分享:
分享到微信朋友圈
X
临床研究
基于双序列的可解释性机器学习模型术前预测浸润性乳腺癌luminal和非luminal分型的价值
张涛 周鹏 王俊 庞志斌 胡云涛

本文引用格式:张涛, 周鹏, 王俊, 等. 基于双序列的可解释性机器学习模型术前预测浸润性乳腺癌luminal和非luminal分型的价值[J]. 磁共振成像, 2025, 16(11): 114-122. DOI:10.12015/issn.1674-8034.2025.11.017.


[摘要] 目的 探讨基于动态对比增强(dynamic contrast enhanced, DCE)及高分辨率延迟期图像的夏普利加性解释(SHapley Additive exPlanations, SHAP)机器学习模型术前预测浸润性乳腺癌luminal和非luminal分型的价值。材料与方法 回顾性收集182例经病理证实为非特殊型浸润性乳腺癌患者的临床-病理-影像资料,依据病理结果分为luminal组(121例)和非luminal组(61例)。利用3D slicer软件在浸润性乳腺癌患者的DCE及高分辨率延迟期乳腺MRI影像上,勾画病灶边缘并提取影像组学特征。按7∶3比例随机分为训练集和测试集。采用单因素t检验或者U检验、最小绝对收缩和选择算子(least absolute shrinkage and selection operator, LASSO)筛选特征。通过logistic、支持向量机(support vector machine, SVM)、AdaBoost算法分别建立临床模型、影像组学模型、联合模型。并通过受试者工作特征曲线下面积(area under the curve, AUC)、准确率、敏感度、特异度评估模型的效能,模型预测效能的比较采用DeLong检验。通过SHAP分析可视化特征在模型中的贡献。结果 组织学分级、糖类抗原-125在两组之间差异有统计学意义(P<0.05)。经过降维后DCE和高分辨率延迟期图像分别剩余2个和4个最佳影像组学特征。基于DCE特征-高分辨率延迟期特征-临床特征的logistic、SVM、AdaBoost联合模型效能较好,在训练集的AUC分别为0.854、0.853、0.962,准确率分别为:71.8%、75.1%、89.4%,敏感度分别为:74.0%、77.3%、85.1%,特异度分别为:69.7%、72.9%、93.6%;在测试集的AUC分别为0.828、0.836、0.802,准确率分别为72.5%、76.3%、72.5%,敏感度分别为:74.1%、77.0%、71.8%,特异度分别为67.5%、74.5%、73.5%。logistic、AdaBoost的联合模型在训练集及测试集之间差异有统计学意义(P值分别为:P=0.044,P<0.001)。SVM的联合模型在训练集和测试集之间差异无统计学意义(P=0.277)。在测试集,SVM的联合模型优于SVM的临床模型,两者之间差异有统计学意义(P<0.001)。结论 可解释性机器学习模型可术前预测浸润性乳腺癌luminal和非luminal分型,对患者制订个性化治疗方案和预后评估有着重要的临床应用价值。
[Abstract] Objective To explore the value of a SHapley Additive exPlanations (SHAP) machine learning model based on dynamic contrast-enhanced (DCE) and high-resolution delayed phase images for the preoperative prediction of luminal and non-luminal subtypes of invasive breast cancer.Materials and Methods Clinical, pathological, and imaging data of 182 patients with pathologically confirmed invasive breast carcinoma of no special type were retrospectively collected and divided into a luminal group (121 cases) and a non-luminal group (61 cases) based on pathological results. Using 3D Slicer software, lesion margins were delineated on DCE and high-resolution delayed phase breast MRI images of invasive breast cancer patients, and radiomic features were extracted. Patients were randomly split into training and test sets in a 7∶3 ratio. Univariate t-test or Mann-Whitney U test and least absolute shrinkage and selection operator (LASSO) were used for feature selection. Clinical models, radiomics models, and combined models were built using logistic regression, support vector machine (SVM), and AdaBoost algorithms, respectively. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Model performance comparisons were conducted using DeLong's test. SHAP analysis was used to visualize feature contributions in the models.Results There were statistically significant differences in histological grade and carbohydrate antigen-125 between the two groups, with P < 0.05. After dimensionality reduction, 2 and 4 optimal radiomics features were respectively retained for DCE and high-resolution delayed-phase images. The combined models of logistic, SVM, and AdaBoost based on DCE features, high-resolution delayed-phase features, and clinical features had better performance. The AUCs in the training set were 0.854, 0.853, and 0.962, respectively, with accuracies of 71.8%, 75.1%, and 89.4%, sensitivities of 74.0%, 77.3%, and 85.1%, and specificities of 69.7%, 72.9%, and 93.6%, respectively. The AUCs in the test set were 0.828, 0.836, and 0.802, respectively, with accuracies of 72.5%, 76.3%, and 72.5%, sensitivities of 74.1%, 77.0%, and 71.8%, and specificities of 67.5%, 74.5%, and 73.5%, respectively. The combined models of logistic and AdaBoost had statistically significant differences between the training set and the test set (P = 0.044, P < 0.001). The combined model of SVM had no statistically significant difference between the training set and the test set (P = 0.277). In the test set, the combined model of SVM was superior to the clinical model of SVM, and the difference was statistically significant (P < 0.001).Conclusions Interpretable machine learning models can preoperatively predict luminal and non-luminal subtypes of invasive breast cancer, holding significant clinical value for formulating personalized treatment plans and prognostic assessments for patients.
[关键词] 磁共振成像;机器学习;夏普利加性解释;乳腺癌;luminal
[Keywords] magnetic resonance imaging;machine learning;SHapley Additive exPlanations;breast cancer;luminal

