基于CT及MRI影像组学在术前预测肝细胞癌微血管侵犯和治疗预后的临床研究

Research on the Preoperative Prediction of Microvascular Invasion and Treatment Prognosis of Hepatocellular Carcinoma Based on CT and MRI Radiomics

摘要: 目的：探讨基于CT及MRI影像组学结合机器学习森林模型对在术前预测肝细胞癌(Hepatocellular Carcinoma, HCC)中微血管侵犯(Microvascular Invasion, MVI)的价值及探索肝细胞癌MVI的影像学特征，建立有效可行的预测模型，为患者提供术前准确的MVI预测，为临床决策提供有力支持，以实现精准医疗。方法：对80例接受CT增强扫描(CECT)和91例完成MRI增强扫描(CE-MRI)的肝细胞癌(HCC)患者进行回顾性收集病史记录、术前一周血清学检测数据和CT/MRI动脉期(AP)、门静脉期(PVP)、延迟期(DP)影像资料。继而根据诊断金标准组织病理学结果，将样本分为MVI阳性和阴性两组，用非参数检验、卡方检验或者Fisher精确检验等方法选择出可能的风险因素，最后用多因素Logistic回归分析确定独立危险因素。对各时相图像分别提取影像组学特征并实施降维筛选，采用三级递进式特征缩减策略逐层剔除冗余变量，筛选出单期最优特征子集，随后整合增强CT三期、增强MRI三期各自筛选后的最优特征，建立多期联合特征集。以此为基础建立综合评分体系，将全部研究对象以8:2的比例进行随机拆分，划分出训练组和测试组。将筛选出的独立临床预测因子与不同扫描时相下提取的影像特征融合，依托随机森林算法在建模集中完成预测模型搭建。借助受试者工作特征(ROC)曲线，对临床基线模型(CF)以及各时相影像组学模型的鉴别能力进行评估。针对同时进行过CECT及CE-MRI检查的57例患者，重复上述特征提取、筛选与分析流程并按8:2随机分组，搭建随机森林模型，利用ROC曲线评价各类模型的预测表现，同时计算得出准确率、灵敏度与特异度三项评价指标。结果：针对80例行CECT各期(动脉期AP、门静脉期PVP、延迟期DP)受试者的影像中筛选出组学特征，这些指标在肝细胞癌MVI阳性与阴性人群中存在显著区别(P < 0.05)。采用完全一致的筛选方案，在91例CE-MRI影像样本里也得到了各组间存在统计学差异的组学特征。最后整合全部入组的114例患者临床数据，开展多因素Logistic回归分析显示，MVI唯一独立危险因素为病灶切面最大直径(P < 0.001)：CECT扫描显示MVI阳/阴性最大径7.2 cm (4.8~11.5 cm)/4.2 cm (3.0~6.2 cm)，CE-MRI扫描显示阳/阴性最大径7.2 cm (4.5~11.0 cm)/3.5 cm (3.0~5.8 cm)，组内差异均显著。随机森林建模后，测试集CT、单纯临床模型(CF) AUC为0.705，MRI临床模型AUC为0.748；CE-MRI各单期及三期联合组学模型AUC依次0.787、0.765、0.764、0.778，普遍优于临床模型。融合临床与组学特征后，CE-MRI联合模型效能高于单一组学模型，三期联合提升明确；CECT提升不显著，仅部分期相小幅上升，证实融合模型可优化HCC MVI预测能力。57例双模态同步检查组中，CECT门静脉期、延迟期及三期联合组学模型数值上高于各CE-MRI组学模型，但差异无统计学意义，提示两种模态预测效能高度相近效能良好。结论：基于增强CT、增强MRI影像组学特征联合随机森林机器学习算法，可实现肝细胞癌MVI的术前有效预判。不同扫描期相所构建的CECT、CE-MRI组学模型，鉴别效能未见显著统计学差异。同时，两种影像模态的组学模型均表现出有效的预测价值，整体预测水平相近、无明显统计学差异。不同期相的增强影像组学模型均可稳定、有效地评估HCC术前MVI状态，可为临床个体化诊疗方案的制定提供可靠的参考依据。

