基于增强CT放射组学预测肝细胞肝癌病理分级

目的 观察增强CT放射组学术前预测肝细胞肝癌（HCC）病理分级的可行性及价值。方法 回顾分析429例经手术病理证实的HCC患者，分为训练组（n=329）和测试组（n=100），记录其临床特征；提取动脉期（AP）及静脉期（VP）CT图像的放射组学特征，应用最小绝对值收敛和选择算子（LASSO）回归分析法对其进行降维，筛选最有价值的组学特征后，构建基于AP、VP、AP+VP图像特征的组学模型，计算2组放射学评分并进行二分类判别。根据病理结果定义高级别和低级别HCC，采用10倍交叉验证训练选择最优组学预测模型，筛选对预测HCC病理分级有意义的临床特征后，构建临床模型以及联合组学特征和临床特征的联合模型。绘制3种模型预测训练组和测试组HCC病理分级的ROC曲线，评估其诊断能力。结果 联合组学模型最优，其判别训练组及测试组高级别和低级别HCC的放射学评分的差异均有统计学意义（Z=8.58、3.24，P均<0.05）。测试组中，联合模型预测HCC病理分级的AUC值（0.70）与组学模型（0.69）和临床模型（0.63）差异均无统计学意义（P均>0.05）。结论 基于增强CT图像的放射组学特征可用于术前预测HCC病理分级。

Prediction of pathological grade of hepatocellular carcinoma based on enhanced CT radiomics

Objective To investigate the feasibility and value of preoperative prediction of pathological grade of hepatocellular carcinoma (HCC) based on enhanced CT radiomics. Methods Imaging and clinical data of 429 HCC patients confirmed by surgical pathology were retrospectively analyzed. The patients were divided into training group (n=329) and test group (n=100), and their clinical characteristics were recorded. Radiology features of arterial-phase (AP) and portal venous-phase (VP) CT images were extracted, the least absolute shrinkage and selection operator method (LASSO) were used to reduce dimension and select the most valuable radiomics features. Then CT radiomics models were built base on AP features, VP features and AP+VP features, respectively. Radiological scores (rad-score) of 2 groups were calculated and then classified. According to surgical pathology results, the pathological grade of HCC was defined as high-grade and low-grade, and the optimal radiomics prediction model was selected through 10-fold cross-validation training. Finally clinical model and combined model (clinical features combined with radiomics) were constructed after screening clinical characteristics for predicting pathological grade of HCC. ROC curves of the above 3 models for predicting pathological grade of HCC in training group and test group were drawn, and their diagnostic efficacy were evaluated. Results Combined radiomics model was the best among 3 models, and the rad-scores of high-grade and low-grade HCC were significantly different in both training group and test group (Z=8.58, 3.24, both P<0.05). In test group,no statistical difference of AUC of combined model(0.70),of radiomics model (0.69) nor clinical model (0.63) was detected for predicting pathological grading of HCC (all P>0.05). Conclusion Radiomics features based on enhanced CT images can be used to preoperative predict pathological grade of HCC, providing reference for diagnosis and treatment of HCC.