【作者】 张鑫；

【导师】 梁斌；

【作者基本信息】 华中科技大学 ， 影像医学与核医学（专业学位）， 2023， 硕士

【摘要】 目的:基于术前临床特征及对比增强计算机断层扫描(contrast-enhanced computer tomography,CECT)的影像组学,构建临床及影像组学预测模型,以期预测肝细胞癌(hepatocellular carcinoma,HCC)患者接受经动脉化疗栓塞术(transarterial chemoembolization,TACE)治疗后靶病灶是否缺乏应答反应。方法:本研究回顾性分析2016年1月至2022年5月就诊于华中科技大学同济医学院附属协和医院与中南大学湘雅医院TACE治疗患者的临床资料。华中科技大学同济医学院附属协和医院作为模型训练集,中南大学湘雅医院作为测试集。分别在平扫、动脉、门静脉及延迟期对病灶进行影像组学特征提取。提取的特征包括一阶统计学特征、形状特征、纹理特征及小波特征。静态特征为四个期相提取出的全部影像组学特征;动态特征为相邻两期图像中后一期与前一期特征一一相减的影像组学特征;混合特征指对筛选的动态特征与静态特征进行二次选择所得到的影像组学特征。在静态影像特征、动态影像特征、混合影像特征及临床特征中筛选出2个排序最高的特征。使用随机森林对筛选的特征进行模型构建,得到3个影像组学模型(静态、动态及混合模型)和1个临床特征模型。最后在测试集中进行相应模型验证。结果:本研究共纳入237例患者,在训练集中,病灶缺乏反应患者为24例;在测试集中病灶缺乏反应患者为10例。在静态特征与动态特征组中,排序最高的2个特征分别是基于延迟期的影像组学特征和基于门脉期与动脉期相减的影像组学特征。混合特征为对上述筛选出的2个动态特征与2个静态特征进行二次选择得到的2个特征。在临床特征中,血小板淋巴细胞比值(platelet-to-lymphocyte ratio,PLR)及天门冬氨酸氨基转移酶(aspartate aminotransferase,AST)的排序最高。在训练集分别构建临床模型、基于延迟期特征的静态影像组学模型、基于门脉期与动脉期特征相减的动态影像组学模型及混合特征的影像组学模型。在训练集与测试集中,临床模型的AUC为0.60、0.64,静态影像组学模型的AUC为0.68、0.72,动态影像组学模型的AUC为0.78、0.66,混合影像组学模型的AUC为0.81、0.68。决策曲线分析结果显示,在阈值为0.13附近混合影像组学模型的净获益高于其他模型。结论:本研究的结果显示基于延迟期的静态影像组学模型、基于门脉期与动脉期相减的动态影像组学模型及混合影像组学模型对TACE治疗后病灶是否缺乏反应具有一定预测作用。更多还原

【Abstract】 Objective: To evaluate the efficiency of radiomics signaturesin predicting the lack of response of target lesions after transarterial chemoembolization(TACE)therapy based on preoperative clinical features and contrast-enhanced computed tomography(CECT).Methods: This study consisted of 237 patients treated with TACE at two institutions from January 2016 to May 2022.The radiomics features was extracted from non-contrast,arterial phase,portal venous phase and delay phase with Pyradiomics.The radiomics features extracted from all the four phases were defined as static radiomics features.In two adjacent phases,the features of the latter phase and the previous phase were subtracted as dynamic radiomics features.Hybrid radiomics features refer to the secondary selection of dynamic features and static features.The two highest ranked features were selected from the static radiomics features,dynamic radiomics features,hybrid radiomics features and clinical features.In training set(institution A,n=142),three radiomics models(static model,dynamic model and hybrid model)and one clinical feature model were obtained by Random Forest.Finally,the corresponding model was validated in the external testing set(institution B,n=95).Results: A total of 237 patients were included in this study.In the training set there were 24 patients who presented with the lack of response of target lesions after TACE,and in the test set 10 patients presented with the response lack of target lesions after the procedure.Two delay-based radiomics features were selected from the static radiomics features group.Two radiomics features based on subtraction between portal venous and arterial phase were selected from the dynamic radiomics features group.Two dynamic radiomics features and two static radiomics features were selected for secondary selection,and two features were selected as hybrid features.Among the clinical features,2 clinical features with the highest ranking were selected.In the training set,the clinical model,the static radiomics model,the dynamic radiomics model,and the hybrid radiomics radiomics model were constructed.In the training set and test set,the AUCs of the clinical model,the static radiomics model,the dynamic radiomics model and the hybrid radiomics model were 0.60 and 0.64,0.68 and0.72,0.78 and 0.66,and 0.81 and 0.68.Decision curve analysis shows that the net benefit of the hybrid radiomics model was higher than other models near the threshold of 0.13.Conclusion: This study shows that the static radiomics model based on delayed phase,the dynamic radiomics model based on the subtraction of portal venous and arterial phase and the hybrid radiomics model can predict the lack of response of target lesions after TACE treatment.更多还原