Preoperative multimodal prediction model for axillary lymph node metastasis in breast cancer based on clinical, conventional ultrasound, and ultrasomics features: a two-center study
-
摘要:目的
构建基于临床特征、常规超声参数及超声组学特征的多模态融合模型用于术前预测乳腺癌腋窝淋巴结转移(ALNM),并评估其跨中心泛化能力及临床应用价值。
方法回顾性纳入2020年1月1日至2025年1月1日联勤保障部队第九一〇医院(230例,按7∶3随机分为训练集和内部验证集,分别为161例、69例)与泉州市第一医院(91例,外部验证集)共321例乳腺癌患者。采用递归特征消除方法筛选临床和常规超声参数,采用最小绝对收缩和选择算子回归筛选超声组学特征,然后基于确立的最优算法glmnet构建临床模型(仅纳入临床和常规超声参数)、组学模型(仅纳入超声组学特征)和融合模型(纳入临床和常规超声参数及超声组学特征),并分别在训练集、内部验证集和外部验证集通过ROC曲线、Hosmer-Lemeshow检验、Brier分数、决策曲线分析(DCA)、净重分类指数(NRI)及综合判别改善指数(IDI)等方法评估各模型预测ALNM的性能与泛化能力。在同时接受前哨淋巴结活检(SLNB)与腋窝淋巴结清扫(ALND)的子集(n=142)中,对比融合模型与SLNB的诊断效能。在外部验证集中,采用ROC曲线的约登指数确定最优截断值并结合DCA临床获益风险比将患者划分为低、中、高危组,评估融合模型指导临床决策的能力;同时按分子分型(Luminal、人表皮生长因子受体2阳性、三阴性)及临床T分期进行亚组效能评价。
结果321例患者中ALNM阳性139例(43.3%)。共筛选出4个关键临床和常规超声特征及18个超声组学特征。在内部验证集与外部验证集中,融合模型AUC值分别为0.829(95%CI 0.779~0.878)与0.854(95%CI 0.810~0.898),均优于临床和组学模型;校准度优异(Hosmer-Lemeshow检验P=0.921、0.365,Brier分数为0.010、0.020),DCA显示融合模型在多数阈值概率范围内的净获益均最高,且以融合模型为参照时临床、组学模型的NRI与IDI均为负值。在对照子集中,融合模型预测ALNM的灵敏度为0.862,与SLNB(0.914)相比差异无统计学意义(McNemar检验P=0.286)。在分子分型与临床T分期亚组分析中,融合模型均能有效检出ALNM高危患者。
结论基于临床信息、常规超声与组学特征的多模态融合模型能够准确预测乳腺癌ALNM,且具备良好的稳定性、校准度与跨中心泛化能力,有望为术前腋窝淋巴结管理策略如是否行SLNB或ALND提供可靠的无创定量依据。
Abstract:ObjectiveTo construct a multimodal fusion model integrating clinical features, conventional ultrasound parameters and ultrasomics features for preoperative prediction of axillary lymph node metastasis (ALNM) in breast cancer, and to evaluate its cross-center generalizability and clinical application value.
MethodsA total of 321 breast cancer patients were retrospectively enrolled from Jan. 1, 2020 to Jan. 1, 2025, including 230 patients from No. 910 Hospital of Joint Logistics Support Force (randomly assigned to training set or internal validation set at a 7∶3 ratio, yielding 161 and 69 cases, respectively) and 91 patients from Quanzhou First Hospital (serving as external validation set). Recursive feature elimination method was adopted to screen clinical and conventional ultrasound parameters, and least absolute shrinkage and selection operator regression was used to select ultrasomics features. Based on the optimal algorithm glmnet, a clinical model (incorporating only clinical and conventional ultrasound parameters), an ultrasomics model (incorporating only ultrasomics features), and a fusion model (incorporating clinical and conventional ultrasound parameters together with ultrasomics features) were established. The predictive performance and generalizability of each model for ALNM were evaluated in the training set, internal validation set, and external validation set using receiver operating characteristic (ROC) curve, Hosmer-Lemeshow test, Brier score, decision curve analysis (DCA), net reclassification improvement (NRI), and integrated discrimination improvement (IDI). In a subgroup of 142 patients who underwent both sentinel lymph node biopsy (SLNB) and axillary lymph node dissection (ALND), the diagnostic efficacy was compared between the fusion model and SLNB. In the external validation set, based on the optimal cut-off value determined by the Youden index of ROC curve and the clinical benefit-to-risk ratio calculated by DCA, the patients were assigned to low-, intermediate-, or high-risk groups to evaluate the ability of the fusion model to guide clinical decision-making. Subgroup analyses stratified by molecular subtype (Luminal, human epidermal growth factor receptor 2-positive, and triple-negative) or clinical T stage were further performed to assess performance of the fusion model.
ResultsAmong the 321 patients, 139 (43.3%) had positive ALNM. Four key clinical and conventional ultrasound parameters and 18 ultrasomics features were screened. In the internal and external validation sets, the area under curve (AUC) values of the fusion model were 0.829 (95% confidence interval [95%CI] 0.779-0.878) and 0.854 (95%CI 0.810-0.898), respectively, which were superior to those of the clinical and ultrasomics models. The fusion model showed excellent calibration, with Hosmer-Lemeshow test P values of 0.921 and 0.365, and Brier scores of 0.010 and 0.020 in the internal and external validation sets, respectively. DCA demonstrated that the fusion model yielded the highest net benefit across most threshold probability ranges. With the fusion model as the reference, the clinical model and ultrasomics model had negative NRI and IDI values. In the comparison subgroup, the sensitivity of the fusion model for predicting ALNM was 0.862, showing no significant difference compared with SLNB (sensitivity=0.914, McNemar test P=0.286). In the subgroup analyses by molecular subtype or clinical T stage, the fusion model effectively identified patients at high risk for ALNM.