Abstract: Objective: To evaluate the value of combining CT and MRI radiomics with machine learning random forest models in preoperatively predicting microvascular invasion (MVI) in hepatocellular carcinoma (HCC), to explore imaging characteristics associated with HCC MVI, and to establish an effective and feasible predictive model. This aims to provide accurate preoperative MVI prediction for patients and offer strong support for clinical decision-making, thereby advancing precision medicine. Methods: Retrospective data were collected from 80 patients with hepatocellular carcinoma (HCC) who underwent contrast-enhanced CT (CECT) and 91 patients who completed contrast-enhanced MRI (CE-MRI), including medical history, preoperative serum laboratory results, and imaging data from the arterial phase (AP), portal venous phase (PVP), and delayed phase (DP) of CT/MRI. Based on histopathological findings as the gold standard for diagnosis, samples were divided into MVI-positive and MVI-negative groups. Potential risk factors were identified using non-parametric tests, chi-square tests, or Fisher’s exact test, followed by multivariate logistic regression analysis to determine independent predictors. Radiomic features were extracted from images at each phase and subjected to dimensionality reduction and feature selection. A three-stage progressive feature reduction strategy was applied to iteratively eliminate redundant variables, identifying optimal feature subsets for each phase. The best features from both CECT and CE-MRI across their respective three phases were then integrated to form a combined multi-phase feature set. Based on this, a comprehensive scoring system was developed. All study subjects were randomly split into training and testing sets in an 8:2 ratio. Independent clinical predictors were fused with radiomic features extracted from different scanning phases, and a predictive model was constructed using the random forest algorithm on the training set. The discriminatory performance of the clinical baseline model (CF) and various radiomic models was evaluated using receiver operating characteristic (ROC) curves. For the 57 patients who underwent both CECT and CE-MRI, the same process of feature extraction, selection, and analysis was repeated, followed by random 8:2 division into training and testing sets. Random forest models were built accordingly, and their predictive performance was assessed via ROC curves, along with calculation of accuracy, sensitivity, and specificity. Results: Omics features were selected from imaging data of 80 patients undergoing CECT at different phases (arterial phase, AP; portal venous phase, PVP; and delayed phase, DP), showing significant differences between microvascular invasion (MVI)-positive and MVI-negative hepatocellular carcinoma groups (P < 0.05). Using an identical selection protocol, similar omics features with statistical differences among groups were also identified in 91 CE-MRI image samples. Integrating clinical data from all 114 enrolled patients, multivariate logistic regression analysis revealed that the maximum diameter of the lesion on cross-section was the only independent risk factor for MVI (P < 0.001): the maximum diameters for MVI-positive and -negative cases were 7.2 cm (4.8~11.5 cm) vs. 4.2 cm (3.0~6.2 cm) on CECT, and 7.2 cm (4.5~11.0 cm) vs. 3.5 cm (3.0~5.8 cm) on CE-MRI, with significant differences within each group. After random forest modeling, the AUCs of the test set were 0.705 for CT alone and the clinical model (CF), and 0.748 for the MRI-clinical model. The AUCs of the single-phase and combined three-phase omics models based on CE-MRI were 0.787, 0.765, 0.764, and 0.778, respectively, generally outperforming the clinical models. When combining clinical and omics features, the performance of the CE-MRI integrated model exceeded that of standalone omics models, with a clear improvement when using the three-phase combination; however, CECT showed no significant enhancement, with only minor improvements in certain phases, confirming that fusion models can optimize prediction of HCC MVI. In the subgroup of 57 patients who underwent dual-modality synchronized examinations, the values of the CECT PVP, delayed phase, and three-phase combined omics models were numerically higher than those of the corresponding CE-MRI omics models, although the differences were not statistically significant, indicating that both modalities have highly comparable and robust predictive performance. Conclusion: Combining enhanced CT and enhanced MRI radiomic features with a random forest machine learning algorithm enables effective preoperative prediction of microvascular invasion (MVI) in hepatocellular carcinoma. There was no significant statistical difference in discriminatory performance among the CECT and CE-MRI radiomic models constructed from different scanning phases. Moreover, both imaging modality-based radiomic models demonstrated effective predictive value, with comparable overall performance and no notable statistical difference. Radiomic models derived from various enhanced imaging phases can stably and effectively assess preoperative MVI status in HCC, providing reliable reference for individualized clinical diagnosis and treatment planning.