张涛    周鹏    王俊    庞志斌    胡云涛 *  

电子科技大学附属肿瘤医院四川省肿瘤医院放射科,成都 610000

通信作者:胡云涛,E-mail:15298217550@163.com

作者贡献声明:胡云涛设计本研究的方案,对稿件重要内容进行了修改;张涛起草和撰写稿件,获取、分析、解释本研究的数据;王俊、庞志斌获取、分析或解释本研究的数据,对稿件重要内容进行了修改;周鹏参与设计本研究方案,并对稿件重要内容进行了修改,获得了国家自然科学基金项目资助;全体作者都同意发表最后的修改稿,同意对本研究的所有方面负责,确保本研究的准确性和诚信。


基金项目: 国家自然科学基金项目 U21A20521
收稿日期:2025-08-02
接受日期:2025-11-03
中图分类号:R445.2  R737.9 
文献标识码:A
DOI: 10.12015/issn.1674-8034.2025.11.017
本文引用格式:张涛, 周鹏, 王俊, 等. 基于双序列的可解释性机器学习模型术前预测浸润性乳腺癌luminal和非luminal分型的价值[J]. 磁共振成像, 2025, 16(11): 114-122. DOI:10.12015/issn.1674-8034.2025.11.017.

0 引言

       据世界卫生组织国际癌症研究机构最新发布数据显示,乳腺癌在女性群体中的发病率逐年上升[1, 2]。目前,临床上主要采用免疫组织化学和荧光原位杂交法明确乳腺癌的细胞受体因子表达来进行分子分型,其中根据ER、PR、HER-2、Ki-67的表达不同可分为luminal型和非luminal型。由于分子分型的不同,其对治疗的预后也不同。对于luminal型常以内分泌治疗为基石,且预后也较好[3]。非luminal型中,HER2阳性型需靶向治疗联合化疗,且预后已显著改善,而三阴性型以化疗为主,预后相对较差[4, 5, 6]。综上所述术前明确乳腺癌的分子分型对乳腺癌的治疗及预后预测有着重要的意义。

       影像组学是一种通过提取高通量图像特征以非侵入性方法反映肿瘤内在异质性的工具[7]。随着影像组学在乳腺癌领域研究的深入,多模态乳腺磁共振成像为术前无创预测乳腺癌分子分型提供了新的途径[8, 9, 10]。目前,虽然已有大量研究基于影像组学与机器学习方法对luminal型和非luminal型乳腺癌进行区分,但其模型构建多依赖于常规影像序列和基础临床指标,且多数局限于模型性能的验证,缺乏对特征贡献机制的可解释性探讨[11, 12, 13]。TOUMAJ等[14]的Meta分析结果显示在癌症检测系统中,约有44.4%的研究采用夏普利加性解释(SHapley Additive exPlanations, SHAP)框架作为可解释性工具,其优势在于能量化每个影像特征对诊断结果的贡献度,这对多模态影像融合分析尤为重要。这种广泛采纳表明SHAP正在成为影像组学研究中解释模型决策的重要标准工具。尽管SHAP解释方法已在部分影像组学研究中有初步应用[15],但其在融合多期相磁共振图像(如动态对比增强与高分辨率延迟期)并结合临床特征进行luminal分型预测中的系统性应用尚未见报道。此外,既往研究往往未能充分整合多期相影像特征与临床信息之间的互补价值[16],亦缺乏对预测过程的可视化与生物学合理性解释,限制了其临床转化潜力。因此,本研究基于DCE与高分辨率延迟期磁共振图像,结合SHAP可解释性分析,构建融合影像组学与临床特征的机器学习模型,通过可视化关键预测特征并阐释其决策机制,旨在探索该模型在术前无创预测浸润性乳腺癌luminal与非luminal分型中的应用价值,以期为临床提供一种精准的分子分型预测工具,辅助个体化治疗决策,推动乳腺癌精准诊疗的发展。

1 材料与方法

1.1 研究对象

       本研究遵守《赫尔辛基宣言》,通过四川省肿瘤医院伦理委员会批准(批准文号:SCCHEC-02-2024-170),免除受试者知情同意。本研究回顾性纳入182例经病理证实为非特殊型浸润性乳腺癌患者,并依据病理结果分为luminal组(121例)和非luminal组(61例)。以7∶3比例随机分为训练集和测试集。纳入标准:(1)MRI检查前未做穿刺及任何治疗;(2)病理结果为MRI检查后,术前超声引导下穿刺获得;(3)所有图像均于术前两周内采集。排除标准:(1)患者临床资料不完整者;(2)图像质量不佳或感兴趣区无法进行勾画者。

1.2 扫描参数

       采用西门子Vida 3.0 T磁共振扫描仪及16通道乳腺专用线圈扫描,扫描采用俯卧位头先进,双侧乳房悬垂于乳腺线圈中央。扫描范围包全两侧整个乳腺。扫描序列及参数如下:

       轴位DCE序列,扫描参数:TR 3.72 ms,TE 1.37 ms,层厚2.5 mm,层间距0.5 mm,视野360 mm×360 mm,每期扫描16 s,一共40期,从扫描的第二期开始注射对比剂。轴位T1WI高分辨率延迟扫描序列,扫描参数:TR 9.00 ms,TE 4.37 ms,视野360 mm×360 mm,层厚0.8 mm,层间距0.8 mm×20%。增强扫描对比剂选用钆喷酸葡胺(北京北陆药业有限公司,中国)。采用高压注射器(安特,ANT065115,深圳安特医疗股份有限公司,中国)以2.5 mL/s速率和0.1 mmol/kg注射剂量进行注射,再以相同速率追加15 mL生理盐水冲管。延迟期扫描均于注射对比剂10 min后开始。

1.3 病理特征

       参考《中国抗癌协会乳腺癌诊治指南与规范(2024版)》,根据ER、PR、HER-2、Ki-67的表达水平分为luminal组(121例)和非luminal组(61例):(1)luminal组包括luminal A型[ER(+)和/或PR(+),HER-2(-),Ki-67低表达(<20%)]和luminal B型[ER(+)和/或PR(+),HER-2(+),Ki-67高表达(≥20%)];(2)非luminal 组包括HER-2过表达型[HER-2过表达,ER和PR均为(-),Ki-67任意表达]和三阴性型[ER、PR和HER-2均为(-),Ki-67任意表达][17, 18]

