探讨基于CT影像组学构建的肝细胞癌（肝癌）微血管侵犯（MVI）预测模型的价值。

回顾性分析2015至2022年昆明理工大学附属医院肝胆外科经肝癌手术切除的129例患者临床资料。其中男108例，女21例；年龄25~83岁，中位年龄52岁。按3∶1随机数表法将患者分成训练集96例，测试集33例。所有患者均在术前1个月内接受增强CT检查。通过影像组学人工智能大数据分析平台运用Lasso算法从临床和影像组学特征中筛选出最优特征。使用增强CT图像数据建立基于术前CT的肿瘤和瘤周0~1、1~2、2~3 cm的单一临床模型（C模型）、单一影像组学模型（R模型）。根据最优特征相对应的加权系数，得到每一个模型的影像组学评分（Rad-score），依据Rad-score得出每个模型的ROC AUC。模型预测能力采用一致性指数（C-index）评估，指数越高预测能力越强。

运用Lasso算法从包含临床和影像组学特征在内的1 818个特征中分别筛选出肿瘤、瘤周0~1、1~2、2~3 cm的3、9、15、50个最优特征指标。训练集中采用五折交叉验证的方法进行惩罚参数优化，构建Lasso-Logistic回归模型。测试集中R模型预测 MVI风险，较C模型的表现更好（AUC=0.883、0.848、0.800、0.848和0.500、0.704、0.500、0.639）。训练集和测试集4个R模型估计的风险和实际MVI发生之间有良好的一致性，R模型均具有很好的预测能力（C-index=0.746和0.883，0.738和0.848，0.732和0.800，0.672和0.848），并有很好的校正能力。

本研究建立了基于术前CT的肿瘤和瘤周0~1、1~2、2~3 cm的R模型，且R模型均优于C模型预测结果。R模型转化的Rad-score可作为MVI发生的独立预测因素。

To investigate the prediction value of CT-based radiomics model for microvascular invasion (MVI) in hepatocellular carcinoma (HCC).

Clinical data of 129 patients who underwent surgical resection in the Affiliated Hospital of Kunming University of Science and Technology from 2015 to 2022 were retrospectively analyzed. Among them, 108 patients were male and 21 female, aged from 25 to 83 years, with a median age of 52 years. According to the 3: 1 ratio using random number table method, all cases were divided into the training (n=96) and test sets (n=33). All patients received enhanced CT scan within preoperative 1 month. Through the artificial intelligence big-data analysis platform of imaging radiomics, Lasso algorithm was used to screen the optimal features from clinical and imaging radiomic features. Using enhanced CT images, a single clinical model (C model) and a single imaging radiomic model (R model) were constructed based on preoperative CT scan of tumors and peritumoral 0-1, 1-2 and 2-3 cm. According to the weighted coefficient corresponding to the optimal features, the radiomic score (Rad-score) of each model was obtained, and the area under the ROC curve (AUC) of each model was calculated according to the Rad-score. The prediction capability of the model was evaluated by the consistency index (C-index). The higher the index, the higher the prediction ability.

Lasso algorithm was employed to screen 3, 9, 15 and 50 optimal features of tumor and peritumoral 0-1, 1-2 and 2-3 cm from 1818 clinical and imaging radiomic features. In the training set, the penalty parameters were optimized by 5-fold cross-validation, and Lasso-Logistic regression model was constructed. In the test set, R model could better predict the risk of MVI than C model (AUC=0.883, 0.848, 0.800, 0.848 and 0.500, 0.704, 0.500 and 0.639). High consistency was found between the risk estimated by 4 R models in the training and test sets and the actual risk of MVI, indicating R models yielded good predictive ability (C-index=0.746 and 0.883, 0.738 and 0.848, 0.732 and 0.800, 0.672 and 0.848) and favorable correction performance.

In this study, R models of tumor and peritumoral 0-1, 1-2 and 2-3 cm are established based on preoperative CT scan. R models perform better than C models. Rad-score transformed by R models can be used as an independent predictor of MVI.