超声联合MRI深度学习组学诺莫图鉴别三阴性乳腺癌的模型建立与验证  

The establishment and verification of a model for identifying TNBC based on the deep learning radiomics nomogram by combining ultrasound with MRI

作　　者：程辰[1] 李洪娥[2] 顾艳[3] 徐峰 王文荣[5] 赵红艳[1] CHENG Chen;LI Hong-e;GU Yan(.Department of Ultrasound,Lianyungang Hospital of Traditional Chinese Medicine,Jiangsu 222004,China)

机构地区：[1]连云港市中医院超声科,江苏222004 [2]连云港市第一人民医院超声科,江苏222000 [3]连云港市第一人民医院影像科,江苏222000 [4]宿迁市第一人民医院影像科,江苏223008 [5]连云港市妇幼保健院超声科,江苏222004

出　　处：《放射学实践》2025年第3期339-348,共10页Radiologic Practice

基　　金：连云港市卫生健康青年科技项目(QN202204);连云港市第六期"521高层次人才培养工程"科研项目(LYG06521202387)。

摘　　要：目的:构建超声联合MRI的深度学习组学诺莫图,以区分三阴性乳腺癌(TNBC)与非三阴性乳腺癌(NTNBC)并进行验证。方法:回顾性收集分析2018年1月-2020年9月在连云港市第一人民医院接受治疗的247例浸润性乳腺癌患者,其中2018年1月-2019年12月的176名乳腺癌患者为训练组(TNBC:55例,NTNBC:121例),2020年1月-2020年9月的71名乳腺癌患者为内部验证组(TNBC:19例,NTNBC:52例),另收集来自连云港市中医院及连云港市妇幼保健院62例浸润性乳腺癌患者作为外部验证集(TNBC:18例,NTNBC:44例)。整理分析所有患者的临床病理资料及影像学资料,采用开源软件ITK-SNAP 4.0.2对病灶进行分割,影像组学采用Pyradiomics软件进行组学特征提取;深度学习采用ResNet50作为卷积神经网络(CNN)学习框架,采用组内相关系数、最小绝对收缩和选择算子(LASSO)回归等进行特征降维和筛选,最后采用结果维度模型融合构建临床-深度学习组学诺莫图。使用受试者操作特征(ROC)曲线评价模型区分度,Delong检验比较各模型间曲线下面积(AUC)的差异,采用校准曲线评估实际观测与预测之间的一致性,采用决策曲线分析(DCA)评估模型的临床有效性。结果:超声联合MRI建立的临床-深度学习组学诺莫图模型在训练组中,AUC为0.923(95%CI:0.880~0.955);内部验证组中,AUC为0.989(95%CI:0.967~1.000);外部验证组中,AUC为0.941(95%CI:0.820~0.944)。所建诺莫图模型的校准预测曲线与标准曲线贴合较好,提示该模型在区分TNBC与NTNBC的预测概率与实际概率具有良好的一致性。决策曲线显示在风险阈值0.00~0.88时,采用临床-深度学习组学诺莫图对患者进行术前区分TNBC与NTNBC预测的临床净获益率最高。结论:超声联合MRI的临床-深度学习组学诺莫图对TNBC与NTNBC诊断效能显著优于其他模型,有较好的临床应用价值。Objective:The aims of this study were to establish a deep learning radiomics nomogram by combing ultrasound with MRI to discriminate TNBC and NTNBC,and to perform prospective validation.Methods:247 patients with invasive breast cancer treated in the First People's Hospital of Lianyungang from January 2018 to September 2020 were retrospectively analyzed.Among them,176 patients treated from January 2018 to December 2019 were put into the training group(with 55 TNBC patients and 121 NTNBC),and the rest 71 patients who received treatment from Jan to Sep 2020 were put into the internal verification group(with 19 TNBC and 52 NTNBC).In addition,62 patients with invasive breast cancer from Lianyungang Hospital of Traditional Chinese Medicine and Lianyungang Maternal and Child Health Hospital were collected as an external validation set(TNBC:18 cases,NTNBC:44 cases).The clinical pathological data and imaging data of all patients were collected and analyzed.The open source software ITK-SNAP 4.0.2(www.itksnap.org)was used for histological lesion segmentation.Radiomics features were extracted using Pyradiomics software(version 3.6.11).Deep learning of imaging:ResNet50 was used as the learning framework of Convolutional Neural Network(CNN).Intra-group correlation coefficient and least absolute shrinkage and selection operator(LASSO)regression were used for feature reduction and screening.The construction of clinical deep learning radiomics nomogram adopted the fusion of result-dimension model.The receiver operating characteristic curve was used to evaluate the discrimination of the model,the Delong test was used to compare the statistical differences between the area under curve(AUC)of each model,the calibration curve was used to evaluate the consistency between the actual observation and prediction,and the decision curve analysis(DCA)was used to evaluate the clinical effectiveness of the model.Results:In the training group,the AUC of the clinical-deep learning radiomics Nomogram model established by ultrasound combined with

关 键 词：超声 磁共振成像 深度学习影像组学诺莫图 乳腺肿瘤 三阴性乳腺癌 

分 类 号：R445.1[医药卫生—影像医学与核医学] R737.9[医药卫生—诊断学]