ConclusionThe multimodal fusion model based on clinical, conventional ultrasound, and ultrasomics features can accurately predict ALNM in breast cancer, with favorable stability, calibration, and cross-center generalizability. This model is expected to provide a reliable non-invasive quantitative basis for preoperative axillary lymph node management strategies, including decision-making regarding SLNB or ALND.
-
乳腺癌已成为全球女性发病率最高的恶性肿瘤,其对公共健康的影响日益凸显[1]。在乳腺癌的综合治疗体系中,腋窝淋巴结状态不仅决定病理分期,也是制定手术方式、术后放化疗及靶向治疗方案的关键依据。当前以腋窝淋巴结清扫(axillary lymph node dissection,ALND)或前哨淋巴结活检(sentinel lymph node biopsy,SLNB)为代表的手术路径虽能提供可靠的病理信息,但60%~70%的临床腋窝淋巴结阴性患者最终被证实无淋巴结转移,却因接受了不必要的外科操作而出现淋巴水肿、感觉障碍及肩关节活动受限等并发症[2]。因此,急需一种准确、无创且可推广的术前评估技术用于更精准地区分腋窝淋巴结转移低风险与高风险患者,从而减少过度治疗。
超声检查因无辐射、可重复、价格低廉且易于临床实施,在乳腺及腋窝淋巴结评估中占据核心地位。然而,传统超声诊断多依赖于淋巴结皮质厚度、髓门结构及血流信号等主观指标,受操作者经验影响明显,难以满足高精度风险分层需求[3]。超声组学(ultrasomics)通过从常规医学超声影像提取大量定量特征,能够在纹理、灰度、形状等多维度捕捉肿瘤内部异质性,有望突破传统超声的局限。近年研究发现,超声组学模型预测乳腺癌腋窝淋巴结转移的AUC值为0.75~0.85,深度学习超声组学模型AUC值甚至达到0.90以上[4-6],显示出超声组学在无创诊断领域的潜力。
尽管相关研究已取得诸多进展,但仍存在以下不足。一是多数研究为单中心、小样本设计,缺乏跨中心验证,模型的稳定性和泛化能力不足。二是许多模型主要关注肿瘤影像特征,较少结合腋窝淋巴结的结构与血流动力学信息,未能全面反映淋巴结转移发生过程中的微环境变化[7]。三是部分研究缺乏严格的特征筛选、标准化影像处理及可解释性分析[如沙普利加性解释(Shapley additive explanation,SHAP)],影响模型的透明度与临床可接受度[8]。因此,在规范化流程下构建融合临床信息、传统超声指标及超声组学特征的多模态模型,并经外部验证证实其可靠性,对于推动该领域向临床应用迈进具有重要意义。
本研究基于联勤保障部队第九一〇医院与泉州市第一医院的双中心样本,构建了融合临床变量、常规超声特征及超声组学特征的多模态乳腺癌腋窝淋巴结转移预测模型,并采用多层次验证策略评价其性能。研究假设多模态融合策略能够提升预测精度,而SHAP方法可增强模型的可解释性;同时,通过外部验证可进一步确认模型的稳定性与临床推广价值。本研究旨在开发一种术前、无创、准确且具备临床可落地性的乳腺癌腋窝淋巴结转移风险评估工具,为乳腺癌个体化腋窝淋巴结管理提供量化支持。
1 资料和方法
1.1 研究设计与病例资料
本研究为回顾性、多中心诊断预测模型研究。连续纳入2020年1月1日至2025年1月1日在联勤保障部队第九一〇医院和泉州市第一医院接受手术治疗并经病理证实为浸润性乳腺癌的女性患者。所有入组病例均于术前完成乳腺及患侧腋窝淋巴结常规超声检查,并以SLNB和/或ALND术后病理结果作为判断腋窝淋巴结转移状态的金标准。纳入标准:(1)女性,年龄≥18岁;(2)术前穿刺或术后病理证实为浸润性乳腺癌;(3)术前2~4周内完成乳腺及腋窝淋巴结超声检查,且超声图像质量满足分析要求;(4)超声检查完成至手术前未接受过任何可能影响肿瘤或淋巴结状态的抗肿瘤治疗,包括新辅助化疗、内分泌治疗、抗人表皮生长因子受体2(human epidermal growth factor receptor 2,HER2)靶向治疗、免疫治疗及放疗等;(5)接受手术治疗并行SLNB和/或ALND,且有完整的腋窝淋巴结病理结果;(6)关键临床与病理资料完整。排除标准:(1)既往乳腺恶性肿瘤史或其他恶性肿瘤史;(2)双侧乳腺癌;(3)复发性乳腺癌(既往同侧手术后复发);(4)多灶/多中心乳腺癌,其中多灶指同一象限内有至少2个肿瘤灶,多中心指不同象限内有肿瘤灶或肿瘤灶间距≥2 cm(以超声测量为准),且上述情况可能导致腋窝淋巴结转移归因/对应肿瘤灶不明确;(5)超声检查至手术间隔>4周;(6)超声图像存在明显伪影、严重声衰减或无法清晰勾画感兴趣区域(region of interest,ROI);(7)关键免疫组织化学指标,如雌激素受体(estrogen receptor,ER)、孕激素受体(progesterone receptor,PR)或HER2数据缺失。
本研究最终纳入321例患者,其中联勤保障部队第九一〇医院230例作为内部队列,泉州市第一医院91例作为外部验证队列。内部队列采用按腋窝淋巴结转移状态分层的随机抽样方法以7∶3比例划分为训练集(n=161)与内部验证集(n=69)。通过R 4.1.2软件使用固定随机种子(set.seed=2020)完成随机化过程。本研究通过联勤保障部队第九一〇医院和泉州市第一医院医学伦理委员会审批(伦理审批号分别为Keshen2025-5-22-01、NO.2025-183);本研究为回顾性研究且不涉及额外干预,故豁免知情同意;所有数据在分析前均进行匿名化处理。
1.2 超声检查方法
2家中心均使用Mindray R9多普勒超声仪(配备5~12 MHz高频线阵探头)完成乳腺及腋窝淋巴结超声检查。为减小跨中心图像采集差异对超声组学特征稳定性的影响,本研究纳入的图像均为关键成像参数采用了初始或默认设置的图像。
1.2.1 超声成像参数设置
乳腺与腋窝淋巴结灰阶图像采集均采用以下参数设置:总体增益初始设定为45 dB(允许在±5 dB范围内微调,以确保脂肪组织回声不饱和、肿瘤病灶/淋巴结边界清晰);成像深度为3~5 cm(以完整覆盖目标病灶/淋巴结及周围组织1~2 cm为原则);聚焦点设置于目标病灶或淋巴结的中心位置后方,优先使用单焦点以保证ROI空间分辨率与纹理一致性。动态范围/压缩、灰阶后处理(如降噪、边缘增强)采用统一默认档位,尽量减少个体化后处理对纹理特征的影响。用于超声组学分析的图像为原始灰阶图像(DICOM格式)。
1.2.2 超声检查与资料采集
(1)乳腺超声检查:获取肿瘤长轴及短轴代表性切面灰阶图像,记录肿瘤最大径、形态(规则/不规则)、边界(清晰/不清晰)、内部回声(均匀/不均匀)、有无钙化、峰值收缩期流速(peak systolic velocity,PSV)、舒张末期流速(end-diastolic velocity,EDV)、阻力指数、搏动指数,以及反映肿瘤血流异质性的PSV/EDV比值。(2)腋窝淋巴结超声检查:系统扫查腋窝Ⅰ~Ⅲ区淋巴结,记录淋巴结长径、短径、皮髓质面积比值、PSV、EDV、阻力指数、搏动指数,以及反映淋巴结血流异质性的PSV/EDV比值。
1.3 临床与病理资料收集
收集患者年龄、绝经状态,以及肿瘤位置、分子分型、临床T分期和免疫组织化学指标(ER、PR、HER2等),结局变量为腋窝淋巴结是否转移。
1.4 超声图像预处理与超声组学特征提取
所有超声图像均按照影像生物标志物标准化倡议(Image Biomarker Standardisation Initiative,IBSI)建议流程进行统一预处理,包括像素间距重采样、灰度强度归一化及灰度离散化(其中灰度离散化采用固定组距法,组距设为25),以降低不同设备与采集参数带来的系统性差异。由2名具有至少5年乳腺超声检查经验的超声科医生在不知晓病理结果的前提下,独立手工勾画ROI:沿腋窝淋巴结整体边界(包括皮质与髓质,不进行皮质、髓质分区)勾画,尽量避开周围脂肪、血管结构及明显伪影区域。为评估ROI勾画一致性,对超声图像进行重复勾画,采用基于双向随机效应模型绝对一致性定义计算的组内相关系数(intraclass correlation coefficient,ICC)进行一致性检验,ICC≥0.75表示一致性良好。对于ROI勾画存在分歧(ICC<0.75)的图像,由第3名具有至少10年乳腺超声检查经验的高级职称医生进行复核并裁决,最终以达到一致意见的ROI版本用于后续特征提取与分析。本研究中同一医生2次勾画的ICC中位数(范围)为0.85(0.78~0.90),2名医生间的ICC中位数(范围)为0.80(0.75~0.86),一致性均达到良好水平(ICC≥0.75)。
基于PyRadiomics 3.0.1软件从ROI中提取形状特征、一阶统计特征及纹理特征,包括灰度共生矩阵(gray-level co-occurrence matrix)、灰度游程矩阵(gray-level run-length matrix)、灰度大小区域矩阵(gray-level size zone matrix)、邻域灰度差矩阵(neighbouring gray-tone difference matrix)和灰度依赖矩阵(gray-level dependence matrix)等,共获得约1 000个组学特征。
1.5 特征筛选与模型构建
1.5.1 临床及常规超声特征筛选
采用递归特征消除(recursive feature elimination,RFE)结合交叉验证方法筛选与结局变量相关的临床和常规超声特征,随后基于SHAP方法对筛选得到的特征进行贡献度分析,并明确各特征对结局的影响方向。
1.5.2 超声组学特征的筛选与降维处理
超声组学特征的筛选与降维处理均在训练集中完成。首先,采用ICC对从2名医生独立勾画的ROI提取的全部组学特征进行一致性评估,保留ICC≥0.75的特征。随后,为降低特征冗余与共线性影响,对保留的特征进行相关性过滤:计算特征两两之间的Spearman相关系数(rs),若|rs|>0.90,则保留该相关特征组中与结局相关性更强的一个特征进入下一步分析。基于上述预筛选后的特征集合,采用最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)正则化logistic回归进行进一步降维,并通过10折交叉验证选择最优惩罚参数λ。将LASSO回归模型中系数非零的特征作为最终超声组学特征,用于组学模型与融合模型构建。
1.5.3 3种对比模型构建
(1)临床模型:临床模型仅纳入1.5.1节筛选出的临床及常规超声特征。变量进入多因素模型的策略为:首先在训练集中进行单因素分析筛选出候选变量,随后采用逐步多因素logistic回归方法确定最终变量组合。(2)组学模型:组学模型仅基于1.5.2节筛选出的超声组学特征构建,即超声组学评分(ultrasomics score)。超声组学评分通过各特征的线性加权求和计算得到。(3)融合模型:融合模型为在临床模型基础上,纳入超声组学评分(作为单一连续变量)并通过多因素logistic回归构建。
1.6 与常规检查策略的效能对比
为评估本研究模型与临床常规检查策略SLNB的诊断效能差异,以同时接受SLNB和ALND(包括SLNB阳性后补充ALND或术中/术后转为ALND)的患者作为对照子集,并以ALND术后最终腋窝淋巴结活检病理结果作为金标准,分别计算SLNB及融合模型在该子集中诊断腋窝淋巴结转移的灵敏度、特异度、准确度、阳性预测值与阴性预测值,并采用Wilson法计算各指标的95%CI。组间灵敏度、特异度的差异分析采用配对资料的McNemar检验。
1.7 风险分层与亚组分析
