基于MRI影像组学的可解释性机器学习模型预测乳腺癌新辅助治疗病理完全缓解  

Interpretable Machine Learning Model Basedon MRI Radiomics for Predicting Pathological Complete Response to Neoadjuvant Therapy in Breast Cancer

作　　者：陈鑫[1] 陈小波 梁演婷 梁俊君 胡恒肖 王宽宏 魏新华[1] CHEN Xin;CHEN Xiaobo;LIANG Yanting(Department of Radiology,GuangzhouFirstPeople’sHospital,Guangzhou,Guangdong,510108,P.R.China)

机构地区：[1]广州市第一人民医院放射科,510108 [2]广东省人民医院放射科,510080

出　　处：《临床放射学杂志》2024年第10期1666-1672,共7页Journal of Clinical Radiology

基　　金：国家自然科学基金项目(编号:82072090,82371954);广州市科技计划项目资助(编号:202201020001,202201010513)。

摘　　要：目的探讨基于动态增强磁共振成像(DCE-MRI)肿瘤和瘤周影像组学联合临床病理特征构建模型预测乳腺癌新辅助治疗(NAT)后病理完全缓解的价值,并通过Shapley(SHapley Additive exPlanation)算法对模型进行可解释性分析。方法回顾性分析来自两家医院656例接受NAT治疗并进行手术切除的乳腺癌患者临床资料(训练组389例,内部验证组166例,外部验证组101例)。基于DCE-MRI肿瘤区域及瘤周5 mm区域提取并筛选影像组学特征,分别构建肿瘤模型、瘤周模型及肿瘤联合瘤周模型。通过单因素及多因素Logistic回归分析构建临床模型。最后,联合临床病理独立预测因子及肿瘤联合瘤周影像组学特征构建联合模型。采用受试者工作特征曲线评估模型效能。利用Shapley算法赋予最佳预测模型可解释性。结果基于肿瘤区域、瘤周区域分别筛选出16个及5个最佳影像组学特征并构建相应的模型,双区域特征联合筛选后保留了15个最佳影像组学特征构建肿瘤联合瘤周模型。临床T分期、HER2表达状态及分子分型均为预测乳腺癌NAT疗效的独立预测因子。联合模型在训练组、内部验证组及外部验证组的曲线下面积分别为0.849(95%CI:0.811~0.886)、0.819(95%CI:0.754~0.881)及0.864(95%CI:0.789~0.928),均高于临床模型、肿瘤模型、瘤周模型、肿瘤联合瘤周模型,差异均有统计学意义(P均<0.05)。Shapley分析显示放射学得分为模型最重要的特征。结论联合肿瘤、瘤周影像组学特征及临床病理信息所构建的临床影像联合模型可有效预测乳腺癌NAT治疗疗效。Shapley算法可提供个体水平的可解释性,保障模型的临床实用性。Objective This study aims to predict pathological complete response(pCR)to neoadjuvant therapy(NAT)in breast cancer using pre-treatmentDCE-MRI intratumoral and peritumoral radiomics features combined with clinical-pathological characteristics.Additionally,it visualizes and analyzes the model using the Shapley Additive Explanation(Shapley)algorithm.Methods Clinical data of 656 patientswith breast cancer who received NAT and surgery from two hospitals were retrospectively analyzed,including 389 patients in the training group,166 patients in the internal validation group,and 101 patients in the external validation group.Radiomics features were extracted and selected based on the volume of interest(VOI)from DCE-MRI intratumoral and peritumoral regions,and the intratumoral,peritumoral and intratumoral combined peritumoral model were constructed,respectively.A clinical model was built through univariate and multivariate Logistic regression analysis.Finally,the combined model was constructed by integrating clinicaland pathological independent predictors and intratumoral combined peritumoral radiomics features.The performance of the model was evaluated using receiver operating characteristic curves.The Shapley algorithm was employed to enhance model interpretability.Results Based on intratumoral and peritumoral VOI,the 16 and 5 best radiomics features were screened and the corresponding models were constructed respectively.After a combined screening of bi-regional radiomic features,15 optimal radiomic features were retained to construct the intratumoral combined peritumoral model.Clinical T-stage,HER2 status and molecular subtypes were identified as independent predictors for predicting NAT efficacy in breast cancer.The areas under the curve of the combined model in the training group,internal validation group and external validation group were 0.849(95%CI:0.811-0.886),0.819(95%CI:0.754-0.881)and 0.864(95%CI:0.789-0.928),respectively,which were higher than that of the clinical model,the intratumoral model,peritumoral mod

关 键 词：乳腺癌 新辅助治疗 磁共振成像 病理完全缓解 影像组学 

分 类 号：R445.2[医药卫生—影像医学与核医学] R737.9[医药卫生—诊断学]