钆塞酸二钠增强MRI影像组学对肝血管瘤腹腔镜肝切除术后疗效的预测价值

目的 探讨钆塞酸二钠（Gd-EOB-DTPA）增强MRI影像组学对肝血管瘤腹腔镜肝切除术疗效的预测价值。方法 选取2018年2月至2020年10月行腹腔镜肝切除术治疗的肝血管瘤患者131例作为研究对象。采用计算机产生随机数法将131例肝血管瘤患者以7:3的比例分为训练集（92例）和验证集（39例）,分别用于构建和验证预测模型效能。根据术后是否出现并发症将训练集患者分为预后不良组和预后良好组,比较两组患者的临床特征资料。使用最小绝对收缩和选择算子（the least absolute shrinkage and selection operator,LASSO）回归5折交叉验证方法筛选训练集临床和影像组学特征,得到最优特征子集。使用机器学习算法构建预测模型,并使用受试者工作特征（receiver operating characteristics,ROC）曲线评价各模型对肝血管瘤术后疗效的预测价值。结果 训练集中预后良好64例,预后不良28例,两组在慢性疾病、肝血管瘤最大径、血管瘤数、血管瘤供血动脉支数、肝切除大小、术中出血量方面差异均有统计学意义（P<0.05）。依据LASSO回归中的最佳λ取值筛选出14个非零系数特征构成的最优特征子集,其中包括12个组学特征和2个临床特征。利用训练集筛选出的最优特征子集构建Logistic回归、CatBoost、XGBoost和LightGBM模型的AUC分别为0.820、0.722、0.784、0.693。验证集构建的Logistic回归、CatBoost、XGBoost和LightGBM模型的ROC曲线下面积（AUC）分别为0.814、0.706、0.785、0.684。4种机器学习模型的预测效果良好,其中Logistic回归的预测效果优于其他3种。结论 血管瘤最大径和血管瘤数目,结合Gd-EOB-DTPA增强MRI影像组学,以此构建的机器学习模型可用于预测肝血管瘤腹腔镜肝切除术后疗效,其中Logistic回归算法构建的模型更加精准。

Value of Gd-EOB-DTPA enhanced MRI in predicting the postoperative efficacy of laparoscopic hepatectomy for hepatic hemangioma

Objective To investigate the value of gadolinium ethoxybenzyl diethylentriamine pentaacetic acid (Gd-EOB-DTPA) enhanced MRI imaging in predicting the postoperative efficacy of laparoscopic hepatectomy for hepatic hemangioma. Methods A total of 131 patients with hepatic hemangioma who underwent laparoscopic hepatectomy between Feb. 2018 and Oct. 2020 were selected as subjects. Patients were randomly divided into training set (n=92) and verification set (n=39) in a ratio of 7:3, to construct and verify the prediction model respectively. According to postoperative complications, patients in training set were divided into poor prognosis group and good prognosis group, and the clinical characteristics between the two groups were compared. The clinical and imaging features in training set were screened by LASSO regression 50% discount cross-validation method, and the optimal feature subset was obtained. The predictive model was constructed by machine learning algorithm, and the ROC curve was used to evaluate the prediction value of each model for postoperative efficacy of laparoscopic hepatectomy. Results In training set, the postoperative prognosis of laparoscopic hepatectomy for hepatic hemangioma was good in 64 cases and poor in 28 cases. There were significant differences in chronic diseases, the maximum diameter of hemangioma, the number of hemangioma, the number of feeding arteries of hemangioma, the size of hepatectomy and the amount of intraoperative blood loss between the two groups (P<0.05). According to the optimal λ value in LASSO regression, the optimal feature subset composed of 14 non-zero coefficient features was screened, which included 12 taxonomic features and 2 clinical features. The AUC of Logistic regression, CatBoost, XGBoost and LightGBM models constructed by the optimal feature subset which selected from the training set were 0.820, 0.722, 0.784 and 0.693, respectively. And the AUC of Logistic regression, CatBoost, XGBoost and LightGBM models from the verification set were 0.814, 0.706, 0.785 and 0.684, respectively. The prediction effect of the above 4 machine algorithms were good, and the prediction effect of Logistic regression is better than the other three. Conclusion The machine learning model constructed by the maximum diameter and the number of hemangiomas, combined with the Gd-EOB-DTPA enhanced MRI imaging features, can be used to predict the postoperative efficacy of laparoscopic hepatectomy for hepatic hemangioma, and the model constructed by Logistic regression algorithm is more accurate.