基于CT的影像组学在肝细胞癌病理分级及肝切除术预后预测的应用研究

发布时间：2018-02-27 15:20

本文关键词： 肝细胞癌 CT 影像组学预后 Edmondson-Steiner分级　出处：《南方医科大学》2017年硕士论文　论文类型：学位论文

【摘要】：目的:探索基于增强CT的影像组学预测肝细胞癌Edmondson-Steiner病理分级,以及预测单发肝细胞癌患者肝切除术后的预后的价值。材料和方法:1.回顾性收集并纳入385例肝细胞癌患者,按检查时间及比例分为训练组及测试组,在门脉期CT图像中肿瘤部分中提取17个一阶统计特征、22个GLCM(灰度共生矩阵)特征、14个GLRLM(灰度行程长度矩阵)特征及8个形态特征,每种特征分别不做平滑、3种高斯平滑,以及分别不做滤波、7种差分滤波、10种Gabor滤波。将训练组的影像组学特征经mRMR(最大相关最小冗余)的特征选择算法进行筛选,通过基于AIC的后退法建立建立影像组学标签,将标签应用于测试组。采用ROC(受试者诊断特征曲线)评估模型的预测效能,在训练组及测试组中生成ROC曲线,在ROC曲线上的阳性似然比最大处确定界值,并计算敏感性和特异性、AUC(ROC曲线下面积)和区分病理分级的准确度。以Logistic回归分析方法分析影像组学标签及临床参数与病理分级的关系。2.在肝细胞癌病例中纳入61例单发病灶且行肝切除术治疗、规律随访的患者。通过拉普拉斯高斯滤波器、GLCM、Gabor转换、Wavelet转换方法提取增强扫描动脉期及门脉期CT图像中肿瘤部分的影像组学特征及肝动脉期和门静脉期纹理之间的差异(Dif)。使用ROC筛选特征参数并获得最佳界值。纳入具有ROC的统计显着性的参数及通过单变量回归筛选出p0.10的临床变量,通过Cox比例风险模型对OS(总生存期)进行多变量回归分析。使用AIC(Akaike信息标准)来确定预测OS的最佳模型。根据多变量回归分析确定的重要因素将患者分为两组,使用Kaplan-Meier方法获得OS和DFS(无病生存期)的生存曲线,并进行对数秩检验以比较差异。结果:1.基于门脉期CT影像组学建立的肝细胞癌Edmondson-Steiner病理分级预测模型由筛选出的八个特征组成:Rad Score = 1.009-0.5107×orig_orig_fos_median+0.3362×g.075_gab.20.1_GLCM_homogeneity2 + 0.3719×g.075_gab.15.2_GLRLM_LRLGLE-0.5798×Compactness2-0.2858 ×orig_diff.4_fos_krutosis +0.3261 ×g.100_gab.10.1_GLCM_homogeneity2 +0.6769×g.075_lap.cut2_GLCM_IDN-0.3208×g.075_gab.15.2_GLCM_IMC2该影像组学标签对肝细胞癌Edmondson-Steiner Ⅰ-Ⅱ级及Ⅲ-Ⅳ级的区分有统计学意义(p0.0001),训练组检测效能AUC值为0.766(95%CI:0.824～0.707),敏感性 0.667,特异性 0.753;测试组 AUC 值为 0.673(95%CI:0.775～0.571),敏感性0.674,特异性0.683。甲胎蛋白等临床参数与病理分级无统计学相关性。2.ROC和Cox回归分析筛选出的预测单发肝细胞癌患者肝切除术后预后的影像组学特征参数中,滤波器参数为1.0(filter 1.0)的影像组学特征实现了最佳性能,其中Dif.Scale1.2是OS的优质独立预测因素(p = 0.05)。Kaplan-Meier分析进一步证明在 filter 0 的 Dif.Scale2.2(p = 0.001),filter 1.0 的 Dif.Scale1.2(p = 0.006),Dif.Scale3.2(p = 0.005),filter 1.5 的 Dif.Wavelet 8(p0.001)与 OS 相关。在filter 0 的 Dif.Scale2.2(p = 0.039),filter 1.0 的 Dif.Scale1.2(p = 0.001)和 filter 1.5的 Dif.Wavelet 8(p = 0.007)与 DFS 相关。而巴塞罗那肝癌分期(BCLC)(p = 0.057)等临床参数与OS没有统计学相关性。结论:建立基于门脉期CT影像组学标签可用于预测肝细胞癌为Edmondson-SteinerⅠ-Ⅱ级或Ⅲ-Ⅳ级,且可作为独立预测因子。对于接受肝切除术治疗的单发肝细胞癌患者,CT图像肿瘤区域的Gabor和Wavelet特征,特别是动脉期及门脉期的特征差异值,可为超越BCLC等传统指标的预后预测因素。
[Abstract]:Objective: To explore the enhanced image group CT prediction Edmondson-Steiner pathological grade of hepatocellular carcinoma and the prognosis based on single liver patients after resection for hepatocellular carcinoma. Materials and methods: 1. were retrospectively collected and included 385 cases of hepatocellular carcinoma patients, according to the inspection time and the proportion is divided into training group and testing group. 17 first order statistical feature extraction on the portal venous tumor CT images in the 22 part, GLCM (GLCM), 14 GLRLM (gray run length matrix) characteristics and 8 morphological characteristics of each kind of features do not smooth respectively, 3 Gauss smoothing, and respectively do filtering, 7 a differential filter, 10 Gabor filter. The image features of the group training group by mRMR (optimization) feature selection algorithm was screened by backward method based on AIC to establish image group label, the label applied to the test group (using ROC. Receiver operating characteristic curve) prediction efficiency evaluation model, ROC curve in the training group and the testing group, the positive likelihood in ROC curve on the boundary than the maximum, and calculate the sensitivity and specificity of AUC (area under the ROC curve) and distinguish the pathological grading accuracy. By Logistic regression analysis method of image