超声影像组学联合临床指标预测乳腺癌新辅助化疗疗效的应用价值
The Application Value of Ultrasound Radiomics Combined with Clinical Indicators in Predicting the Efficacy of Neoadjuvant Chemotherapy for Breast Cancer
冯晓丹 1石岩 2杜瑶 1吴萌 1徐宇缘 1刘菲菲1
作者信息
- 1. 滨州医学院附属医院超声医学科 山东省滨州市,256603
- 2. 威海市立医院超声医学科 山东省威海市,264299
- 折叠
摘要
目的 比较基于超声影像组学特征联合临床病理特征构建的8种机器学习模型对乳腺癌新辅助化疗(NAC)后病理完全缓解(pCR)的预测价值,并探索最佳机器学习算法.方法 回顾性分析经病理证实为乳腺癌患者的化疗前超声图像,提取并降维筛选乳腺癌原发灶的超声影像组学特征,基于多因素Logistic回归分析、极端梯度提升、轻量梯度提升、随机森林、朴素贝叶斯、决策树、k近邻、支持向量机分类器,结合差异有统计学意义的临床病理信息和筛选出的最优组学特征,构建8种临床-影像组学联合模型.绘制受试者工作特征曲线比较各模型的预测性能.采用沙普利可加性模型解释方法(SHAP)赋予最佳预测模型可解释性.结果 在训练组和测试组中,随机森林模型的预测性能最优,其受试者工作特征曲线下面积(AUC)、准确度、特异度均高于其他模型.SHAP算法分析随机森林模型内各特征的决策权重,结果显示,临床特征人表皮生长因子受体-2、雌激素受体、细胞增殖抗原(Ki-67)及影像组学特征灰度游程矩阵特征(GLRLM-LRE)、一阶特征(lbp.2D_firstorder)是预测乳腺癌NAC后pCR的关键决策因素.结论 基于乳腺癌原发灶影像组学特征和临床特征构建的8种机器学习模型对乳腺癌NAC后pCR具有较高的预测效能,且以随机森林模型的表现能力较强,模型有望在NAC前准确预测患者pCR情况,为临床治疗决策提供有价值的参考.
Abstract
Objective To compare the predictive value of eight machine learning models based on ultrasound ra-diomics features combined with clinicopathological features in predicting pathological complete response(pCR)in breast cancer patients after neoadjuvant chemotherapy(NAC).Methods The pre-chemotherapy ultrasound images of patients with pathologically confirmed breast cancer were retrospectively analyzed,and the ultrasound radiomics fea-tures of the primary breast cancer lesions were extracted and screened for dimensionality reduction.Eight clinical-radiomics combined models were constructed based on multivariate Logistic regression analysis,XGBoost,LightGBM,random forest,naive Bayesian,decision tree,k-nearest neighbor and support vector machine classifiers,combined with the clinicopathological information with statistically significant differences of patients and the selected optimal radiomics features.Receiver operating characteristic(ROC)curves were drawn to compare the predictive per-formance of each model.The Shapley Additive exPlanation(SHAP)algorithm was used to give the best prediction model interpretability.Results In the training group and the test group,the random forest model had the best predic-tion performance,and its area under the receiver operating characteristic curve,accuracy and specificity were higher than those of other models.SHAP algorithm was used to analyze the decision weight of each feature in the random forest model,and the results showed that the clinical features human epidermal growth factor receptor-2,estrogen re-ceptor,cell proliferation antigen(Ki-67)and the radiomics features original-glrlm_LongRunEmphasis,lbp.2D_fir-storder were the key decision factors for predicting pCR after NAC in breast cancer.Conclusions Eight machine learning models based on radiomics and clinical features of primary breast cancer have high prediction efficiency for pCR after NAC,and the random forest model has the higher performance.These models are expected to accurately predict pCR before NAC and provide valuable reference for clinical treatment decisions.
关键词
超声检查/乳腺肿瘤/新辅助化疗/影像组学Key words
Ultrasonography/Breast cancer/Neoadjuvant chemotherapy/Radiomics引用本文复制引用
出版年
2025
NETL
NSTL
万方数据