在外部验证集中,采用融合模型输出的腋窝淋巴结转移概率进行临床风险分层。首先通过ROC曲线的约登指数确定融合模型的最佳分类阈值,然后结合决策曲线分析(decision curve analysis,DCA)提示的净获益区间,将患者分为低危(预测概率<0.15)、中危(0.15~0.45)与高危(>0.45)3层,评估不同风险层的实际腋窝淋巴结转移发生率及潜在临床管理意义。
亚组分析按照乳腺癌分子分型[Luminal型、HER2阳性、三阴性乳腺癌(triple-negative breast cancer,TNBC)]及临床T分期(T1、T2、T3)分层,在各亚组内计算融合模型预测腋窝淋巴结转移的AUC值及95%CI,以评估融合模型在不同人群中的稳定性与适用范围。
1.8 统计学处理
所有数据由双人双录入核对后纳入分析。计量资料若符合正态分布以$\bar{x} \pm s$表示,若不符合正态分布则以M(Q1,Q3)表示,两组间比较分别采用独立样本t检验和Mann-Whitney U检验,多组间比较分别采用单因素方差分析和Kruskal-Wallis H检验;计数资料以例数和百分数表示,组间比较采用χ2检验或Fisher确切概率法。对于缺失比例<5%的变量采用多重插补法进行补全。
模型构建阶段,在训练集中开发了多种机器学习算法,包括广义线性模型(generalized linear model,GLM)、广义线性模型正则化路径算法包/弹性网络模型glmnet、线性核函数支持向量机(support vector machine with a linear kernel,svmLinear)、径向基核函数支持向量机(support vector machine with a radial basis function kernel,svmRadial)、随机森林(random forest,RF)、梯度提升机(gradient boosting machine,GBM)、极端梯度提升(extreme gradient boosting,XGBoost)、自适应提升(adaptive boosting,AdaBoost)、C5.0决策树(C5.0 decision tree)、朴素贝叶斯(naive Bayes)、k近邻算法(k-nearest neighbor,kNN)、神经网络(neural network,NNet)、递归分割与回归树(recursive partitioning and regression tree,rpart)、线性判别分析(linear discriminant analysis,LDA)、灵活判别分析(flexible discriminant analysis,FDA)及混合判别分析(mixture discriminant analysis,MDA)等,并采用10折交叉验证优化各算法超参数,以ROC曲线AUC值作为主要性能评价指标。在综合比较各算法表现后,本研究最终选择glmnet作为最优基础算法。
采用glmnet算法分别构建临床模型、组学模型及融合模型,并在内部验证集和外部验证集中进行模型验证。通过ROC曲线的AUC值及95%CI评价模型区分度,采用DeLong检验比较各模型间AUC值,并采用Bonferroni法进行多重比较校正[每个队列中分别进行融合模型与临床模型、组学模型的比较,因此校正后P值(Padj)=min(PDeLong×2, 1)]。采用Hosmer-Lemeshow检验与Brier分数(Brier分数取值范围为0~1,得分越低表示预测误差越小)评价模型校准度。通过自助法(Bootstrap法,重复抽样1 000次)绘制校准曲线,并按预测概率十分位分组报告各模型预测概率均值与实际发生率的平均偏差和平均绝对偏差,以量化预测概率与实际发生率的偏离程度;进一步计算校准截距与校准斜率以定量表征校准度,其中截距接近0、斜率接近1表示模型校准度良好。为避免仅依赖Hosmer-Lerneshow检验与Brier分数导致的校准度评价不充分,同时采用Spiegelhalter Z检验等补充校准指标进行综合评估。基于DCA评价各模型在不同阈值概率下的临床净获益。此外,以融合模型为参照,计算净重分类指数(net reclassification improvement,NRI)和综合判别改善指数(integrated discrimination improvement,IDI),评价融合模型相较临床模型及组学模型在风险重分类能力方面的改善程度。
本研究所有统计学分析均采用R 4.1.2软件完成,主要使用caret、glmnet、pROC、rms、rmda、ggDCA及survIDINRI等工具包实现。检验水准(α)为0.05。
2 结果
2.1 患者一般情况比较
本研究共纳入321例乳腺癌患者,包括内部训练集161例、内部验证集69例及外部验证集91例。3个队列间患者年龄、绝经状态、肿瘤位置、肿瘤形态、肿瘤边界、肿瘤内部回声、肿瘤内钙化情况、肿瘤直径、肿瘤及淋巴结的血流动力学指标(PSV、EDV、阻力指数、搏动指数等)、淋巴结形态学参数(长径、宽径、皮髓质面积比值)以及结局变量(腋窝淋巴结转移发生率)差异均无统计学意义(均P>0.05)。见表 1。
表 1 3个队列乳腺癌患者的临床及常规超声特征比较Table 1 Comparison of clinical and conventional ultrasound characteristics of breast cancer patients across 3 cohortsVariable Training set N=161 Internal validation set N=69 External validation set N=91 Statistic P value Age/year, $\bar{x} \pm s$ 49.53±8.76 49.29±8.47 50.58±8.50 F=0.570 0.566 Lymph node metastasis, n (%) χ2=1.994 0.369 No 94 (58.4) 34 (49.3) 54 (59.3) Yes 67 (41.6) 35 (50.7) 37 (40.7) Menopausal status, n (%) χ2=2.469 0.291 No 82 (50.9) 33 (47.8) 37 (40.7) Yes 79 (49.1) 36 (52.2) 54 (59.3) Tumor Location, n (%) 0.550a Upper outer 41 (25.5) 20 (29.0) 26 (28.6) Lower outer 26 (16.1) 7 (10.1) 10 (11.0) Lower inner 34 (21.1) 22 (31.9) 26 (28.6) Upper inner 49 (30.4) 16 (23.2) 21 (23.1) Areolar region 11 (6.8) 4 (5.8) 8 (8.8) Regular shape, n (%) χ2=0.053 0.974 Yes 82 (50.9) 34 (49.3) 46 (50.5) No 79 (49.1) 35 (50.7) 45 (49.5) Clear margin, n (%) χ2=0.825 0.662 Yes 82 (50.9) 33 (47.8) 41 (45.1) No 79 (49.1) 36 (52.2) 50 (54.9) Homo-echo, n (%) χ2=4.901 0.086 Yes 79 (49.1) 41 (59.4) 38 (41.8) No 82 (50.9) 28 (40.6) 53 (58.2) Calcification, n (%) χ2=0.572 0.751 Yes 83 (51.6) 33 (47.8) 49 (53.8) No 78 (48.4) 36 (52.2) 42 (46.2) Diameter/mm, M (Q1, Q3) 12.80 (10.01, 16.37) 13.98 (10.90, 17.94) 13.60 (10.67, 17.34) H=3.219 0.200 PSV/(cm·s-1), $\bar{x} \pm s$ 37.50±6.26 39.05±5.70 38.88±6.28 F=2.251 0.107 EDV/(cm·s-1), $\bar{x} \pm s$ 9.94±2.00 10.20±1.92 10.21±2.42 F=0.627 0.535 PSV/EDV ratio, M (Q1, Q3) 3.78 (3.10, 4.61) 3.85 (3.21, 4.63) 3.88 (3.25, 4.62) H=0.165 0.921 Resistive index, $\bar{x} \pm s$ 0.57±0.19 0.59±0.19 0.60±0.20 F=0.490 0.613 Pulsatility index, $\bar{x} \pm s$ 1.76±0.31 1.84±0.26 1.84±0.28 F=2.789 0.063 Lymph node Short diameter/mm, $\bar{x} \pm s$ 5.91±1.34 5.90±1.21 5.93±1.24 F=0.016 0.984 Long diameter/mm, $\bar{x} \pm s$ 11.57±2.02 11.51±2.08 11.77±2.05 F=0.394 0.675 CMR, $\bar{x} \pm s$ 1.93±0.30 1.95±0.27 1.89±0.26 F=1.128 0.325 PSV/(cm·s-1), M (Q1, Q3) 16.96 (13.69, 21.02) 15.58 (12.13, 20.03) 16.23 (12.69, 20.76) H=1.683 0.431 EDV/(cm·s-1), M (Q1, Q3) 8.20 (6.89, 9.74) 7.69 (6.57, 8.99) 7.14 (6.06, 8.41) H=2.694 0.260 PSV/EDV ratio, M (Q1, Q3) 2.57 (1.79, 3.71) 3.19 (2.53, 4.01) 3.00 (2.13, 4.22) H=5.077 0.079 Resistive index, $\bar{x} \pm s$ 0.60±0.19 0.62±0.20 0.60±0.18 F=0.232 0.793 Pulsatility index, $\bar{x} \pm s$ 1.69±0.31 1.67±0.26 1.67±0.27 F=0.238 0.788 a: Fisher exact test. Homo-echo: Homogeneous echogenicity; PSV: Peak systolic velocity; EDV: End-diastolic velocity; CMR: Cortex-to-medulla area ratio. 在3个队列中,腋窝淋巴结转移患者的肿瘤直径均大于无腋窝淋巴结转移者(均P<0.001),并伴随更高的肿瘤阻力指数和搏动指数(均P<0.05);肿瘤形态、边界、内部回声在多数队列中与腋窝淋巴结转移无关(均P>0.05),而肿瘤内钙化在训练集与内部验证集中均与腋窝淋巴结转移相关(P=0.026、0.022)。淋巴结相关指标中,腋窝淋巴结转移患者的淋巴结短径、皮髓质面积比值与淋巴结阻力指数在3个队列中均高于无腋窝淋巴结转移者(均P<0.05),淋巴结PSV/EDV比值在内部训练集和内部验证集中亦与腋窝淋巴结转移相关(P=0.007、0.002)。见表 2。
