摘要
目的:探讨基于MRI动态增强扫描(DCE-MRI)影像组学及深度迁移学习特征结合临床特征预测乳腺癌患者前哨淋巴结(SLN)转移状态的价值。方法:回顾性分析本院2020年1月-2023年4月经临床病理证实为浸润性乳腺癌患者MRI影像及临床资料,按8:2比例随机划分训练集145例(SLN阴性95例、SLN阳性50例)和验证集37例(SLN阴性28、SLN阳性9例)。于DCE-MRI图像上逐层手工勾画肿瘤感兴趣区并提取影像组学(Rad)特征。裁剪出最大二维矩形横截面感兴趣区(ROI),输入残差网络(ResNet50)深度神经网络模型行预训练,提取平均池化层的深度迁移学习(DTL)特征,并将Rad特征与DTL特征融合。以LASSO回归分别筛选出最优的Rad特征、DTL特征及二者融合(DLR)特征,选出最佳模型。将临床病理指标及MRI影像特征通过单-多因素Logistic回归分析筛选独立预测因素并参与构建联合模型,采用LR分类器构建其预测模型。采用受试者操作特征(ROC)曲线、校准曲线和决策曲线分析(DCA)评价模型预测效能、校准度及临床获益度。结果:Rad模型在训练集和验证集的AUC分别为0.686(95%CI:0.594~0.777)和0.671(95%CI:0.455~0.885),DTL模型在训练集和验证集的AUC分别为0.739(95%CI:0.657~0.821)和0.643(95%CI:0.409~0.876),DLR模型在训练集和验证集的AUC分别为0.886(95%CI:0.830~0.941)和0.730(95%CI:0.515~0.945)。结合临床病理及MRI影像特征构建联合模型在训练集和验证集的AUC分别为0.891(95%CI:0.835~0.946)、0.762(95%CI:0.579~0.944)。结论:基于DCE-MRI图像的DTL特征、Rad特征以及临床特征构建的联合模型能较好地预测乳腺癌患者SLN转移状态,而联合模型对其预测SLN转移状态有增益作用,并有望为乳腺癌患者的临床决策提供更有价值的信息。
Objective:To explore the value of radiomics(Rad)and deep transfer learning(DTL)features derived from dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI)combined with clinical features in predicting the sentinel lymph node(SLN)metastasis status in breast cancer patients.Methods:A retrospective analysis was conducted on MRI images and clinical data of patients with pathologically confirmed invasive breast cancer in our hospital from January 2020 to April 2023.The patients were randomly divided into a training set(145 cases:95 with negative SLN,50 with positive SLN)and a validation set(37 cases:28 with negative SLN,9 with positive SLN)at an 8:2 ratio.Tumor regions of interest(ROIs)were manually delineated layer by layer on the DCE-MRI images and radiomic(Rad)features were extracted.The largest two-dimensional(2D)rectangular cross-sectional ROI was cropped and input into the residual network(ResNet50)deep neural network model for pre-training.DTL features of the average pooling layer were extracted,and the Rad features were fused with the DTL features.The least absolute shrinkage and selection operator(LASSO)regression was used to screen the optimal Rad,DTL,and their fused(DLR)features,respectively,and the best model was selected.Clinicopathological indicators and MRI imaging features were subjected to univariate and multivariate Logistic regression analyses to screen independent factors,which were then used to construct the predictive model.Receiver operating characteristic(ROC)curve,calibration curve and decision curve analysis(DCA)were used to evaluate the model's predictive performance,calibration,and clinical benefits.Results:The AUC of Rad model in the training set and validation set were 0.686(95%CI:0.594~0.777)and 0.671(95%CI:0.455~0.885),respectively.The AUC of DTL model in the training set and validation set were 0.739(95%CI:0.657~0.821)and 0.643(95%CI:0.409~0.876),respectively.The AUC of DLR model in training set and validation set were 0.886(95%CI:0.830~0.941)and 0.730(95%CI:0.515~0.945),respectively.The AUC of the combined model constructed by combining clinicopathological and MRI imaging features in the training set and validation set were 0.891(95%CI:0.835~0.946)and 0.762(95%CI:0.579~0.944),respectively.Conclusion:The combined model constructed based on DTL features,Rad features derived from DCE-MRI images,and clinical features can effectively predict the SLN metastasis status in breast cancer patients.Moreover,the combined model provides added value for predicting SLN metastasis status and is expected to offer more valuable information for clinical decision-making in breast cancer patients.
作者
谢汉民
胡蝶
黄煌
程佳玲
谭昱
王思月
XIE Han-min;HU Die;HUANG Huang(Department of Radiology,Guangdong Women and Children Hospital,Guangzhou 511400,China)
出处
《放射学实践》
北大核心
2025年第11期1435-1442,共8页
Radiologic Practice
基金
广东省医学科研基金(A2024533)。
关键词
乳腺肿瘤
前哨淋巴结
磁共振成像
深度学习
Breast neoplasms
Sentinel lymph node
Magnetic resonance imaging
Deep learning