analysis group labels and clinical parameters and pathologic grading of.2. in 61 patients with single lesion and liver resection in hepatocellular carcinoma cases, regular follow-up of patients. Through the Laplasse Gauss filter, GLCM, Gabor conversion, Wavelet conversion method to extract the difference contrast enhancement in artery phase and portal vein phase scanning group tumor the CT image characteristics and the part between the hepatic arterial phase and portal venous phase texture (Dif). Screening parameters and obtain the best value of ROC. ROC is included in the statistical significant parameters and by single variable Selected clinical variables p0.10 regression, using the Cox proportional hazards model for OS (overall survival) multivariate regression analysis. Using AIC (Akaike information criterion) to determine the best model for predicting OS. According to the multivariate regression analysis the patients were divided into two groups of important factors identified, using the Kaplan-Meier method to obtain OS and DFS (DFS) survival curves and log rank test to compare the differences. Results: 1. hepatocellular carcinoma Edmondson-Steiner grading portal CT images group establishment forecast based on eight features selected by the model: Rad Score = 1.009-0.5107 * orig_orig_fos_median+0.3362 * g.075_gab.20.1_GLCM_homogeneity2 * g.075_gab.15.2_GLRLM_LRLGLE-0.5798 * Compactness2-0.2858 + 0.3719 * orig_diff.4_fos_krutosis +0.3261 g.100_gab.10.1_GLCM_homogeneity2 * +0.6769 * g.075_lap.cut2_GLCM_IDN-0. 3208 x g.075_gab.15.2_GLCM_IMC2 the image group label to distinguish hepatocellular carcinoma Edmondson-Steiner grade I - II and III - IV was statistically significant (P0.0001), training group detection efficiency of AUC value is 0.766 (95%CI:0.824 ~ 0.707), the sensitivity was 0.667, specificity was 0.753; test group AUC = 0.673 (95%CI:0.775 ~ 0.571). The sensitivity was 0.674, specificity for forecasting 0.683. AFP clinical parameters and pathological grading was no correlation between.2.ROC and Cox regression analysis showed that the patients with solitary liver resection for hepatocellular carcinoma prognosis after image group of characteristic parameters, the filter parameter is 1 (filter 1) of the image group characteristics to achieve the best performance, including Dif.Scale1.2 is the quality of the independent predictive factors of OS (P = 0.05).Kaplan-Meier analysis further proved in filter 0 Dif.Scale2.2 (P = 0.001), filter 1 Dif.Scale1.2 (P = 0.006), Dif.Sca Le3.2 (P = 0.005), filter 1.5 Dif.Wavelet 8 (p0.001) and OS. In filter 0 Dif.Scale2.2 (P = 0.039), filter 1 Dif.Scale1.2 (P = 0.001) and filter 1.5 Dif.Wavelet 8 (P = 0.007) and DFS. Barcelona liver cancer (BCLC) (P = 0.057) there were no statistical correlation between clinical parameters and OS. Conclusion: the establishment of portal CT images group label can be used for prediction of hepatocellular carcinoma for Edmondson-Steiner grade I - II or III - IV based, and can be used as an independent predictor. For liver resection for patients with solitary hepatocellular carcinoma, and Gabor Wavelet features of tumor CT image region, especially the differences in arterial phase and portal venous phase value for prognosis of BCLC beyond the traditional indicators such as predictive factors.

【学位授予单位】：南方医科大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R735.7;R730.44

【相似文献】