表 2 3个队列中ALNM+与ALNM-组乳腺癌患者常规超声特征比较Table 2 Comparison in conventional ultrasound characteristics of breast cancer patients between ALNM+ and ALNM- groups in 3 cohortsVariable Training set Internal validation set External validation set ALNM-N=94 ALNM+N=67 P value ALNM-N=34 ALNM+N=35 Pvalue ALNM-N=54 ALNM+N=37 P value Tumor Irregular shape, n (%) 40 (42.6) 39 (58.2) 0.072 15 (44.1) 20 (57.1) 0.400 24 (44.4) 21 (56.8) 0.347 Unclear margin, n (%) 40 (42.6) 39 (58.2) 0.072 16 (47.1) 20 (57.1) 0.550 29 (53.7) 21 (56.8) 0.942 Het-echo, n (%) 43 (45.7) 39 (58.2) 0.162 8 (23.5) 20 (57.1) 0.009 32 (59.3) 21 (56.8) 0.983 Calcification, n (%) 41 (43.6) 42 (62.7) 0.026 11 (32.4) 22 (62.9) 0.022 26 (48.1) 23 (62.2) 0.270 Diameter/mm, $\bar{x} \pm s$ 11.54±5.11 16.68±5.29 <0.001 11.88±5.65 17.97±5.85 <0.001 12.45±5.31 17.51±5.56 <0.001 PSV/ (cm·s-1), $\bar{x} \pm s$ 36.64±6.20 38.70±6.39 0.044 37.81±5.63 40.25±5.92 0.085 38.06±6.17 40.08±6.46 0.139 EDV/ (cm·s-1), $\bar{x} \pm s$ 9.80±1.99 10.14±2.06 0.293 9.99±1.92 10.40±1.99 0.392 10.07±2.39 10.41±2.48 0.520 PSV/EDV ratio, $\bar{x} \pm s$ 3.84±1.19 4.10±1.22 0.186 3.85±1.10 4.15±1.15 0.265 3.91±1.04 4.16±1.10 0.274 Resistive index, $\bar{x} \pm s$ 0.53±0.19 0.63±0.20 0.001 0.53±0.19 0.65±0.20 0.011 0.56±0.20 0.66±0.21 0.023 Pulsatility index, $\bar{x} \pm s$ 1.69±0.31 1.86±0.32 0.001 1.74±0.26 1.94±0.27 0.002 1.77±0.28 1.94±0.29 0.007 Lymph node Short diameter/mm, $\bar{x} \pm s$ 5.27±1.33 6.81±1.39 <0.001 4.97±1.21 6.80±1.24 <0.001 5.31±1.24 6.83±1.27 <0.001 Long diameter/mm, $\bar{x} \pm s$ 11.36±1.98 11.87±2.07 0.117 11.20±2.04 11.81±2.16 0.232 11.56±2.04 12.07±2.10 0.257 CMR, $\bar{x} \pm s$ 1.75±0.30 2.18±0.31 <0.001 1.69±0.27 2.20±0.28 <0.001 1.72±0.26 2.14±0.27 <0.001 PSV/ (cm·s-1), $\bar{x} \pm s$ 17.13±5.76 18.84±5.97 0.071 15.67±6.33 17.70±6.66 0.199 16.66±6.52 18.35±6.78 0.241 EDV/ (cm·s-1), $\bar{x} \pm s$ 8.68±2.20 8.17±2.28 0.154 8.21±1.84 7.60±1.92 0.183 7.56±1.77 7.05±1.85 0.197 PSV/EDV ratio, $\bar{x} \pm s$ 2.66±1.73 3.43±1.79 0.007 2.92±1.17 3.83±1.22 0.002 3.10±1.82 3.86±1.90 0.061 Resistive index, $\bar{x} \pm s$ 0.56±0.19 0.65±0.20 0.006 0.57±0.20 0.67±0.21 0.040 0.57±0.18 0.65±0.18 0.034 Pulsatility index, $\bar{x} \pm s$ 1.65±0.31 1.74±0.32 0.089 1.62±0.26 1.72±0.27 0.113 1.64±0.27 1.72±0.28 0.152 ALNM+: Positive axillary lymph node metastasis; ALNM-: Negative axillary lymph node metastasis; Het-echo: Heterogeneous echogenicity; PSV: Peak systolic velocity; EDV: End-diastolic velocity; CMR: Cortex-to-medulla area ratio. 2.2 临床和常规超声特征的筛选与SHAP重要性分析
经RFE方法对临床及常规超声参数进行筛选,最终确定了4个对腋窝淋巴结转移具有独立预测价值的关键特征。按照SHAP重要性排序,淋巴结PSV/EDV比值对模型预测效能的贡献度最高,其次依次为肿瘤直径、淋巴结皮髓质面积比值和肿瘤阻力指数。根据SHAP值分布,上述4个特征呈现出明确的正向驱动特征,即随着指标数值的增加模型判定腋窝淋巴结转移的概率增加(图 1)。典型病例分析也表明,同时具备肿瘤直径大、肿瘤阻力指数高、腋窝淋巴结皮髓质面积比值高与淋巴结PSV/EDV比值高等表现的患者,术后病理证实腋窝淋巴结转移阳性(图 2A);而上述指标均处于低风险水平的患者,术后病理证实腋窝淋巴结转移阴性(图 2B)。
图 2 高风险与低风险腋窝淋巴结转移的代表性乳腺癌病例的超声图像Fig. 2 Ultrasound images of representative breast cancer patients with high- or low-risk axillary lymph node metastasisA: A high-risk (axillary lymph node metastasis positive) case. The left panels showed ultrasonography and spectral Doppler imaging of the primary breast tumor, with a maximum tumor diameter >3.0 cm and a tumor resistive index >0.80. The right panels showed ultrasonography and spectral Doppler imaging of the ipsilateral axillary lymph node, with a lymph node cortex-to-medulla area ratio >5.0and a lymph node PSV/EDV ratio of approximately 4.8. Postoperative pathology confirmed axillary lymph node metastasis. B: A low-risk(axillary lymph node metastasis negative) case. The left panels showed ultrasonography and spectral Doppler imaging of the primary breast tumor, with a tumor diameter of approximately 1.3 cm and a tumor resistive index of approximately 0.64. The right panels showed ultrasonography and spectral Doppler imaging of the ipsilateral axillary lymph node, with a lymph node cortex-to-medulla area ratio of approximately 1.1 and a lymph node PSV/EDV ratio of approximately 2.2. Postoperative pathology confirmed no axillary lymph node metastasis. PSV: Peak systolic velocity; EDV: End-diastolic velocity.2.3 超声组学特征的筛选与SHAP重要性分析
采用LASSO回归对提取的超声组学特征进行降维筛选(图 3A),并基于10折交叉验证(图 3B)确定的最优惩罚参数,最终筛选出18个具有非零系数的关键组学特征。根据SHAP重要性排序GLCM_JointEntropy(联合熵)和Shape_Sphericity(球形度)等特征占据较大的权重,提示肿瘤的形态规则性及内部纹理的复杂性是模型预测腋窝淋巴结转移的关键影像学依据(图 3C)。
图 3 超声组学特征筛选及重要性分析Fig. 3 Ultrasomics feature selection and importance analysisA: LASSO regression coefficient path plot. The plot showed the convergence trajectories of the partial regression coefficients of the 18 ultrasomics features as λ changes. As λ increased, the coefficients of redundant features were gradually shrunk to 0. B: Ten-fold cross-validation AUC curve. The 2 vertical dashed lines indicated the optimal λ for maximum AUC (right) and the λ under 1-standard error rule (left). C: SHAP importance beeswarm plot. The plot showed the direction and magnitude of the contributions of the included features to the predicted probability of axillary lymph node metastasis. LASSO: Least absolute shrinkage and selection operator; AUC: Area under area; SHAP: Shapley additive explanation.2.4 多种机器学习算法的构建与效能比较