1.4 特征提取及降维

       所有患者的靶区勾画分别由两位具有5年诊断经验的放射科主治医师采用3Dslicer软件(版本号5.8.1,National Alliance for Medical Imaging Computing,美国)逐层勾画整个病灶体积,勾画时避开水肿区域、血管及正常组织(图1),在特征提取前先通过重采样,X、Y、Z参数分别为1,1,1进行标准化,并对靶区提取1223个影像组学特征。采用组内相关系数(intra-class correlation coefficient, ICC)比较两名医师提取特征的一致性,ICC≥0.75的特征被保留。通过单因素t检验或曼-惠特尼U检验进行第一次降维,再采用十折交叉验证的最小绝对收缩和选择算子(least absolute shrinkage and selection operator, LASSO)进行第二次降维。

图1  靶区勾画图。女,56岁,非特殊型浸润性乳腺癌。1A:患者高分辨率延迟期原始病灶图;1B:高分辨率延迟期靶区勾画图。1C:患者动态对比增强原始病灶图;1D:动态对比增强图靶区勾画图。人工逐层勾画病灶(蓝箭)边缘,勾画时避开水肿区域、血管及正常组织。
Fig. 1  Target volume delineation diagram. Female, 56 years old, invasive breast carcinoma of no special type. 1A: Original lesion image from the patient's high-resolution delayed phase; 1B: Target volume delineation on the high-resolution delayed phase image. 1C: Original lesion image from the patient's dynamic contrast-enhanced scan; 1D: Target volume delineation on the dynamic contrast-enhanced image. The lesion (blue arrows) margins are manually contoured slice-by-slice, avoiding areas of edema, blood vessels, and normal tissue during the delineation process.

1.5 模型建立

       采用十折交叉验证,构建了基于支持向量机(support vector machine, SVM)、logistic回归和AdaBoost算法的以下模型:DCE影像组学模型、高分辨率延迟期影像组学模型、DCE-高分辨率延迟期影像组学模型、临床模型,以及DCE特征-高分辨率延迟期特征-临床特征的联合模型。通过准确率、受试者工作特征曲线下面积(area under the curve, AUC)、敏感度、特异度来评估模型性能,模型之间的比较采用DeLong检验,通过SHAP分析将联合模型进行可视化。

1.6 样本量估计算

       采用机器学习模型特征经验法则[19, 20],样本量估算公式为:samplesize=n×10/(1-p),其中n为特征筛选后剩余的特征数,p代表模型预期的分类错误率。由于本研究是二分类问题,随机猜测的错误率为50%。因此本研究模型p取值0.5[21]

1.7 统计学分析

       本研究采用R语言4.4.0版本、python版本3.9.0对数据进行统计学分析。符合正态分布的定量资料采用两个独立样本的t检验,否则采用U检验。采用χ2检验或Fisher's检验分析计数资料,其中对于理论数T<5但T≥1,并且n≥40,用连续性校正的卡方进行检验,理论数T<1或n<40,则用Fisher's检验。P<0.05差异有统计学意义。

2 结果

2.1 样本量计算结果

       根据本研究通过LASSO降维后剩余的特征数估算样本量,高分辨率延迟期图像剩余4个影像组学特征,DCE图像剩余2个影像组学特征。为使估算样本量尽可能大,n取值为高分辨率延迟期剩余特征数+DCE剩余特征数,模型精确要求最低50%,代入公式估算样本量至少为120例。

2.2 一般资料

       在luminal组和非luminal 组中年龄分别为(51.00±9.81)和(51.20±9.98)岁。年龄、腋窝淋巴结转移、月经状态在两组之间差异无统计学意义(P>0.05)。组织学分级、糖类抗原-125在两组之间差异有统计学意义(P<0.05)(表1)。

表1  两组患者的一般资料
Tab. 1  General information of the two groups of patients

2.3 临床模型建立

       基于组织学分级的logistic、SVM、AdaBoost临床特征模型,在训练集的AUC均为0.597,敏感度均为79.2%,特异度均为40.1%,准确率均为59.7%;在测试集的AUC均为0.588,敏感度均为79.2%,特异度均为38.5%,准确率均为57.8%(表2)。

表2  不同算法的临床特征模型结果
Tab. 2  Results of clinical feature models of different algorithms

2.4 动态对比增强影像组学模型建立

       经过降维后,DCE图像剩余2个影像组学特征(图2)。基于DCE的影像组学特征logistic、SVM、AdaBoost模型在训练集的AUC分别为0.678、0.678、0.771,敏感度分别为62.0%、59.3%、66.0%,特异度分别为65.0%、68.1%、73.5%,准确率分别为63.5%、63.7%、69.7%;在测试集的AUC分别为0.676、0.665、0.663,敏感度分别为61.6%、59.1%、58.9%,特异度分别为64.2%、64.2%、64.7%,准确率分别为62.6%、61.0%、60.9%(表3)。

图2  动态对比增强影像组学特征的特征选择图。2A:图中曲线代表特征的变化轨迹,黑色垂直线定义了λ的最佳值为0.093。2B:左侧和右侧的虚线分别代表二项式偏差最小模型原则和距离二项式偏差最小值的一个标准差范围内的最简洁模型原则,本研究中采用最小模型原则。
Fig. 2  Feature selection diagram for dynamic contrast-enhanced radiomics features. 2A: The curve in the diagram represents the trajectory of feature changes, and the black vertical line defines the optimal value of λ as 0.093; 2B: The dashed lines on the left and right represent the principle of the minimal binomial deviation model and the principle of the simplest model within one standard deviation of the minimal binomial deviation, respectively. In this study, the minimal model principle is adopted.
表3  不同算法的动态对比增强影像组学模型结果
Tab. 3  Results of dynamic contrast-enhanced radiomics models of different algorithms

