目的：探讨动态增强MRI影像组学列线图预测肝细胞癌（HCC）切除术后3年复发的价值。方法：回顾性收集2007年1月—2016年9月于我院行肝切除术且病理证实为HCC的80例患者（90个HCC病灶）的资料，其中64例作为训练集（以术后3年为研究终点，复发35例，未复发29例），26例作为测试集（复发14例，未复发12例）。患者于术前均行MRI平扫及动态增强扫描。在动脉期、门静脉期和延迟期MR图像上沿肿瘤边缘手动勾画，获得肿瘤三维容积感兴趣区，并提取增强3期各1 029个影像组学特征。利用最大相关最小冗余（Maximal relevance and minimal redundancy，mRMR）算法、最小绝对收缩和选择算子（Least absolute shrinkage and selection operator，LASSO）方法对特征进行特征降维，以建立基于不同期相的影像组学评分。由两位放射科医生确认与预后相关的临床因素，并建立临床评分模型（包括性别、肿瘤大小、病理分级），然后使用多变量Logistic回归构建包含预测效能最佳期相的影像组学评分和临床危险因素的列线图。通过ROC曲线和决策曲线衡量诊断效能和临床应用价值。结果：基于动脉期影像组学评分在测试集的曲线下面积（AUC）为0.82，敏感度和特异度分别为0.83、0.86；临床评分模型在测试集的AUC为0.61，敏感度和特异度分别为0.63、0.60；联合影像组学评分和临床危险因素后的列线图的AUC显著优于临床评分模型（P=0.019），在测试集的AUC为0.83，敏感度和特异度分别为0.85、0.77。结论：基于术前动脉期MRI的影像组学列线图预测HCC切除术后3年复发具有价值，影像组学评分的预测效能与列线图模型的预测效能相当。

Objective: To explore the value of dynamic enhanced MRI-based radiomics nomogram in preoperatively predicting the recurrence(within three years after hepatectomy) of hepatocellular carcinoma(HCC). Methods: A total of 80 HCC patients(90 HCC lesions) who underwent partial hepatectomy in our hospital from January 2007 to September 2016 were enrolled in this retrospective study. A training set consisted of 64 patients(three years after hepatectomy as the endpoint of the study, 35 cases of recurrent HCC lesions and 29 cases of non-recurrent HCC lesions) and a testing set consisted of 26 patients(14 cases of recurrent HCC lesions and 12 cases of non-recurrent HCC lesions). All patients underwent preoperative non-enhanced MR scanning and liver acquisition with volume acceleratio(LAVA) enhanced scanning within 2 weeks preoperatively. Based on the arterial, portal and delayed phases of MRI enhanced images, 1 029 radiomics features based on the three-dimensional volume of the tumors were extracted. The maximal relevance and minimal redundancy(mRMR) and least absolute shrinkage and selection operator(LASSO) methods were used for data dimension reduction to establish radiomics socre(radscore) based on different phases of enhanced MR images. Meanwhile, the preoperative clinical characteristics associated with prognosis were recorded by two radiologists and then clinical score(including gender, tumor size and pathological grading) was built. Multivariate logistic regression was used to build a nomogram which integrated the optimal radscore and clinical risk factors. Predictive performance and clinical usefulness were evaluated by the area under the curve(AUC) of receiver operating characteristics(ROC) and decision curves. Results: The AUC of the radscore based on the arterial phase in the testing set was 0.82, sensitivity of 0.83 and specificity of 0.86. The AUC of the clinical score in the testing set was 0.61, sensitivity of 0.63 and specificity of 0.60. The nomogram integrating radscore and clinical risk factors showed better predictive performance than the clinical score(P=0.019). The AUC of the nomogram in the testing set was 0.83, sensitivity of 0.85 and specificity of 0.77. Conclusion: The radiomics nomogram based on the arterial phase of enhanced MRI can be used to preoperatively predict the recurrence(within three years after hepatectomy) of HCC, and the predictive performance of radscore is similar to that of the radiomics nomogram.