为筛选出预测腋窝淋巴结转移的最优模型,本研究在训练集中基于筛选出的特征构建了多种机器学习算法模型。根据ROC曲线分析结果,大多数算法模型在训练集中均表现出了良好的区分度,其中glmnet算法模型展现出了最优的综合预测性能,其AUC值高达0.979 0,同时保持了极高的灵敏度(0.970 1)和特异度(0.989 4),见表 3。因此,本研究最终选择glmnet算法构建腋窝淋巴结转移预测模型。
表 3 不同机器学习算法模型在训练集中的效能分析Table 3 Performance of different machine learning algorithm models in training setAlgorithm AUC Sensitivity Specificity Accuracy PPV NPV F1 score glmnet 0.979 0 0.970 1 0.989 4 0.981 4 0.984 8 0.978 9 0.977 4 NNet 0.976 7 0.970 1 0.989 4 0.981 4 0.984 8 0.978 9 0.977 4 svmLinear 0.972 5 0.970 1 0.989 4 0.981 4 0.984 8 0.978 9 0.977 4 XGBoost 0.970 2 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 C5.0 0.969 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 GBM 0.967 7 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 FDA 0.963 1 0.940 3 0.989 4 0.968 9 0.984 4 0.958 8 0.961 8 AdaBoost 0.960 5 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 LDA 0.957 2 0.910 4 0.968 1 0.944 1 0.953 1 0.938 1 0.931 3 RF 0.952 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 MDA 0.938 3 0.910 4 0.968 1 0.944 1 0.953 1 0.938 1 0.931 3 svmRadial 0.934 9 0.955 2 0.978 7 0.968 9 0.969 7 0.968 4 0.962 4 GLM 0.922 2 0.970 1 0.989 4 0.981 4 0.984 8 0.978 9 0.977 4 naive Bayes 0.913 9 0.895 5 0.968 1 0.937 9 0.952 4 0.928 6 0.923 1 kNN 0.884 8 0.850 7 0.946 8 0.906 8 0.919 4 0.899 0 0.883 7 rpart 0.867 4 0.865 7 0.946 8 0.913 0 0.920 6 0.908 2 0.892 3 glmnet: Generalized linear model via elastic-net regularization; NNet: Neural network; svmLinear: Support vector machine with a linear kernel; XGBoost: Extreme gradient boosting; C5.0: C5.0 decision tree; GBM: Gradient boosting machine; FDA: Flexible discriminant analysis; AdaBoost: Adaptive boosting; LDA: Linear discriminant analysis; RF: Random forest; MDA: Mixture discriminant analysis; svmRadial: Support vector machine with a radial basis function kernel; GLM: Generalized linear model; kNN: K-nearest neighbor; rpart: Recursive partitioning and regression tree; AUC: Area under curve; PPV: Positive predictive value; NPV: Negative predictive value. 2.5 3个预测模型在不同队列中的预测效能比较与验证
(1)在训练集中,与组学模型(AUC=0.774)与临床模型(AUC=0.745)相比,融合模型显示出更佳的腋窝淋巴结转移区分度(AUC=0.845,95%CI 0.818~0.872);经Bonferroni法校正后,融合模型与组学模型、临床模型的AUC值比较差异均有统计学意义(Padj=0.036、0.002)。在校准度方面,融合模型的Hosmer-Lemeshow检验P=0.855,Brier分数为0.014;基于预测概率十分位分组数据,融合模型的平均偏差为0.012,平均绝对偏差为0.019,校准截距为0.01,校准斜率为0.99,提示融合模型的预测概率与实际发生率总体一致;组学模型与临床模型的Hosmer-Lemeshow检验P<0.05且Brier分数较高(分别为0.243、0242),平均绝对偏差分别为0.075与0.083,提示两者均存在一定校准偏差。DCA显示,在较宽阈值概率范围内,融合模型净获益高于临床模型和组学模型,提示其具有更好的临床应用潜能。见图 4A~4C、表 4。
图 4 融合模型、临床模型及组学模型在不同队列的表现Fig. 4 Performance of fusion model, clinical model, and ultrasomics model across different cohortsA: ROC curve in the training set; B: Calibration curve in the training set; C: DCA curve in the training set; D: ROC curve in the internal validation set; E: Calibration curve in the internal validation set; F: DCA curve in the internal validation set; G: ROC curve in the external validation set; H: Calibration curve in the external validation set; I: DCA curve in the external validation set. ROC: Receiver operating characteristic; DCA: Decision curve analysis.表 4 各模型在不同队列的预测效能比较Table 4 Comparison of predictive performance of each model across different cohortsCohort and model AUC (95%CI) Sensitivity Specificity Accuracy PPV NPV DeLong P valuea DeLong Padj valuea HL P value Training set Fusion model 0.845 (0.818, 0.872) 0.782 0.749 0.770 0.842 0.668 0.855 Ultrasomics model 0.774 (0.739, 0.808) 0.707 0.719 0.711 0.811 0.589 0.018 0.036 0.018 Clinical model 0.745 (0.710, 0.781) 0.646 0.753 0.685 0.817 0.554 0.001 0.002 0.022 Internal validation set Fusion model 0.829 (0.779, 0.878) 0.851 0.714 0.797 0.819 0.759 0.921 Ultrasomics model 0.788 (0.735, 0.840) 0.669 0.807 0.723 0.840 0.615 0.032 0.064 <0.001 Clinical model 0.724 (0.665, 0.784) 0.696 0.706 0.700 0.783 0.604 0.004 0.008 <0.001 External validation set Fusion model 0.854 (0.810, 0.898) 0.798 0.830 0.810 0.888 0.710 0.365 Ultrasomics model 0.775 (0.717, 0.834) 0.761 0.723 0.747 0.822 0.643 0.021 0.042 <0.001 Clinical model 0.784 (0.730, 0.838) 0.840 0.625 0.760 0.790 0.700 0.006 0.012 <0.001 Cohort and model Brier score Calibration intercept Calibration slope Mean bias Eavg Spiegelhalter Z test NRIb IDIb Z value P value Training set Fusion model 0.014 0.01 0.99 0.012 0.019 0.72 0.472 Ultrasomics model 0.243 0.10 0.85 0.041 0.075 2.80 0.005 -0.25 -0.26 Clinical model 0.242 0.12 0.82 0.049 0.083 2.95 0.003 -0.22 -0.22 Internal validation set Fusion model 0.010 0.02 0.98 0.015 0.031 0.41 0.682 Ultrasomics model 0.235 0.10 0.86 0.045 0.079 3.35 0.001 -0.07 -0.14 Clinical model 0.251 0.14 0.80 0.060 0.088 3.60 <0.001 -0.08 -0.14 External validation set Fusion model 0.020 