2.5 高分辨率延迟期影像组学模型建立

       经过降维后,高分辨率延迟期剩余4个影像组学特征(图3)。基于高分辨率延迟期的影像组学特征的logistic、SVM、AdaBoost模型在训练集的AUC分别为0.717、0.720、0.870,敏感度分别为68.0%、68.6%、70.7%,特异度分别为63.5%、62.8%、85.9%,准确率分别为65.7%、65.7%、78.3%;在测试集的AUC分别为0.696、0.707、0.653,敏感度分别为66.8%、67.6%、57.9%,特异度分别为65.2%、61.9%、64.2%,准确率分别为66.3%、65.8%、59.9%(表4)。

图3  高分辨率延迟期影像组学特征的特征选择图。3A:图中曲线代表特征的变化轨迹,黑色垂直线定义了λ 的最佳值为0.088。3B:左侧和右侧的虚线分别代表二项式偏差最小模型原则和距离二项式偏差最小值的一个标准差范围内的最简洁模型原则,本研究中采用最小模型原则。
Fig. 3  Feature selection map of high-resolution delayed-period radiomics features. 3A: the curve represents the trajectory of feature changes, with the black vertical line indicating the optimal λ value of 0.088. 3B: the dashed lines on the left and right correspond to the principle of the minimum model under binomial deviance and the principle of the most parsimonious model within one standard deviation of the minimum binomial deviance, respectively; this study adopted the minimum model principle.
表4  不同算法的高分辨率延迟期影像组学模型结果
Tab. 4  Results of high-resolution delayed-phase radiomics models of different algorithms

2.6 DCE-高分辨率延迟期影像组学模型建立

       基于DCE-高分辨率延迟期的影像组学特征logistic、SVM、AdaBoost模型在训练集的AUC分别为0.780、0.783、0.919,敏感度分别为69.5%、71.0%、79.8%,特异度分别为65.3%、64.2%、85.9%,准确率分别为67.4%、67.6%、82.8%;在测试集的AUC分别为0.756、0.753、0.714,敏感度分别为67.7%、66.9%、70.1%,特异度分别为59.0%、54.2%、57.6%,准确率64.8%、62.6%、65.9%(表5)。

表5  不同算法的DCE-高分辨率延迟期影像组学模型结果
Tab. 5  Results of DCE- high-resolution delay radiomics models of different algorithms

2.7 DCE特征-高分辨率延迟期特征-临床特征的联合模型建立

       基于DCE特征-高分辨率延迟期特征-临床特征的联合模型logistic、SVM、AdaBoost在训练集的AUC分别为0.854、0.853、0.962,敏感度分别为74.0%、77.3%、85.1%,特异度分别为69.7%、72.9%、93.6%,准确率分别为71.8%、75.1%、89.4%;在测试集的AUC分别为0.828、0.836、0.802,敏感度分别为74.1%、77.0%、71.8%,特异度分别为67.5%、74.5%、73.5%,准确率分别为72.5%、76.3%、72.5%(表6)。

表6  不同算法的DCE特征-高分辨率延迟期特征-临床特征的联合模型结果
Tab. 6  Results of the combined model of DCE features - high-resolution delay period features - clinical features of different algorithms

2.8 模型比较及可视化

       在测试集logistic的联合模型与SVM、AdaBoost的联合模型之间差异无统计学意义(P值分别为:P=0.086,P=0.0965)。logistic、AdaBoost的联合模型存在过拟合(图45),训练集及测试集之间差异有统计学意义(P值分别为:P=0.044,P<0.01)。SVM的联合模型在训练集和测试集之间差异无统计学意义(P=0.277)(图6)。在测试集,SVM的联合模型优于SVM的临床模型,两者之间差异有统计学意义(P<0.01)。

       三种联合模型中SVM的校正曲线拟合度较好(图7)。决策曲线显示当阈值在0.3左右时,联合模型的净收益大于临床模型(图8)。根据SHAP值排序显示,联合模型中联合平均特征对模型的贡献度最大,其次是最大相关系数特征、组织学分级、总体能量特征、高灰度值强调特征、峰度特征、小依赖低灰度强调特征(图9)。

图4  logistic联合模型在训练集和测试集的AUC曲线随迭代次数变化图。AUC:曲线下面积。
Fig. 4  The AUC curves of the logistic combined model on the training set and the test set change with the number of iterations. AUC: area under the curve.
图5  AdaBoost联合模型在训练集和测试集的AUC曲线随迭代次数变化图。AUC:曲线下面积。
Fig. 5  The AUC curves of the AdaBoost combined model on the training set and the test set vary with the number of iterations. AUC: area under the curve.
图6  SVM联合模型在训练集和测试集的AUC曲线随迭代次数变化图。AUC:曲线下面积。
Fig. 6  The AUC curves of the SVM combined model on the training set and the test set change with the number of iterations. AUC: area under the curve.
图7  基于DCE特征-高分辨率延迟期特征-临床特征的支持向量机联合模型校正曲线。蓝色实线代表训练集校正曲线,红色实线代表测试集校正曲线。黑色虚线代表理想线。蓝色和红色实线越接近红色虚线代表模型的预测概率与实际发生概率相匹配。DCE:动态对比增强。
Fig. 7  The calibration curve of the support vector machine combined model based on DCE features, high-resolution delayed phase features, and clinical characteristics. The solid blue line represents the calibration curve of the training set, while the solid red line represents the calibration curve of the test set. The black dashed line represents the ideal line. The closer the blue and red solid lines are to the black dashed line, the better the model's predicted probabilities match the actual occurrence probabilities. DCE: dynamic contrast enhancement.
图8  SVM联合模型的决策曲线。当纵坐标为净收益,横坐标为阈值概率。当阈值在0.3左右时,联合模型获得的净收益大于临床模型。SVM:支持向量机。
Fig. 8  Decision Curve of the SVM combined model. When the vertical axis represents the net income and the horizontal axis represents the threshold probability. When the threshold is around 0.3, the net benefit obtained by the combined model is greater than that of the clinical model. SVM: support vector machine.
图9  SHAP图。9A:根据SHAP值排序的特征重要性。9B:根据SHAP值特征贡献度的热图。SHAP:夏普利加性解释。
Fig. 9  SHAP plot. 9A: Feature importance ranked according to SHAP values. 9B: Heatmap of feature contributions based on SHAP values. SHAP: SHapley Additive exPlanation.

