基于MR超分重建图像的影像组学模型术前预测早期宫颈癌淋巴血管浸润的价值  

The value of radiomics model based on MR hyperresolution reconstructed images in preoperative prediction of lymphatic vascular infiltration in early cervical cancer

作　　者：何月明[1] 罗锦文[2] 张婷[1] 戚莹莹 HE Yue-ming;LUO Jin-wen;ZHANG Ting(Department of Gynecology,the Fifth Affiliated Hospital of Guangzhou Medical University,510735 Guangzhou,China)

机构地区：[1]广州医科大学附属第五医院妇科,广州510735 [2]广州医科大学附属第五医院影像科,广州510735

出　　处：《放射学实践》2024年第9期1213-1220,共8页Radiologic Practice

基　　金：广东省医学科学技术研究基金(编号:A2024300);广州市基础与应用基础研究项目(项目编号202102080250)。

摘　　要：目的:探讨基于MR深度迁移学习超分重建图像的影像组学模型术前预测早期宫颈癌(CC)淋巴血管间隙浸润(LVSI)的价值。方法:回顾性分析经术后病理证实的100例早期CC患者的MRI及临床资料,对矢状面T 2WI抑脂非增强序列原始图像(OI)进行深度迁移学习超分重建图像(SRI),并采用ITK-SNAP软件在OI及SRI上对全肿瘤区域进行3D标注,根据病理结果分为LVSI阳性与LVSI阴性组,并按照8:2比例随机分为训练集(80例)和验证集(20例)。对OI、SRI标注图像3D VOI进行特征提取及最小绝对收缩与选择算子(LASSO)回归筛选影像组学特征,并分别建立LightGBM影像组学模型,使用AUC评估模型的诊断效能,使用决策曲线分析(DCA)评估模型的临床价值。结果:基于OI影像组学模型预测宫颈鳞癌LVSI状态,训练集AUC=0.795(95%CI:0.696~0.894),敏感度为0.533,特异度为0.920;验证集AUC=0.637(95%CI:0.350~0.924),敏感度为0.429,特异度为0.923。基于SRI影像组学模型预测宫颈鳞癌LVSI状态,训练集AUC=0.817(95%CI:0.722~0.913),敏感度为0.920,特异度为0.717;验证集AUC=0.815(95%CI:0.625~1.000),敏感度为0.667,特异度为0.786。两组图像训练集和验证集中均显示出良好的校准和区分能力,SRI较OI影像组学模型的诊断效能明显提高,DCA结果表明模型具有较高的临床价值。结论:基于MR深度迁移学习SRI影像组学模型对术前预测宫颈癌LVSI状态具有良好的应用价值,较OI影像组学模型的诊断效能有所提高,有助于更好地指导临床治疗决策。Objective:To explore the value of a radiomics model based on deep transfer learning super-reconstructed images of MR for preoperative prediction of lymphatic vessel infiltration(LVSI)of early cervical cancer(CC).Methods:A retrospective analysis was conducted on the MR images and clinical data of 100 early CC patients confirmed by postoperative pathology.The original images(OI)of sagittal T 2WI lipid-pressure non enhanced sequence were reconstructed using deep transfer learning(SRI).The entire tumor area was labeled in 3D on both OI and SRI using ITK-SNAP software.According to the pathological results,the patients were divided into LVSI positive and LVSI negative groups,and randomly divided into a training set(80 cases)and a validation set(20 cases)in a 8:2 ratio.Feature extraction and Least Absolute Shrinkage and Selection Operator(LASSO)regression were performed on the annotated 3D VOI images from OI and SRI to screen for radiomics features.LightGBM radiomics models were established,and the diagnostic efficacy of the models was evaluated using AUC.The clinical value of the models was evaluated using Decision Curve Analysis(DCA).Results:The diagnostic efficacy of the OI radiomics model in CC was evaluated with a training set AUC of 0.795(95%CI:0.696~0.894),sensitivity of 0.533,and specificity of 0.920.The validation set AUC=0.637(95%CI:0.350~0.924),with a sensitivity of 0.429 and a specificity of 0.923.The diagnostic efficacy of SRI radiomics classification in CC was evaluated with training set AUC=0.817(95%CI:0.722~0.913),sensitivity of 0.920,specificity of 0.717,validation set AUC=0.815(95%CI:0.625~1.000),sensitivity of 0.667,specificity of 0.786.The two sets of image training sets and validation sets showed good calibration and discrimination abilities,and the diagnostic efficiency of SRI compared to OI's radiomics model was significantly improved.The DCA results showed that the model had high clinical value.Conclusion:The SRI radiomics model based on MR deep transfer learning has good application value in pre

关 键 词：宫颈癌 磁共振成像 超分重建 影像组学 淋巴血管浸润 

分 类 号：R737.33[医药卫生—肿瘤] R445.2[医药卫生—临床医学]