0.01 1.02 0.018 0.035 1.05 0.294 Ultrasomics model 0.246 0.09 0.88 0.043 0.073 3.05 0.002 -0.09 -0.11 Clinical model 0.239 0.11 0.86 0.047 0.078 2.70 0.007 -0.11 -0.13 a: The P value for the AUC value was obtained by the DeLong test. b: The NRI and IDI results were obtained using the fusion model as the reference. AUC: Area under curve; 95%CI: 95% confidence interval; PPV: Positive predictive value; NPV: Negative predictive value; Padj: P corrected by Bonferroni method; HL: Hosmer-Lemeshow test; Eavg: Average absolute error; NRI: Net reclassification improvement; IDI: Integrated discrimination improvement. (2)在内部验证集中,融合模型对乳腺癌患者腋窝淋巴结转移仍保持较高区分度(AUC=0.829,95%CI 0.779~0.878),优于组学模型(AUC=0.788)与临床模型(AUC=0.724);经Bonferroni法校正后,融合模型与临床模型的AUC值比较差异有统计学意义(Padj=0.008),而与组学模型比较差异无统计学意义(Padj=0.064)。在校准度方面,融合模型的Hosmer-Lemeshow检验P=0.921,Brier分数为0.010;基于十分位分组的校准曲线关键数据中,融合模型的平均偏差为0.015,平均绝对偏差为0.031,校准截距为0.02,校准斜率为0.98,提示融合模型在内部验证集中校准度良好;组学模型和临床模型的Hosmer-Lemeshow检验P<0.001且Brier分数较高,平均绝对偏差分别为0.079与0.088,提示两者校准度不足。DCA显示,融合模型在主要阈值区间内可提供更高的临床净获益。见图 4D~4F、表 4。
(3)在外部验证集中,融合模型预测乳腺癌患者腋窝淋巴结转移的AUC值为0.854(95%CI0.810~0.898),优于组学模型(AUC=0.775)与临床模型(AUC=0.784);经Bonferroni法校正后,融合模型与组学模型、临床模型的AUC值比较差异均有统计学意义(Padj=0.042、0.012),提示融合模型在外部数据中具有稳健的泛化表现。在校准度方面,融合模型的Hosmer-Lemeshow检验P=0.365,Brier分数为0.020;基于十分位分组的校准曲线关键数据中,融合模型的平均偏差为0.018,平均绝对偏差为0.035,校准截距为0.01,校准斜率为1.02,提示融合模型在外部队列中的预测概率与实际发生率仍保持较好的一致性;组学模型和临床模型Hosmer-Lemeshow检验P<0.001且Brier分数较高,平均绝对偏差分别为0.073与0.078,提示两者校准度仍存在一定偏差。DCA显示融合模型在多数阈值区间内净获益高于组学模型和临床模型,且以融合模型为参照时组学模型和临床模型的NRI、IDI均为负值,提示融合模型可作为术前腋窝淋巴结风险分层与决策支持的潜在工具。见图 4G~4I、表 4。
2.6 融合模型与SLNB的效能对比
在同时接受SLNB的ALND的对照子集(n=142)中,以最终腋窝淋巴结术后病理结果为金标准,SLNB诊断乳腺癌患者腋窝淋巴结转移的灵敏度、特异度与准确度分别为0.914、1.000与0.965,融合模型的灵敏度、特异度与准确度分别为0.862、0.821与0.838。McNemar检验显示,融合模型在灵敏度方面与SLNB相比差异无统计学意义(P=0.286)。上述结果提示,融合模型在术前无创条件下可提供与SLNB相近的风险分层信息,具有作为术前筛选工具、辅助临床决策的潜在价值。见表 5。
表 5 融合模型与SLNB在对照子集中的诊断效能比较Table 5 Comparison of diagnostic performance between fusion model and SLNB in comparison subgroupDiagnostic metric Fusion model SLNB P value Formula, n/N Result (95%CI) Formula, n/N Result (95%CI) Sensitivity 50/58 0.862 (0.746, 0.939) 53/58 0.914 (0.810, 0.971) 0.286 Specificity 69/84 0.821 (0.723, 0.896) 84/84 1.000 (0.957, 1.000) <0.001 Accuracy (50+69)/142 0.838 (0.767, 0.894) (53+84)/142 0.965 (0.920, 0.989) <0.001 PPV 50/65 0.769 (0.648, 0.865) 53/53 1.000 (0.933, 1.000) <0.001 NPV 69/77 0.896 (0.806, 0.954) 84/89 0.944 (0.875, 0.982) <0.001 SLNB: Sentinel lymph node biopsy; 95%CI: 95% confidence interval; PPV: Positive predictive value; NPV: Negative predictive value. 2.7 融合模型的临床阈值与风险分层表现
在外部验证集中,基于约登指数确定融合模型用于二分类(腋窝淋巴结转移与否)判定的最佳截断值为0.29。为满足术前分流管理需求,进一步结合DCA的净获益区间设定风险分层阈值:低危(预测概率<0.15)、中危(0.15~0.45)与高危(>0.45)。在外部验证集中,低危、中危与高危组分别占33.0%(30/91)、47.3%(43/91)与19.8%(18/91),腋窝淋巴结转移阳性率分别为16.7%(5/30)、41.9%(18/43)与77.8%(14/18),呈现随风险分层递增的清晰梯度。提示融合模型可用于术前风险分层与分流管理。
2.8 融合模型亚组分析
在外部验证集中,按乳腺癌分子分型与临床T分期进行亚组分析。按乳腺癌分子分型分层,融合模型在Luminal、HER2阳性与TNBC亚组中预测腋窝淋巴结转移的AUC值分别为0.838、0.914与0.864;按临床T分期分层,融合模型在T1、T2与T3亚组的AUC值分别为0.848、0.873与0.700。总体而言,多数亚组AUC值处于可接受范围,但样本量较小的亚组(如T3亚组)95%CI较宽(95%CI 0.250~1.000),提示预测不确定性较高。见表 6。
表 6 融合模型在外部验证集不同亚组中的判别效能Table 6 Discriminative performance of fusion model across subgroups in external validation setSubgroup N ALNM+, n AUC (95%CI) Molecular subtype Luminal 55 19 0.838 (0.731, 0.933) HER2+ 18 9 0.914 (0.740, 1.000) TNBC 18 9 0.864 (0.642, 1.000) Clinical T stage T1 36 9 0.848 (0.691, 0.959) T2 46 24 0.873 (0.758, 0.970) T3 9 4 0.700 (0.250, 1.000) ALNM+: Positive axillary lymph node metastasis; AUC: Area under curve; 95%CI: 95% confidence interval; HER2+: Positive human epidermal growth factor receptor 2; TNBC: Triple-negative breast cancer. 3 讨论
本研究开发并验证了一个基于超声组学与临床和常规超声特征的融合预测模型,用于术前评估乳腺癌腋窝淋巴结转移风险。本研究的优势主要体现在:(1)采用双中心队列,引入独立外部验证集,系统评估了模型的跨中心泛化能力;(2)在IBSI标准化预处理与稳定性筛选基础上构建了可复用的超声组学评分,并与可获得的临床和常规超声变量融合,兼顾性能与可解释性;(3)除区分度外,同时报告了校准度与临床净获益,使模型的“能用性”评价更贴近真实临床场景。在模型性能方面,融合模型在内部与外部验证集中均表现出较高的区分度,整体优于单纯组学模型与临床模型。同时,融合模型在不同队列中显示出较好的校准一致性与较低的预测误差,校准曲线更接近理想参考线。DCA进一步提示融合模型在较宽的阈值范围内可获得更高的净获益,表明其有望用于术前腋窝淋巴结风险分层与决策支持,减少过度处理与漏治风险。
在关键特征方面,本研究筛选出的临床和常规超声变量包括淋巴结PSV/EDV比值、肿瘤直径、淋巴结皮髓质面积比值和肿瘤阻力指数。淋巴结PSV/EDV比值量化淋巴结内部回声的异质性,主要反映血流信号分布的变异性,常与淋巴结微循环异常相关[9];肿瘤直径直接衡量肿瘤大小,作为T分期的核心指标;淋巴结皮髓质面积比值评估淋巴结皮质相对于髓质的相对厚度,测量值升高提示淋巴结结构重塑;肿瘤阻力指数则通过计算PSV与EDV的差值和PSV之比表征肿瘤血管阻力的变化,其升高表示高阻力血流模式[8]。SHAP分析进一步证实,这些特征呈正向风险驱动,即淋巴结PSV/EDV比值增大、肿瘤直径增加、淋巴结皮髓质面积比值升高和肿瘤阻力指数升高均显著提升腋窝淋巴结转移风险。本研究采用LASSO回归降维筛选确定了用于超声组学评分构建的超声组学特征,并在外部队列验证以降低过拟合风险。组学参数中,18个非零系数特征以纹理和形状为主,其中GLCM_JointEntropy(联合熵)反映灰度共生矩阵的纹理复杂性,量化肿瘤内部像素分布的随机性和异质性;Shape_Sphericity(球形度)则评估肿瘤三维形状的球形程度,低值表示边界不规则[10],这些特征权重最高,突显了影像学在捕捉肿瘤微观异质性方面的价值。