3 讨论

       本研究基于乳腺癌临床特征、DCE及高分辨率延迟期图像上,建立了15种不同的预测模型用于术前评估浸润性乳腺癌的luminal型和非luminal型,并对临床特征模型、DCE影像组学特征模型、高分辨率延迟期影像组学特征模型、基于DCE-高分辨率延迟期的影像组学特征模型、基于DCE特征-高分辨率延迟期特征-临床特征的联合模型进行了比较,联合模型在区分luminal型和非luminal型效果较好,且本研究通过SHAP分析将模型进行了可视化,这将有助于临床对浸润性乳腺癌患者进行个性化治疗。

3.1 本研究与既往研究的差异性及创新性

       既往研究,如王世科等[22]在42例乳腺癌患者的DCE影像组学特征上建立了4种不同的模型,发现基于随机森林的影像组学模型区分luminal型和非luminal型的准确率和AUC分别为85.3%和0.876,此研究在小样本的基础上论证了影像组学在乳腺癌分子分型中的应用。胥豪等[23]通过将DCE的影像组学特征与血细胞参数结合建立预测luminal型和非luminal型的模型,发现影像组学-血细胞联合模型、独立的影像组学模型均优于血细胞参数模型。HUANG等[24]从DCE的第二期图像提取影像组学特征建立预测luminal型和非luminal型的影像组学模型,其在训练组及验证组的效能较好,AUC分别为0.86和0.80。上述研究均证实了基于DCE-MRI的影像组学在乳腺癌分子分型中具有巨大潜力,但它们大多局限于单一时间序列或单一模态的特征分析。本研究的主要创新点在于,首次整合了DCE序列的动力学信息与高分辨率延迟期序列的精细形态学信息,构建了双序列联合模型[25, 26]

       此外,程卫群等[27]将T2WI、DWI、DCE图像与生境分析相结合,从原始靶区上分割出不同的亚区,并从不同亚区中提取影像组学特征建立生境影像组学联合模型,结果发现生境分析的SVM模型区分luminal型和非luminal型效能最佳。张丁懿等[28]探讨了扩散加权成像(diffusion weighted imaging, DWI)模型、DCE模型、DWI-DCE联合模型区分luminal型和非luminal型效能,结果联合模型预测效能高达0.821。与既往研究相比,本研究引入了具备超高空间分辨率及各向同性三维成像能力的高分辨率延迟期序列。该序列凭借其超薄层厚的技术特点,在捕捉肿瘤空间解剖结构方面展现出显著优势,能够清晰显示内部细微结构(如微小分隔、钙化点),并精确刻画边缘特征(如毛刺征、分叶状轮廓)。这些精细的形态学信息,与DCE图像形成了有力互补。本研究结果显示基于SVM的DCE特征-高分辨率延迟期特征-临床特征联合模型取得了较好的预测效能,在测试集AUC值为0.836。表明通过融合不同增强时期的信息,可更全面地捕捉肿瘤异质性,从而提高分子分型预测的准确率。这一发现与张丁懿等[28]提出的多序列联合思路一致,但在序列选择、特征融合方法及模型构建方面体现出差异性与创新性。

3.2 主要研究成果分析

       既往研究结果发现非luminal型的三阴性乳腺癌中,组织学分级3级占比最多[29, 30],组织学分级结果可能是潜在初步判断乳腺患者luminal分型的因素之一。而本研究也得到了类似的结果,在非luminal型中高级别占比42.6%,相对于luminal型较高。糖类抗原-125主要用于卵巢癌的诊断和监测,在乳腺癌中并没有特异性和敏感性。但在本研究中糖类抗原-125在luminal型和非luminal型之间差异有统计学意义,导致以上研究结果的原因可能与本研究的样本量有关。所以本研究的糖类抗原-125不具备太大实际意义,只能作为辅助性参考,故本研究未将糖类抗原-125纳入模型。而本研究中基于组织学分级的三种临床特征模型预测效能较差,在训练集和测试集的AUC均分别为0.597、0.588。导致以上结果的原因可能是样本量较小,得到有统计学意义的临床特征较少,而特征与结局的关系很简单,SVM和AdaBoost模型可能无法学习到有效模式,在训练集和测试集上表现不佳。

       本研究构建并比较了基于不同特征组合(临床特征、DCE影像组学特征、高分辨率延迟期影像组学特征及其联合特征)与多种机器学习算法(logistic回归、SVM、AdaBoost)的15个预测模型,旨在区分乳腺肿瘤的luminal与非luminal亚型。结果表明,基于DCE特征-高分辨率延迟期特征-临床特征的联合模型在测试集上整体表现最优,其中logistic回归与SVM模型的AUC分别达到0.828和0.836,显示出良好的泛化能力。然而,AdaBoost算法虽然在训练集上取得了极高的性能(AUC为0.962,准确率为89.4%),但其测试集性能出现明显下降(AUC为0.802,准确率为72.5%),表明该模型存在过拟合现象。AdaBoost是一种迭代集成算法,通过序列化训练多个弱分类器,并在每一轮迭代中增加被误分类样本的权重,使模型后续更关注难以学习的样本。在有限数据集条件下,尤其是当某一类别(如非luminal型)样本量较少时,该机制容易导致模型过度拟合训练数据中的噪声和异常值。其原因是算法以降低训练误差为优化目标,可能过度适应个别样本的特定模式。因此,模型在训练集上表现优异,但在测试集上泛化性能显著下降,反映出其存在过拟合问题。

