group(n=65).The percentages of patients with age≥60years,smoking habits,alcohol consumption history,family history of liver cancer,fatty liver and diabetes,male ratio,liver hardness value,body mass index,and laboratory indexes one week before histopathological examination including platelet count,serum albumin,prothrombin time,glutamic-pyruvic transaminase,glutamic-oxaloacetic transaminase,total bilirubin,alpha fetoprotein,and HBV-DNA positive rate were compared between the training set and the testing set,and between malignant nodules group and benign nodules group.Multivariate logistic regression analysis was done to assess the clinical influencing factors of malignant cirrhotic nodules,and to construct a clinical predictive model.Pyradiomics software was used to extract quantitative features from liver and gallbladder MRI images in the training set,and 1423radiomics features were extracted for each patient.Totally 823features were removed through repeatability analysis,and the feature variables were screened in the remaining 600image features by a 10-fold cross validation lasso regression algorithm.Multivariate logistic regression analysis was used to construct a radiomics predictive model for the malignant cirrhotic nodules,and the radiomics score was calculated.MRI radiomics scores and clinical features were integrated into multivariate logistic regression analysis to construct a joint predictive model.ROC curves were plotted to evaluate the efficiencies of clinical predictive model,radiomics predictive model,and jointed predictive model on predicting malignant nodules in the testing set.The calibration curves and clinical decision curves were used to evaluate the values of three models to the prediction of malignant nodules in the testing set.Results(1)There were no significant differences in the percentages of patients with age≥60years,smoking habits,alcohol consumption history,family history of liver cancer,fatty liver and diabetes,male ratio,HBV-DNA positive rate,liver hardness value,body mass index,platelet count,serum albumin,prothrombin time,glutamic-pyruvic transaminase,glutamic-oxaloacetic transaminase,total bilirubin and alpha fetoprotein between the training set and the testing set(P>0.05).(2)The percentages of patients with age≥60years,alcohol consumption history and family history of liver cancer,male ratio,HBV-DNA positive rate,liver hardness value,serum albumin,and alpha fetoprotein levels were higher in malignant nodules group than those in benign nodules group(P<0.05),and there were no significant differences in the percentages of patients with smoking habits,fatty liver and diabetes,body mass index,total bilirubin,glutamic-pyruvic transaminase,glutamic-oxaloacetic transaminase,platelet count and prothrombin time between two groups(P>0.05).Age(OR=2.993,95%CI:1.681-3.341,P<0.001),gender(OR=2.223,95%CI:1.569-3.867,P=0.002),alcohol consumption history(OR=1.298,95%CI:1.005-1.977,P=0.016),family history of liver cancer(OR=1.236,95%CI:1.005-2.112,P=0.021),and HBV-DNA(OR=3.032,95%CI:1.005-4.968,P<0.001)were the influencing factors of malignant nodules.The clinical predictive model=1.256+1.096×Age+0.799×Gender+0.261×Alcohol consumption history+0.212×Family history of liver cancer+1.109×HBV-DNA.(3)The lasso regression screening results showed that when the optimalλwas 0.056,orig_shape_MALs,waves_LHH_GLSZM_GLNU,wave_HHH_GLDM_DE,wave_HLL_FO_Minimum,and wave_HHH_FO_TE were the five feature variables with the most generalization ability.orig_shape_MALs(OR=4.101,95%CI:2.321-6.297,P=0.023),waves_LHH_GLSZM_GLNU(OR=3.568,95%CI:1.863-4.448,P=0.001),wave_HHH_GLDM_DE(OR=2.512,95%CI:1.278-4.006,P=0.014),wave_HLL_FO_Minimum(OR=2.115,95%CI:1.119-3.238,P=0.018),and wave_HHH_FO_TE(OR=3.205,95%CI:2.009-4.317,P=0.025)were the influencing factors of malignant nodules.The radiomics score=1.360+1.411×orig_shape_MAL+1.272×wave_LHH_GLSZM_GLNU+0.921×wave_HHH_GLDM_DE+0.749×wave_HLL_FO_Minimum+1.165×wave_HHH_FO_TE.(4)The joint predictive model=1.943+1.016×Age+0.993×Gender+0.310×Alcohol consumption history+0.257×Family history of liver cancer+1.244×HBV-DNA+1.280×radiomics score.(5)When the optimal cut-offvalues of clinical predictive model,radiomics predictive model and joint predictive model were 5.689,4.339and 4.727,the AUCs for predicting the malignant nodules in the testing set were 0.711(95%CI:0.532-0.866,P<0.001),0.836(95%CI:0.665-0.956,P<0.001)and 0.937(95%CI:0.851-0.993,P<0.001),respectively.The AUCof joint predictive model in the testing set was greater than that of radiomics predictive model and clinical predictive model(Z=2.374,P<0.001;Z=2.124,P<0.001),and the AUCof radiomics predictive model was greater than that of clinical predictive model(Z=1.964,P<0.001).The calibration curve showed that the calibration curves of the clinical predictive model,radiomics predictive model,and joint predictive model in the testing set were consistent with the ideal curves.The decision curve showed that when the threshold probability was greater than 20%,the net profit of joint predictive model was higher than that of radiomics predictive model and clinical predictive model in the testing set,and the net profit of radiomics predictive model was higher than that of clinical predictive model.Conclusion The predictive model constructed by Gd-EOB-DTPA-enhanced MRI radiomics score combined with clinical characteristics has a high predictive value for malignant cirrhotic nodules.