建立一个基于MRI影像组学机器学习的列线图模型术前预测肝细胞癌（肝癌）微血管侵犯（MVI）风险，协助个性化治疗决策。

回顾性分析2019年6月至2022年12月在中国医科大学附属第一医院行手术切除且术后病理组织学证实为肝癌的176例患者。患者均签署知情同意书，符合医学伦理学规定。其中男148例，女28例；年龄32~82岁，中位年龄61岁。收集并整理176例肝癌患者的临床病理学资料，按7∶3随机（纯随机采样法）分为训练集（123例）和测试集（53例）。使用Python中的sklearn软件包在训练集中拟合K最近邻、随机森林、逻辑回归、支持向量机、朴素贝叶斯5种机器学习模型，在训练集和测试集中比较准确率、灵敏度、特异度、F1值和ROC AUC评估各机器学习模型的预测效能，选取综合表现最佳的模型生成影像组学评分。MVI阳性患者临床病理学特征分析采用χ²检验。采用Logistic多因素回归分析MVI的危险因素并建立列线图预测模型。应用ROC、决策曲线和校准曲线评估模型的预测能力和临床价值。

176例患者中MVI阳性54例（30.7%），其中M1为42例，M2为12例；MVI阴性122例（69.3%）。MVI阳性与年龄≤50岁、AFP升高、最大肿瘤直径>5 cm、肿瘤多发、动脉期瘤周强化和肿瘤内坏死有关（χ²=0.049，0.047，0.002，0.049，0.031，0.016；P<0.05）。最大肿瘤直径>5 cm、动脉期瘤周强化和影像组学评分是肝癌发生MVI的独立危险因素（OR=3.733，3.130，2.007；P<0.05）。列线图模型的ROC AUC为0.856（训练集）和0.772（测试集）。决策曲线和校准曲线证明了模型具有良好的临床适用性。

基于MRI影像组学机器学习构建的列线图模型对于术前预测肝癌MVI、指导诊疗决策具有良好的临床价值。

To construct a nomogram model based on machine learning of MRI radiomics to predict the risk of microvascular invasion (MVI) of hepatocellular carcinoma (HCC) before surgery, and to assist the selection of individualized treatment.

Clinical data of 176 patients pathologically diagnosed with HCC who underwent surgical resection in the First Affiliated Hospital of China Medical University from June 2019 to December 2022 were retrospectively analyzed. The informed consents of all patients were obtained and the local ethical committee approval was received. Among them, 148 patients were male and 28 female, aged from 32 to 82 years, with a median age of 61 years. Clinicopathological data of 176 HCC patients were collected and analyzed. All patients were randomly divided into the training set (n=123) and test set (n=53) according to a ratio of 7∶3 (simple random sampling method). The sklearn software package in Python was used to fit five machine learning models: K-nearest neighbor, random forest, logistic regression, support vector machine and naive Bayes in the training set. The accuracy, sensitivity, specificity, F1 score and the area under the ROC curve (AUC) were compared between the training and test sets to evaluate the prediction efficiency of each machine learning model, and the model with the optimal comprehensive performance was selected to generate the radiomics score. Clinicopathological features of MVI-positive patients were analyzed by Chi-square test. Multivariate Logistic regression analysis was employed to analyze the risk factors of MVI and establish a nomogram prediction model. ROC, decision curve and calibration curve were utilized to evaluate the predictive ability and clinical value of these models.

Among 176 patients, 54 cases (30.7%) were positive for MVI, including 42 cases of grade M1 and 12 of grade M2. 122 patients (69.3%) were negative for MVI. Positive MVI was associated with age ≤50 years, increased AFP, the maximum tumor diameter >5 cm, multiple tumors, peritumoral enhancement in arterial phase and intra-tumoral necrosis (χ²=0.049, 0.047, 0.002, 0.049, 0.031, 0.016; all P<0.05). The maximum tumor diameter >5 cm, peritumoral enhancement in arterial phase and radiomics score were the independent risk factors for MVI in HCC (OR=3.733,3.130, 2.007; all P<0.05). The AUC of each nomogram model was 0.856 (training set) and 0.772 (test set). The decision curve and calibration curve indicated that the models possessed high clinical practicability.

The nomogram model based on machine learning of MRI radiomics has high clinical value for predicting the risk of MVI of HCC before surgery and guiding decision-making of diagnosis and treatment.