乳腺钼靶影像组学结合深度学习预测乳腺癌HER-2三种表达状态的价值  

The Value of Mammography Radiomics Combinedwith Deep Learning in PredictingThree Expression States of HER-2 in Breast Cancer

作　　者：胡雪超 周永 刘文[2] 付国庆 马明瑞 张曦 HU Xuechao;ZHOU Yong;LIU Wen(The 3rd Affiliated Teaching Hospital of Xinjiang Medical University(Affiliated Cancer Hospital),Urumqi,Xinjiang Uygur Autonomous Region 830011,P.R.China)

机构地区：[1]新疆医科大学第三临床医学院(附属肿瘤医院),830011 [2]新疆工程学院,830023

出　　处：《临床放射学杂志》2025年第5期844-849,共6页Journal of Clinical Radiology

摘　　要：目的探讨利用深度学习和随机森林算法建立钼靶影像组学模型对乳腺癌人表皮生长因子受体-2(HER-2)三种表达状态的预测价值。方法回顾并分析2019年1月至2022年1月于新疆医科大学附属肿瘤医院行全数字化乳腺摄影(FFDM)检查并经病理学检查证实为乳腺癌的病例726例,以7∶3比例随机分配为训练集(n=508)与测试集(n=218)。每例患者FFDM影像包含头尾(CC)位和内外斜(MLO)位。通过深度学习提取726例基于FFDM检查病灶的全瘤组学特征,经降维和筛选后,将保留的特征放入随机森林(RF)机器学习模型,以受试者工作特征(ROC)曲线评价模型对乳腺癌三种HER-2表达状态的预测效能。结果经病理学检查证实的726例病灶中,HER-2零表达状态为175例,HER-2低表达状态病例为316例,HER-2高表达状态病例为235例。RF模型在训练集中预测乳腺癌HER-2的不同表达状态的AUC值分别为0.81、0.74和0.71,在测试集中的AUC值分别为0.88、0.74和0.76,影像组学模型在基于FFDM测试集中的准确度、灵敏度和特异度分别为74.00%、70.30%和79.02%。结论基于FFDM检查利用深度学习和RF算法建立影像组学模型对乳腺癌HER-2三种表达状态有较好的预测价值。Objective To explore the predictive value of a radiomics model based on deep learning and random forest algorithms for the three expression states of human epidermal growth factor receptor 2(HER-2)in breast cancer.Methods A total of 726 cases of breast cancer confirmed by pathology after full-field digital mammography(FFDM)examination at the Affiliated Cancer Hospital of Xinjiang Medical University from January 2019 to January 2022 were retrospectively analyzed and randomly assigned to the training set(n=508)and test set(n=218)at a ratio of 7∶3.Each case of FFDM images included craniocaudal(CC)and mediolateral oblique(MLO)views.A total of 726 radiomics features based on FFDM were extracted using deep learning.After dimensionality reduction and feature selection,the remaining features were input into the random forest(RF)machine learning model to evaluate the model's predictive power for the three HER-2 expression states in breast cancer using receiver operating characteristic(ROC)curves.Results Of the 726 lesions confirmed by pathological examination,175 cases were HER-2 zero expression status,316 cases were HER-2 low expression status,and 235 cases were HER-2 high expression status.The AUC values of the RF model in the training set for different HER-2 expression states were 0.81,0.74,and 0.71,respectively,and the AUC values in the test set were 0.88,0.74,and 0.76,respectively.The accuracy,sensitivity,and specificity of the radiomics model based on the FFDM test set were 74.00%,70.30%,and 79.02%,respectively.Conclusion Based on FFDM examination,the radiomics model established using deep learning and random forest algorithms has good predictive value for the three expression states of HER-2 in breast cancer.

关 键 词：乳腺癌 影像组学 全视野数字乳腺X线摄影 

分 类 号：R737.9[医药卫生—肿瘤]