上述特征的发现为理解腋窝淋巴结转移机制提供了更全面的视角,与既有研究在一定程度上一致,但也存在细微差异。肿瘤直径作为独立预测因子在本研究模型中的贡献突出,与Qiu等[11]的超声模型高度一致,后者报道肿瘤大小是腋窝淋巴结转移的强相关因素(OR=1.45,P<0.01),因为较大的肿瘤往往伴随更高的侵袭潜力。淋巴结皮髓质面积比值升高提示淋巴结皮质增厚,反映了转移的肿瘤细胞浸润导致淋巴结结构发生改变,Zhang等[4]的多中心研究也表明淋巴结皮髓质面积比值>2.0与腋窝淋巴结转移相关(灵敏度为85%)。然而,本研究中淋巴结PSV/EDV比值的突出作用(SHAP值最高)与既往研究有所不同,例如,Zhang等[8]的列线图模型更强调搏动指数和阻力指数等血流参数,但未突出淋巴结PSV/EDV比值的异质性,这可能源于其基于2D静态图像分析,而本研究整合了频谱血流图像评估以捕捉动态变异。肿瘤阻力指数升高在本研究中也被确认为腋窝淋巴结转移的危险因素,与Guo等[12]的早期超声特征研究结果一致(阻力指数>0.7与腋窝淋巴结转移相关,AUC值为0.76),但与部分MRI研究形成对比,如Dong等[13]报告低阻力指数者更易发生腋窝淋巴结转移,可能因为MRI更侧重宏观血管网络,而超声聚焦于微循环异质性。在组学特征方面,联合熵和球形度的关键作用与Han等[14]的列线图模型一致,该模型中灰度共生矩阵熵是显著的纹理预测因子(AUC值提升0.08),反映了肿瘤微环境异质性促进淋巴播散;球形度低值与腋窝淋巴结转移相关,也与Yu等[15]的超声组学列线图相似(球形度<0.8预测腋窝淋巴结转移的OR为2.1),提示不规则形态增强了肿瘤的侵袭性。然而,与深度学习组学模型[6]比较,本研究中纹理特征的权重高于形状特征,这可能归因于超声的高分辨率更易捕捉细微纹理细节,而深度学习模型偏好全局形态提取。总体差异的原因可能包括不同成像模态的灵敏性差异、ROI勾画方法及标准化流程不同、特征筛选与模型融合策略差异,以及样本来源和病例构成不同等[5, 16]。
从机制和底层逻辑层面,这些因素与腋窝淋巴结转移的相关性根植于肿瘤微环境和转移“种子-土壤”模型[17]。淋巴结PSV/EDV比值增大反映淋巴结血流异常,可能由转移的肿瘤细胞诱导的新生血管异常引起[18],这促进了淋巴管渗透和癌细胞播散。肿瘤直径增加提示可能存在更多“种子”细胞,通过上调血管内皮生长因子通路激活淋巴生成[19],从而扩大转移窗口。淋巴结皮髓质面积比值升高提示肿瘤细胞优先浸润淋巴结皮质区,破坏正常门结构,类似于“土壤”准备阶段[20],这与淋巴结免疫逃逸相关。肿瘤阻力指数升高提示高阻力微血管形成,反映上皮-间质转化增强,导致肿瘤侵袭性和血管生成增加[21]。在组学层面,联合熵高值反映肿瘤微环境异质性,如免疫细胞浸润和基质重塑[22],且往往与Ki-67增殖指数相关,促进淋巴结转移[23];低球形度则关联上皮-间质转化诱导的形态不规则,增加肿瘤细胞的迁移和附着能力[24]。这些机制解释了为什么肿瘤异质性和形态改变是腋窝淋巴结转移的核心驱动因素,融合临床和组学特征能够从宏观(直径、阻力指数)到微观(纹理、形状)层面捕捉生物学过程信息,从而提升模型的可解释性和临床相关性。
本研究融合模型的整体性能(AUC=0.854)在外部验证集中优于对照模型(临床模型和组学模型),提示该融合模型借助glmnet算法和多模态参数融合进一步提升了泛化能力;相较于Zha等[25]的前哨淋巴结转移预测模型(AUC值为0.80),本研究的融合模型覆盖了全腋窝淋巴结范围并进行了外部验证,增强了鲁棒性。多中心设计可有效减少过拟合风险,本研究融合模型的性能虽然与Zhang等[4]的6家医院深度学习组学模型(AUC值为0.87~0.90)相近,但本研究采用传统超声而非深度学习,降低了计算成本并提高了可及性[26]。基于SHAP和LASSO回归筛选的18个组学特征(以纹理为主)与Han等[14]的列线图一致,强调熵和形状的作用,但本研究的SHAP重要性蜂群图更直观地揭示了各组学特征的权重。在机器学习算法比较中,glmnet在ROC曲线分析中的AUC值达0.979 0,优于支持向量机、RF和XGBoost,这证明了glmnet在中等样本(321例)条件下的适用性和稳健性。这些结果表明,本研究融合模型在性能和可解释性上均优于单一模态或传统方法,为超声组学在腋窝淋巴结转移预测中的应用提供了依据。
在临床意义方面,本研究融合模型可指导SLNB、ALND决策,有助于减少不必要手术,特别是在概率阈值0.03~0.59范围内净获益高于对照方案。该融合模型较既往模型[11, 25]AUC值提升了约0.015,且通过LASSO回归减少了共线性,进一步强化了跨中心泛化能力。基于本研究分析,融合模型可在术前提供无创且与SLNB相近的风险分层能力,其更适合定位为SLNB的术前筛选与分流工具:对于模型预测为腋窝淋巴结转移低风险的患者,可在多学科讨论基础上减少不必要的SLNB;对于模型提示高风险者,则可强化术中腋窝淋巴结评估或优化手术策略。需要强调的是,该融合模型旨在辅助决策与风险提示,并不能在当前证据水平下直接替代SLNB,未来仍需前瞻性、多中心研究进一步验证其对临床结局与手术决策的真实影响。
本研究的优势在于回顾性多中心设计、IBSI标准化预处理、16种机器学习算法的全面比较和多维度验证(ROC曲线、Hosmer-Lemeshow检验、Brier分数、DCA、NRI、IDI等),同时基于PyRadiomics 3.0.1软件提取的约1 000个特征经ICC(≥0.75)筛选确保了稳定性。然而也存在局限性:(1)样本量中等,需前瞻性大样本试验验证以排除回顾性偏差;(2)模型未纳入HER2、Ki-67等分子标志物,可能限制了其在分子层面的稳定性;(3)超声设备统一(Mindray R9),未来需通过多设备测试以进一步提高泛化能力;(4)未覆盖新辅助治疗人群,限制了适用范围。
总之,本研究建立的融合模型作为可靠的非侵入工具,对于乳腺癌患者具有较好的术前腋窝淋巴结转移风险预测能力,有助于指导精准医疗。未来研究方向包括整合深度学习(如DenseNet)或多模态(超声+MRI)参数以进一步提升模型性能,开展前瞻性多中心试验验证模型的临床影响,以及探索超声组学评分与肿瘤微环境的分子关联以推动从影像学向机制转化的研究范式。
-
图 2 高风险与低风险腋窝淋巴结转移的代表性乳腺癌病例的超声图像
Fig. 2 Ultrasound images of representative breast cancer patients with high- or low-risk axillary lymph node metastasis
A: A high-risk (axillary lymph node metastasis positive) case. The left panels showed ultrasonography and spectral Doppler imaging of the primary breast tumor, with a maximum tumor diameter >3.0 cm and a tumor resistive index >0.80. The right panels showed ultrasonography and spectral Doppler imaging of the ipsilateral axillary lymph node, with a lymph node cortex-to-medulla area ratio >5.0and a lymph node PSV/EDV ratio of approximately 4.8. Postoperative pathology confirmed axillary lymph node metastasis. B: A low-risk(axillary lymph node metastasis negative) case. The left panels showed ultrasonography and spectral Doppler imaging of the primary breast tumor, with a tumor diameter of approximately 1.3 cm and a tumor resistive index of approximately 0.64. The right panels showed ultrasonography and spectral Doppler imaging of the ipsilateral axillary lymph node, with a lymph node cortex-to-medulla area ratio of approximately 1.1 and a lymph node PSV/EDV ratio of approximately 2.2. Postoperative pathology confirmed no axillary lymph node metastasis. PSV: Peak systolic velocity; EDV: End-diastolic velocity.
图 3 超声组学特征筛选及重要性分析
Fig. 3 Ultrasomics feature selection and importance analysis
A: LASSO regression coefficient path plot. The plot showed the convergence trajectories of the partial regression coefficients of the 18 ultrasomics features as λ changes. As λ increased, the coefficients of redundant features were gradually shrunk to 0. B: Ten-fold cross-validation AUC curve. The 2 vertical dashed lines indicated the optimal λ for maximum AUC (right) and the λ under 1-standard error rule (left). C: SHAP importance beeswarm plot. The plot showed the direction and magnitude of the contributions of the included features to the predicted probability of axillary lymph node metastasis. LASSO: Least absolute shrinkage and selection operator; AUC: Area under area; SHAP: Shapley additive explanation.
图 4 融合模型、临床模型及组学模型在不同队列的表现
Fig. 4 Performance of fusion model, clinical model, and ultrasomics model across different cohorts
A: ROC curve in the training set; B: Calibration curve in the training set; C: DCA curve in the training set; D: ROC curve in the internal validation set; E: Calibration curve in the internal validation set; F: DCA curve in the internal validation set; G: ROC curve in the external validation set; H: Calibration curve in the external validation set; I: DCA curve in the external validation set. ROC: Receiver operating characteristic; DCA: Decision curve analysis.