       此外,本研究中DCE的影像组学特征经过降维后,剩余总体能量、高灰度值强调两个影像组学特征,其中总体能量用于量化图像中感兴趣区内体素强度的总体大小。高总体能量可能提示肿瘤内细胞密度高或坏死区域少,低总体能量可能与囊变、坏死或低活性组织相关。高的灰度值强调可能提示肿瘤内存在高代谢、高增殖或坏死区域[31, 32, 33]。高分辨率延迟期图像经过降维后剩余峰度、联合平均、小依赖低灰度强调、最大相关系数。其中峰度、小依赖低灰度强调主要用于反映肿瘤或组织的微观结构异质性。联合平均用于量化图像中相邻像素对的灰度值联合分布特性,反映肿瘤或组织的局部均匀性。最大相关系数用于量化图像中灰度值之间的线性相关性,反映纹理模式的规律性[34, 35, 36]

       为进一步增强DCE特征-高分辨率延迟期特征-临床特征联合模型的可解释性,本研究采用SHAP分析进行可视化及贡献度评估[37, 38]。SHAP分析显示,联合平均值、最大相关系数特征值越低对模型正向输出的影响越大,提示联合平均值、最大相关系数特征值可能与肿瘤内部微观结构的无序性相关。峰度特征、小依赖低灰度强调特征对模型输出的影响较小,这可能提示luminal型和非luminal型的肿瘤在整体密度分布和低灰度区域特性上相似。总体能量、高灰度值强调值越高对模型负向输出影响越大,这可能提示在影像上看起来纹理均匀、规则,并且在增强扫描中呈现高强化的肿块,更有可能是一个生物学行为相对“温和”的luminal型乳腺癌。

3.3 本研究的局限性

       (1)本研究为单中心数据,结果可能会存在部分偏倚;(2)非luminal分型的病例较少,并缺乏外部验证;(3)本研究仅提取了DCE及高分辨率延迟期的影像组学特征,未将T2WI、DWI序列纳入研究;(4)未提取瘤周特征进行比较。因此,未来的研究中还需纳入平扫图像及瘤周影像组学特征进行建模,并在此基础上应用外部数据来进行验证。

4 结论

       综上所述,基于可解释性机器学习模型可以术前预测浸润性乳腺癌luminal和非luminal分型,未来这种非侵入性方法,有望成为临床无创鉴别乳腺癌分子分型的新方法,对患者制订个性化治疗方案和预后评估都有着重要的临床应用价值。

