基于深度学习的磁共振成像影像组学预测胎盘植入性疾病  

Prediction of Placenta Accreta Spectrum by MRI Imaging Based on Deep Learning

作　　者：令潇 胡玉瑞 王颖超 李洁[1] 李浩源 张静[1] 岳松虹 LING Xiao;HU Yurui;WANG Yingchao;et al(Department of Magnetic Resonance,The Second Hospital&Clinical Medical School,Lanzhou University,Lanzhou Gansu 730030,China;Hebei North University,Zhangjiakou Hebei 075000,China;Department of Medical Imaging,Zhangye People′s Hospital Affiliated to Hexi University,Zhangye Gansu 734000,China)

机构地区：[1]兰州大学第二医院(第二临床医学院)核磁共振科,甘肃兰州730030 [2]河北北方学院,河北张家口075000 [3]河西学院附属张掖人民医院医学影像科,甘肃张掖734000

出　　处：《实用妇产科杂志》2025年第3期230-236,共7页Journal of Practical Obstetrics and Gynecology

基　　金：甘肃省科技计划项目(编号:21JR7RA438);2024年高校教师创新基金项目(编号:2024A-153);兰州市人才创新创业项目(编号:2022-RC-74)。

摘　　要：目的:探讨基于磁共振成像(MRI)矢状位T2加权成像(T2WI)的深度学习影像组学预测高危孕妇胎盘植入性疾病(PAS)的诊断价值。方法:回顾性分析兰州大学第二医院(第二临床医学院)和河西学院附属张掖人民医院2019年1月至2023年12月265例因可疑胎盘植入行MRI检查的孕妇的完整资料,按7∶3将患者随机分为训练组(n=172)与验证组(n=93),并根据术中是否诊断PAS分为PAS组和正常组。采用多因素Logistic回归分析筛选临床影像特征独立危险因素。分别基于矢状位T2WI图像提取影像组学特征,基于密集连接卷积神经网络-121(DenseNet-121)模型作为深度学习特征提取的基础模型,构建传统的临床模型、影像组学模型、深度学习模型预测PAS,采用受试者工作特征(ROC)曲线下面积(AUC)评价各模型的诊断效能,AUC值最大者确定为最优模型。结果:训练组及验证组中,PAS组与正常组在剖宫产次数≥2次、存在前置胎盘及胎盘厚度>40 mm差异均有统计学意义(P<0.05)。多因素Logistic回归分析得出剖宫产次数≥2次、胎盘厚度>40 mm及存在前置胎盘为发生PAS的独立危险因素,构建的所有模型中深度学习联合临床的组合模型的诊断效能显著优于其他3种模型,其在训练组和验证组中的AUC分别为0.96(95%CI 0.93~0.98)、0.91(95%CI 0.87~0.95)。结论:基于MRI的深度学习联合临床模型在诊断PAS方面可能比临床或传统影像组学模型表现出更好的性能。Objective:To explore the value of deep learning imageomics based on MRI sagittal T2WI images in predicting placenta accreta spectrum in high-risk pregnant women.Methods:The complete data of 265 pregnant women who underwent MRI due to suspected placenta implantation in The Second Hospital&Clinical Medical School,Lanzhou University and Zhangye People′s Hospital Affiliated to Hexi University from January 2019 to December 2023 were analyzed retrospectively.The patients were randomly divided into training group(n=172)and validation group(n=93)at 7∶3.Multivariate Logistic regression analysis was used to screen the independent risk factors among clinical and imaging characteristics.Radiomics features were extracted based on sagittal T2WI images.Using the DenseNet-121 model as the basic model for deep learning feature extraction,traditional clinical model,radiomic model and deep learning model were constructed to predict PAS.The diagnostic efficiency of each model was evaluated by the area under the receiver operating characteristic(ROC)curve(AUC).Finally,the model with the highest performance was determined as the optimal model.Results:In both the training and validation groups,the PAS group and normal group exhibited statistically significant differences(P<0.05)in terms of the number of cesarean section≥2,history of placenta previa,and placental thickness>40 mm.Multivariate Logistic regression analysis revealed that cesarean section history,placental thickness and placenta previa were independent risk factors for predicting PAS.Among all the models constructed,the diagnostic performance of the combination model of deep learning combined with clinic was higher than the other three models.The AUC in training group and verification group were 0.96(95%CI 0.93-0.98)and 0.91(95%CI 0.87-0.95)respectively.Conclusions:The combined clinical model of deep learning based on MRI may have better performance in the diagnosis of PAS than clinical or traditional radiomic models.

关 键 词：磁共振成像 深度学习 胎盘植入性疾病 影像组学 

分 类 号：R714.56[医药卫生—妇产科学]