基于深度学习的3D超分辨率重建技术的MRI影像组学预测TACE联合分子靶向药物治疗不可切除肝癌的疗效  

MRI radiomics based on deep learning 3D super-resolution reconstruction technology for predicting the efficacy of TACE combined with molecular targeted drugs in the treatment of unresectable hepatocellular carcinoma

作　　者：董亚宁 朱菊芳 毛珂 翟晓阳 段金辉 韩东明[1] DONG Yaning;ZHU Jufang;MAO Ke;ZHAI Xiaoyang;DUAN Jinhui;HAN Dongming(Department of Magnetic Resonance,the First Affiliated Hospital of Xinxiang Medical University,Xinxiang 453100,China)

机构地区：[1]新乡医学院第一附属医院磁共振科,新乡453100

出　　处：《磁共振成像》2024年第12期124-130,共7页Chinese Journal of Magnetic Resonance Imaging

基　　金：北京康盟慈善基金会医学科研发展基金项目(编号:B21145AN)。

摘　　要：目的探讨基于深度学习的3D超分辨率重建技术的MRI影像组学在预测经动脉化疗栓塞术(transcatheter arterial chemoembolization,TACE)联合分子靶向药物治疗不可切除肝癌疗效中的价值。材料与方法回顾性分析122例原发性肝癌(hepatocellular carcinoma,HCC)患者的资料,根据改良实体瘤疗效评价标准(modified response evaluation criteria in solid tumors,mRECIST)分为客观反应组(完全缓解+部分缓解)68例和无客观反应组(疾病进展+病情稳定)54例。基于生成对抗网络的3D超分辨率重建技术将MRI增强动脉早期图像分辨率提高至原来的2倍。以8∶2的比例随机分为训练集及验证集,于重建前后图像分别勾画感兴趣区体积进行影像组学分析筛选影像组学特征并计算影像组学评分。采用logistic回归建立重建前及重建后影像组学模型并筛选临床特征建立临床模型。采用受试者工作特征(receiver operating characteristic,ROC)曲线评估模型性能,采用DeLong检验比较曲线下面积(area under the curve,AUC)。采用决策曲线(decision curve analysis,DCA)分析各模型的临床价值。结果logistic回归分析表明肿瘤直径[比值比(odds ratio,OR)值=1.311,95%置信区间(confidence interval,CI)=1.112~1.547,P<0.001]与动脉期强化(OR值=9.466,95%CI=2.489~36.001,P<0.001)是联合治疗疗效的独立预测因素。重建后影像组学模型预测性能最佳,训练集与验证集AUC分别为0.883(95%CI:0.814~0.952)、0.844(95%CI:0.656~0.999),高于重建前影像组学模型0.847(95%CI:0.765~0.928)、0.753(95%CI:0.554~0.953)及临床模型0.834(95%CI:0.754~0.914)、0.760(95%CI:0.564-0.956)。然而,训练集及验证集中各个模型的AUC差异不具有统计学意义(P均>0.05)。DCA表明训练集在大于0.34、验证集在0.36~0.59以及大于0.71阈值范围内,重建后影像组学模型临床净收益最大。结论采用基于生成对抗网络的3D超分辨率重建技术的MRI影像组学在预测TACE联合分子靶向Objective:To explore the value of MRI radiomics based on deep learning three-dimensional(3D)super-resolution reconstruction technology in predicting the efficacy of transcatheter arterial chemoembolization(TACE)combined with molecularly targeted drugs in treating unresectable hepatocellular carcinoma(HCC).Materials and Methods:A retrospective analysis was conducted on data from 122 patients with primary HCC,divided into an objective response group(complete remission+partial remission,n=68)and a non-objective response group(progressive disease+stable disease,n=54)according to the modified response evaluation criteria in solid tumors(mRECIST).A 3D super-resolution reconstruction technique based on generative adversarial networks was used to double the resolution of MRI-enhanced arterial early images.The dataset was randomly divided into training and validation sets in an 8∶2 ratio.Radiomic features were extracted from volume of interests delineated on both pre-and post-reconstructed images,and subsequently,radiomic scores were calculated.Logistic regression classifiers were used to establish radiomic models for both preand post-reconstructed images.Multivariable logistic regression was employed to screen clinical characteristics and establish a clinical model.Model performance was evaluated using receiver operating characteristic(ROC)curves,with area under the curve(AUC)compared via DeLong's test.Decision curve analysis(DCA)was used to assess the clinical value of each model.Results:Logistic regression analysis identified tumor diameter[odds ratio(OR)=1.311,95%confidence interval(CI)=1.112-1.547,P<0.001]and arterial phase enhancement(OR=9.466,95%CI=2.489-36.001,P<0.001)as independent predictors of treatment efficacy for HCC.The post-reconstruction radiomic model exhibited the best predictive performance,with an AUC of 0.883(95%CI:0.814-0.952)in the training set and 0.844(95%CI:0.656-0.999)in the validation set.These results surpassed those of the pre-reconstruction radiomic model,which had AUC values of 0.847(95%

关 键 词：肝癌 经动脉化疗栓塞术 磁共振成像 影像组学 深度学习 

分 类 号：R445.2[医药卫生—影像医学与核医学] R735.7[医药卫生—诊断学]