[1]
殷雨来, 江昕, 何晓阳, 等. 1990—2021年中国和全球乳腺癌疾病负担及未来15年预测[J]. 中华肿瘤防治杂志, 2025, 32(5): 275-283. DOI: 10.16073/j.cnki.cjcpt.2025.05.03.
YIN Y L, JIANG X, HE X Y, et al. Disease burden of breast cancer in China and the world from 1990 to 2021 and projections for the next 15 years[J]. Chin J Cancer Prev Treat, 2025, 32(5): 275-283. DOI: 10.16073/j.cnki.cjcpt.2025.05.03.
[2]
AUNE D, MARKOZANNES G, ABAR L, et al. Physical activity and health-related quality of life in women with breast cancer: a meta-analysis[J/OL]. JNCI Cancer Spectr, 2022, 6(6): pkac072 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/36474321/. DOI: 10.1093/jncics/pkac072.
[3]
NIU Z P, WU J Y, ZHAO Q C, et al. CAR-based immunotherapy for breast cancer: peculiarities, ongoing investigations, and future strategies[J/OL]. Front Immunol, 2024, 15: 1385571 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/38680498/. DOI: 10.3389/fimmu.2024.1385571.
[4]
LI J, WU Y, LI Y P, et al. Glutathione-disrupting nanotherapeutics potentiate ferroptosis for treating luminal androgen receptor-positive triple-negative breast cancer[J]. ACS Nano, 2024, 18(39): 26585-26599. DOI: 10.1021/acsnano.4c04322.
[5]
HAN J M, ZHAI X H, TAO X F, et al. Pharmacovigilance study of adverse reactions of anti-HER-2 drugs for the treatment of HER-2-positive breast cancer based on the FAERS database[J/OL]. Breast Cancer Res, 2025, 27(1): 54 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/40205546/. DOI: 10.1186/s13058-025-02013-w.
[6]
LI X Y, QIAN X H, ZHU J, et al. Synthesis and evaluation of novel HER-2 inhibitors to exert anti-breast cancer ability through epithelial-mesenchymal transition (EMT) pathway[J/OL]. Eur J Med Chem, 2022, 237: 114325 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/35452936/. DOI: 10.1016/j.ejmech.2022.114325.
[7]
LAMBIN P, LEIJENAAR R T H, DEIST T M, et al. Radiomics: the bridge between medical imaging and personalized medicine[J]. Nat Rev Clin Oncol, 2017, 14(12): 749-762. DOI: 10.1038/nrclinonc.2017.141.
[8]
HUANG G L, DU S Y, GAO S, et al. Molecular subtypes of breast cancer identified by dynamically enhanced MRI radiomics: the delayed phase cannot be ignored[J/OL]. Insights Imaging, 2024, 15(1): 127 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/38816553/. DOI: 10.1186/s13244-024-01713-9.
[9]
ZHANG S H, WANG X L, YANG Z, et al. Intra- and peritumoral radiomics model based on early DCE-MRI for preoperative prediction of molecular subtypes in invasive ductal breast carcinoma: a multitask machine learning study[J/OL]. Front Oncol, 2022, 12: 905551 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/35814460/. DOI: 10.3389/fonc.2022.905551.
[10]
HAN Y X, HUANG M X, XIE L Z, et al. The value of intratumoral and peritumoral radiomics features based on multiparametric MRI for predicting molecular staging of breast cancer[J/OL]. Front Oncol, 2025, 15: 1379048 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/40134601/. DOI: 10.3389/fonc.2025.1379048.
[11]
CAO M T, LIU X Y, YANG A R, et al. Prediction of HER-2 expression status in breast cancer based on multi-parameter MRI intratumoral and peritumoral radiomics[J/OL]. Magn Reson Imaging, 2025, 122: 110434 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/40460947/. DOI: 10.1016/j.mri.2025.110434.
[12]
LAFCı O, CELEPLI P, SEHER ÖZTEKIN P, et al. DCE-MRI radiomics analysis in differentiating luminal a and luminal B breast cancer molecular subtypes[J]. Acad Radiol, 2023, 30(1): 22-29. DOI: 10.1016/j.acra.2022.04.004.
[13]
MO S J, LUO H, WANG M Y, et al. Machine learning radiomics based on intra and peri tumor PA/US images distinguish between luminal and non-luminal tumors in breast cancers[J/OL]. Photoacoustics, 2024, 40: 100653 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/39399393/. DOI: 10.1016/j.pacs.2024.100653.
[14]
TOUMAJ S, HEIDARI A, JAFARI NAVIMIPOUR N. Leveraging explainable artificial intelligence for transparent and trustworthy cancer detection systems[J/OL]. Artif Intell Med, 2025, 169: 103243 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/40839960/. DOI: 10.1016/j.artmed.2025.103243.
[15]
LIU Z W, LUO C, CHEN X J, et al. Noninvasive prediction of perineural invasion in intrahepatic cholangiocarcinoma by clinicoradiological features and computed tomography radiomics based on interpretable machine learning: a multicenter cohort study[J]. Int J Surg, 2024, 110(2): 1039-1051. DOI: 10.1097/JS9.0000000000000881.
[16]
ZHANG J, ZHAN C N, ZHANG C X, et al. Fully automatic classification of breast lesions on multi-parameter MRI using a radiomics model with minimal number of stable, interpretable features[J]. Radiol Med, 2023, 128(2): 160-170. DOI: 10.1007/s11547-023-01594-w.
[17]
中国抗癌协会乳腺癌专业委员会, 中华医学会肿瘤学分会乳腺肿瘤学组. 中国抗癌协会乳腺癌诊治指南与规范(2024年版)[J]. 中国癌症杂志, 2023, 33(12): 1092-1187. DOI: 10.19401/j.cnki.1007-3639.2023.12.004.
The Society of Breast Cancer China AntiCancer Association;Breast Oncology Group of the Oncology Branch of the Chinese Medical Association. Guidelines for breast cancer diagnosis and treatment by China Anti-cancer Association(2024 edition)[J]. China Oncol, 2023, 33(12): 1092-1187. DOI: 10.19401/j.cnki.1007-3639.2023.12.004.
[18]
HUANG Y H, WEI L H, HU Y L, et al. Multi-parametric MRI-based radiomics models for predicting molecular subtype and androgen receptor expression in breast cancer[J/OL]. Front Oncol, 2021, 11: 706733 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/34490107/. DOI: 10.3389/fonc.2021.706733.
[19]
PEDUZZI P, CONCATO J, KEMPER E, et al. A simulation study of the number of events per variable in logistic regression analysis[J]. J Clin Epidemiol, 1996, 49(12): 1373-1379. DOI: 10.1016/s0895-4356(96)00236-3.
[20]
RILEY R D, SNELL K I, ENSOR J, et al. Minimum sample size for developing a multivariable prediction model: PART II - binary and time-to-event outcomes[J]. Stat Med, 2019, 38(7): 1276-1296. DOI: 10.1002/sim.7992.
[21]
