摘要
目的 本研究旨在基于术前多参数MRI的深度迁移学习(deep transfer learning,DTL)特征建立模型,以预测子宫内膜癌(endometrial carcinoma,EC)患者淋巴脉管间隙浸润(lymphovascular space invasion,LVSI)状态。材料与方法 回顾性收集2016年2月至2023年7月本院187例经手术病理确诊的EC患者的临床信息及术前MRI图像。并按7∶3比例随机分为训练集(131例)和测试集(56例)。在轴位T2加权成像、扩散加权成像、表观扩散系数(apparent diffusion coefficient,ADC)图及对比增强T1加权图像上手动勾画病灶感兴趣区。采用ResNet50、ResNet101及DenseNet121网络建立12种DTL模型,接着,采用平均值、最大值及最小值三种决策级融合方法建立融合模型,并从中选取最佳模型作为最终的DTL模型。通过单因素和多因素logistic回归分析筛选出临床特征后建立临床模型,并使用logistic回归联合DTL和临床特征建立DTL-临床联合模型。采用受试者工作特征曲线评估模型诊断EC患者LVSI的效能,通过DeLong检验比较曲线下面积(area under the curve,AUC),校准曲线分析模型的拟合优度,决策曲线分析(decision curve analysis,DCA)探讨模型的临床适用性。结果 测试集中,基于ADC图像建立的ResNet101模型在诊断EC患者LVSI时表现出最高的AUC值,为0.850 [95%置信区间(confidence interval,CI):0.736~0.963]。。采用平均值融合方法建立的融合模型,在测试集中AUC值最高,达到了0.932(95%CI:0.868~0.996),为最佳DTL模型。logistic回归分析表明年龄是EC患者LVSI的独立危险因素。DTL-临床联合模型在测试集中AUC为0.934(95%CI:0.871~0.997),诊断效能优于临床模型[AUC为0.554 (95%CI:0.436~0.671),P<0.001],与DTL模型比较差异无统计学意义(P=0.909)。Hosmer-Lemeshow检验显示联合模型在训练集和测试集中均具有较好的拟合效果(P=0.814及0.402),DCA显示临床净获益更大。结论 基于术前多参数MRI建立的DTL模型,以及将DTL特征与临床特征相结合建立的联合模型,能有效预测EC患者LVSI状态,诊断效能优于临床模型。DTL在小样本EC的MRI数据中表现优异,为LVSI术前预测提供重要临床辅助工具。
Objective:This study aimed to develop a model based on deep transfer learning(DTL) features from preoperative multiparametric magnetic resonance imaging(MRI) to predict lymphovascular space invasion(LVSI) status in patients with endometrial carcinoma(EC).Materials and Methods:A retrospective analysis was conducted on clinical information and preoperative MRI images of 187 EC patients who were surgically and pathologically confirmed in our hospital from February 2016 to July 2023.The patients were randomly divided into a training set(131 patients) and a test set(56 patients) in a 7∶3 ratio.Regions of interest were delineated on axial T2-weighted imaging,diffusion-weighted imaging,apparent diffusion coefficient(ADC) maps,and contrast-enhanced T1-weighted imaging,manually.Subsequently,12 DTL models were established using ResNet50,ResNet101,and DenseNet121 networks.Fusion models were then established using three decision-level fusion methods:mean,maximum,and minimum,with the best model selected as the final DTL model.A clinical model was established after screening clinical features through univariate and multivariate logistic regression analysis,and a DTL-clinical combined model was developed using logistic regression incorporating DTL and clinical features.The receiver operating characteristic curve was used to assess the diagnostic performance of the models for LVSI in EC patients,the area under the curve(AUC) was compared using the DeLong test.The calibration curve was used to analyze the goodness of fit of the models,and the decision curve was used to explore the clinical applicability of the models.Results:In the test set,the ResNet101 model based on the ADC images showed the highest AUC value of 0.850 [95% confidence interval(CI):0.736 to 0.963] for diagnosing LVSI in EC patients.The fusion model established using the mean fusion method had the highest AUC value of 0.932(95% CI:0.868 to 0.996) in the test set,representing the best DTL model.Logistic regression analysis indicated that age was an independent risk factor for LVSI.The DTL-clinical combined model had an AUC of 0.934(95% CI:0.871 to 0.997) in the test set,with significantly better diagnostic performance than the clinical model [AUC:0.554(95% CI:0.436 to 0.671),P < 0.001] and no statistical difference compared to the DTL model(P = 0.909).The combined model demonstrated good fit in both the training and test sets(Hosmer-Lemeshow test:P = 0.814 and 0.402,respectively) and offered greater clinical net benefit.Conclusions:The DTL model based on preoperative multiparametric MRI,as well as the combined model integrating DTL features with clinical features,can effectively predict the LVSI status of EC patients,outperforming clinical models.DTL demonstrates excellent performance on our small-sample EC MRI data,providing important clinical assistance for preoperative LVSI prediction.
作者
郭冉
彭如臣
李艳翠
沈秀芝
郝攀
信瑞强
GUO Ran;PENG Ruchen;LI Yancui;SHEN Xiuzhi;HAO Pan;XIN Ruiqiang(Department of Radiology,Beijing Luhe Hospital,Capital Medical University,Beijing 101149,China)
出处
《磁共振成像》
北大核心
2025年第3期70-76,82,共8页
Chinese Journal of Magnetic Resonance Imaging
基金
2023年度首都医科大学附属北京潞河医院青年科研孵育专项(编号:LHYY2023-LC209)。
关键词
子宫内膜癌
淋巴脉管间隙浸润
多参数磁共振成像
深度学习
迁移学习
endometrial carcinoma
lymphvascular space invasion
multiparametric magnetic resonance imaging
deep learning
transfer learning