匡静, 黄松涛, 黄小华, 等. 磁共振高分辨率延迟期的可解释性机器学习模型术前预测浸润性乳腺癌组织学分级[J]. 磁共振成像, 2025, 16(5): 164-169, 216. DOI: 10.12015/issn.1674-8034.2025.05.025.
KUANG J, HUANG S T, HUANG X H, et al. Interpretable machine learning model for predicting preoperative histological grade of invasive breast cancer based on high resolution delay period of magnetic resonance imaging[J]. Chin J Magn Reson Imag, 2025, 16(5): 164-169, 216. DOI: 10.12015/issn.1674-8034.2025.05.025.
[22]
王世科, 孙冬, 郭大静, 等. 基于DCE-MRI的影像组学在预测乳腺癌Luminal型和非Luminal型中的价值[J]. 重庆医科大学学报, 2019, 44(7): 933-937. DOI: 10.13406/j.cnki.cyxb.002175.
WANG S K, SUN D, GUO D J, et al. Value of radiomics based on dynamic contrast-enhanced magnetic resonance imaging in predicting luminal and non-luminal breast cancer[J]. J Chongqing Med Univ, 2019, 44(7): 933-937. DOI: 10.13406/j.cnki.cyxb.002175.
[23]
胥豪, 杨澳, 胡云涛, 等. DCE-MRI影像组学联合血细胞参数鉴别管腔型和非管腔型乳腺癌的价值[J]. 磁共振成像, 2025, 16(4): 33-40, 150. DOI: 10.12015/issn.1674-8034.2025.04.006.
XU H, YANG A, HU Y T, et al. Value of radiomics model based on DCE-MRI combined with blood cell parameters in differentiating Luminal and Non-Luminal breast cancer[J]. Chin J Magn Reson Imag, 2025, 16(4): 33-40, 150. DOI: 10.12015/issn.1674-8034.2025.04.006.
[24]
HUANG T, FAN B, QIU Y Y, et al. Application of DCE-MRI radiomics signature analysis in differentiating molecular subtypes of luminal and non-luminal breast cancer[J/OL]. Front Med (Lausanne), 2023, 10: 1140514 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/37181350/. DOI: 10.3389/fmed.2023.1140514.
[25]
CHATTERJI M, MERCADO C L, MOY L. Optimizing 1.5-Tesla and 3-Tesla dynamic contrast-enhanced magnetic resonance imaging of the breasts[J]. Magn Reson Imaging Clin N Am, 2010, 18(2): 207-224, viii. DOI: 10.1016/j.mric.2010.02.011.
[26]
MONTEMEZZI S, CAMERA L, GIRI M G, et al. Is there a correlation between 3T multiparametric MRI and molecular subtypes of breast cancer?[J/OL]. Eur J Radiol, 2018, 108: 120-127 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/30396643/. DOI: 10.1016/j.ejrad.2018.09.024.
[27]
程卫群, 戚轩, 杨宏楷, 等. 基于多参数MR生境成像鉴别Luminal型和非Luminal型乳腺癌[J]. 磁共振成像, 2025, 16(5): 170-180. DOI: 10.12015/issn.1674-8034.2025.05.026.
CHENG W Q, QI X, YANG H K, et al. Identification of type Luminal and non-type Luminal breast cancers based on multiparametric MR habitat imaging[J]. Chin J Magn Reson Imag, 2025, 16(5): 170-180. DOI: 10.12015/issn.1674-8034.2025.05.026.
[28]
张丁懿, 黄小华, 沈梦伊, 等. MRI影像组学预测非特殊型浸润性乳腺癌分子分型的价值[J]. 磁共振成像, 2024, 15(3): 100-106. DOI: 10.12015/issn.1674-8034.2024.03.017.
ZHANG D Y, HUANG X H, SHEN M Y, et al. Value of MRI radiomics in predicting molecular subtypes of invasive breast carcinoma of no special type[J]. Chin J Magn Reson Imag, 2024, 15(3): 100-106. DOI: 10.12015/issn.1674-8034.2024.03.017.
[29]
GROSSE C, NOACK P, GROSSE A, et al. Prognostic impact of histological subtyping in triple-negative breast cancer[J/OL]. Hum Pathol, 2024, 152: 105640 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/39128557/. DOI: 10.1016/j.humpath.2024.105640.
[30]
SANGES F, FLORIS M, COSSU-ROCCA P, et al. Histologic subtyping affecting outcome of triple negative breast cancer: a large Sardinian population-based analysis[J/OL]. BMC Cancer, 2020, 20(1): 491 [2025-08-01]. https://pubmed.ncbi.nlm.nih.gov/32487046/. DOI: 10.1186/s12885-020-06998-9.
[31]
刘云泽, 李宬润, 郭俊唐, 等. 基于临床-影像组学列线图模型鉴别局灶性机化性肺炎与肺腺癌[J]. 南方医科大学学报, 2024, 44(2): 397-404. DOI: 10.12122/j.issn.1673-4254.2024.02.23.
LIU Y Z, LI C R, GUO J T, et al. A clinical-radiomics nomogram for differentiating focal organizing pneumonia and lung adenocarcinoma[J]. J South Med Univ, 2024, 44(2): 397-404. DOI: 10.12122/j.issn.1673-4254.2024.02.23.
[32]
MA M W, LIU R Y, WEN C J, et al. Predicting the molecular subtype of breast cancer and identifying interpretable imaging features using machine learning algorithms[J]. Eur Radiol, 2022, 32(3): 1652-1662. DOI: 10.1007/s00330-021-08271-4.
[33]
SHAFIQ-UL-HASSAN M, ZHANG G G, LATIFI K, et al. Intrinsic dependencies of CT radiomic features on voxel size and number of gray levels[J]. Med Phys, 2017, 44(3): 1050-1062. DOI: 10.1002/mp.12123.
[34]
MAHMOOD U, APTE A P, DEASY J O, et al. Investigating the robustness neighborhood gray tone difference matrix and gray level co-occurrence matrix radiomic features on clinical computed tomography systems using anthropomorphic phantoms: evidence from a multivendor study[J]. J Comput Assist Tomogr, 2017, 41(6): 995-1001. DOI: 10.1097/RCT.0000000000000632.
[35]
CASTELLOTE-HUGUET P, RUIZ-ESPAÑA S, GALÁN-AUGE C, et al. Breast cancer diagnosis using texture and shape features in MRI[C]//2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). July 24-27, 2023, Sydney, Australia. IEEE, 2023: 1-4. DOI: 10.1109/EMBC40787.2023.10340385.
[36]
O'SHEA R, WITHEY S J, OWCZARCZYK K, et al. Multicentre validation of CT grey-level co-occurrence matrix features for overall survival in primary oesophageal adenocarcinoma[J]. Eur Radiol, 2024, 34(10): 6919-6928. DOI: 10.1007/s00330-024-10666-y.
[37]
舒予静, 何子龙, 曾伟雄, 等. 基于双模态乳腺影像特征的HER2表达状态多分类机器学习预测模型构建及可解释性分析研究[J]. 临床放射学杂志, 2024, 43(8): 1310-1316. DOI: 10.13437/j.cnki.jcr.2024.08.016.
SHU Y J, HE Z L, ZENG W X, et al. Multi-class machine learning predictive model for HER2 expression status based on dual-modal breast imaging features[J]. J Clin Radiol, 2024, 43(8): 1310-1316. DOI: 10.13437/j.cnki.jcr.2024.08.016.
[38]
冯晓丹, 刘菲菲, 杜瑶, 等. 基于超声影像组学及临床特征的可解释性机器学习模型预测乳腺癌新辅助化疗疗效[J]. 中国超声医学杂志, 2024, 40(3): 278-282. DOI: 10.3969/j.issn.1002-0101.2024.03.010.
FENG X D, LIU F F, DU Y, et al. Predicting the effect of interpretable machine learning model based on ultrasound radiomics and clinical features on neoadjuvant chemotherapy for breast cancer[J]. Chin J Ultrasound Med, 2024, 40(3): 278-282. DOI: 10.3969/j.issn.1002-0101.2024.03.010.

上一篇 多时相自由呼吸实时电影成像在心脏磁共振左心室功能评估中的应用
下一篇 基于钆塞酸二钠增强MRI的Delta影像组学模型预测肝细胞癌微血管侵犯
  
诚聘英才 | 广告合作 | 免责声明 | 版权声明
联系电话:010-67113815
京ICP备19028836号-2