表 1 3个队列乳腺癌患者的临床及常规超声特征比较
Table 1 Comparison of clinical and conventional ultrasound characteristics of breast cancer patients across 3 cohorts
Variable Training set N=161 Internal validation set N=69 External validation set N=91 Statistic P value Age/year, $\bar{x} \pm s$ 49.53±8.76 49.29±8.47 50.58±8.50 F=0.570 0.566 Lymph node metastasis, n (%) χ2=1.994 0.369 No 94 (58.4) 34 (49.3) 54 (59.3) Yes 67 (41.6) 35 (50.7) 37 (40.7) Menopausal status, n (%) χ2=2.469 0.291 No 82 (50.9) 33 (47.8) 37 (40.7) Yes 79 (49.1) 36 (52.2) 54 (59.3) Tumor Location, n (%) 0.550a Upper outer 41 (25.5) 20 (29.0) 26 (28.6) Lower outer 26 (16.1) 7 (10.1) 10 (11.0) Lower inner 34 (21.1) 22 (31.9) 26 (28.6) Upper inner 49 (30.4) 16 (23.2) 21 (23.1) Areolar region 11 (6.8) 4 (5.8) 8 (8.8) Regular shape, n (%) χ2=0.053 0.974 Yes 82 (50.9) 34 (49.3) 46 (50.5) No 79 (49.1) 35 (50.7) 45 (49.5) Clear margin, n (%) χ2=0.825 0.662 Yes 82 (50.9) 33 (47.8) 41 (45.1) No 79 (49.1) 36 (52.2) 50 (54.9) Homo-echo, n (%) χ2=4.901 0.086 Yes 79 (49.1) 41 (59.4) 38 (41.8) No 82 (50.9) 28 (40.6) 53 (58.2) Calcification, n (%) χ2=0.572 0.751 Yes 83 (51.6) 33 (47.8) 49 (53.8) No 78 (48.4) 36 (52.2) 42 (46.2) Diameter/mm, M (Q1, Q3) 12.80 (10.01, 16.37) 13.98 (10.90, 17.94) 13.60 (10.67, 17.34) H=3.219 0.200 PSV/(cm·s-1), $\bar{x} \pm s$ 37.50±6.26 39.05±5.70 38.88±6.28 F=2.251 0.107 EDV/(cm·s-1), $\bar{x} \pm s$ 9.94±2.00 10.20±1.92 10.21±2.42 F=0.627 0.535 PSV/EDV ratio, M (Q1, Q3) 3.78 (3.10, 4.61) 3.85 (3.21, 4.63) 3.88 (3.25, 4.62) H=0.165 0.921 Resistive index, $\bar{x} \pm s$ 0.57±0.19 0.59±0.19 0.60±0.20 F=0.490 0.613 Pulsatility index, $\bar{x} \pm s$ 1.76±0.31 1.84±0.26 1.84±0.28 F=2.789 0.063 Lymph node Short diameter/mm, $\bar{x} \pm s$ 5.91±1.34 5.90±1.21 5.93±1.24 F=0.016 0.984 Long diameter/mm, $\bar{x} \pm s$ 11.57±2.02 11.51±2.08 11.77±2.05 F=0.394 0.675 CMR, $\bar{x} \pm s$ 1.93±0.30 1.95±0.27 1.89±0.26 F=1.128 0.325 PSV/(cm·s-1), M (Q1, Q3) 16.96 (13.69, 21.02) 15.58 (12.13, 20.03) 16.23 (12.69, 20.76) H=1.683 0.431 EDV/(cm·s-1), M (Q1, Q3) 8.20 (6.89, 9.74) 7.69 (6.57, 8.99) 7.14 (6.06, 8.41) H=2.694 0.260 PSV/EDV ratio, M (Q1, Q3) 2.57 (1.79, 3.71) 3.19 (2.53, 4.01) 3.00 (2.13, 4.22) H=5.077 0.079 Resistive index, $\bar{x} \pm s$ 0.60±0.19 0.62±0.20 0.60±0.18 F=0.232 0.793 Pulsatility index, $\bar{x} \pm s$ 1.69±0.31 1.67±0.26 1.67±0.27 F=0.238 0.788 a: Fisher exact test. Homo-echo: Homogeneous echogenicity; PSV: Peak systolic velocity; EDV: End-diastolic velocity; CMR: Cortex-to-medulla area ratio. 表 2 3个队列中ALNM+与ALNM-组乳腺癌患者常规超声特征比较
Table 2 Comparison in conventional ultrasound characteristics of breast cancer patients between ALNM+ and ALNM- groups in 3 cohorts
Variable Training set Internal validation set External validation set ALNM-N=94 ALNM+N=67 P value ALNM-N=34 ALNM+N=35 Pvalue ALNM-N=54 ALNM+N=37 P value Tumor Irregular shape, n (%) 40 (42.6) 39 (58.2) 0.072 15 (44.1) 20 (57.1) 0.400 24 (44.4) 21 (56.8) 0.347 Unclear margin, n (%) 40 (42.6) 39 (58.2) 0.072 16 (47.1) 20 (57.1) 0.550 29 (53.7) 21 (56.8) 0.942 Het-echo, n (%) 43 (45.7) 39 (58.2) 0.162 8 (23.5) 20 (57.1) 0.009 32 (59.3) 21 (56.8) 0.983 Calcification, n (%) 41 (43.6) 42 (62.7) 0.026 11 (32.4) 22 (62.9) 0.022 26 (48.1) 23 (62.2) 0.270 Diameter/mm, $\bar{x} \pm s$ 11.54±5.11 16.68±5.29 <0.001 11.88±5.65 17.97±5.85 <0.001 12.45±5.31 17.51±5.56 <0.001 PSV/ (cm·s-1), $\bar{x} \pm s$ 36.64±6.20 38.70±6.39 0.044 37.81±5.63 40.25±5.92 0.085 38.06±6.17 40.08±6.46 0.139 EDV/ (cm·s-1), $\bar{x} \pm s$ 9.80±1.99 10.14±2.06 0.293 9.99±1.92 10.40±1.99 0.392 10.07±2.39 10.41±2.48 0.520 PSV/EDV ratio, $\bar{x} \pm s$ 3.84±1.19 4.10±1.22 0.186 3.85±1.10 4.15±1.15 0.265 3.91±1.04 4.16±1.10 0.274 Resistive index, $\bar{x} \pm s$ 0.53±0.19 0.63±0.20 0.001 0.53±0.19 0.65±0.20 0.011 0.56±0.20 0.66±0.21 0.023 Pulsatility index, $\bar{x} \pm s$ 1.69±0.31 1.86±0.32 0.001 1.74±0.26 1.94±0.27 0.002 1.77±0.28 1.94±0.29 0.007 Lymph node Short diameter/mm, $\bar{x} \pm s$ 5.27±1.33 6.81±1.39 <0.001 4.97±1.21 6.80±1.24 <0.001 5.31±1.24 6.83±1.27 <0.001 Long diameter/mm, $\bar{x} \pm s$ 11.36±1.98 11.87±2.07 0.117 11.20±2.04 11.81±2.16 0.232 11.56±2.04 12.07±2.10 0.257 CMR, $\bar{x} \pm s$ 1.75±0.30 2.18±0.31 <0.001 1.69±0.27 2.20±0.28 <0.001 1.72±0.26 2.14±0.27 <0.001 PSV/ (cm·s-1), $\bar{x} \pm s$ 17.13±5.76 18.84±5.97 0.071 15.67±6.33 17.70±6.66 0.199 16.66±6.52 18.35±6.78 0.241 EDV/ (cm·s-1), $\bar{x} \pm s$ 8.68±2.20 8.17±2.28 0.154 8.21±1.84 7.60±1.92 0.183 7.56±1.77 7.05±1.85 0.197 PSV/EDV ratio, $\bar{x} \pm s$ 2.66±1.73 3.43±1.79 0.007 2.92±1.17 3.83±1.22 0.002 3.10±1.82 3.86±1.90 0.061 Resistive index, $\bar{x} \pm s$ 0.56±0.19 0.65±0.20 0.006 0.57±0.20 0.67±0.21 0.040 0.57±0.18 0.65±0.18 0.034 Pulsatility index, $\bar{x} \pm s$ 1.65±0.31 1.74±0.32 0.089 1.62±0.26 1.72±0.27 0.113 1.64±0.27 1.72±0.28 0.152 ALNM+: Positive axillary lymph node metastasis; ALNM-: Negative axillary lymph node metastasis; Het-echo: Heterogeneous echogenicity; PSV: Peak systolic velocity; EDV: End-diastolic velocity; CMR: Cortex-to-medulla area ratio. 表 3 不同机器学习算法模型在训练集中的效能分析
Table 3 Performance of different machine learning algorithm models in training set
Algorithm AUC Sensitivity Specificity Accuracy PPV NPV F1 score glmnet 0.979 0 0.970 1 0.989 4 0.981 4 0.984 8 0.978 9 0.977 4 NNet 0.976 7 0.970 1 0.989 4 0.981 4 0.984 8 0.978 9 0.977 4 svmLinear 0.972 5 0.970 1 0.989 4 0.981 4 0.984 8 0.978 9 0.977 4 XGBoost 0.970 2 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 C5.0 0.969 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 GBM 0.967 7 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 FDA 0.963 1 0.940 3 0.989 4 0.968 9 0.984 4 0.958 8 0.961 8 AdaBoost 0.960 5 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 LDA 0.957 2 0.910 4 0.968 1 0.944 1 0.953 1 0.938 1 0.931 3 RF 0.952 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 1.000 0 MDA 0.938 3 0.910 4 0.968 1 0.944 1 0.953 1 0.938 1 0.931 3 svmRadial 0.934 9 0.955 2 0.978 7 0.968 9 0.969 7 0.968 4 0.962 4 GLM 0.922 2 0.970 1 0.989 4 0.981 4 0.984 8 0.978 9 0.977 4 naive Bayes 0.913 9 0.895 5 0.968 1 0.937 9 0.952 4 0.928 6 0.923 1 kNN 0.884 8 0.850 7 0.946 8 0.906 8 0.919 4 0.899 0 0.883 7 rpart 0.867 4 0.865 7 0.946 8 0.913 0 0.920 6 0.908 2 0.892 3 glmnet: Generalized linear model via elastic-net regularization; NNet: Neural network; svmLinear: Support vector machine with a linear kernel; XGBoost: Extreme gradient boosting; C5.0: C5.0 decision tree; GBM: Gradient boosting machine; FDA: Flexible discriminant analysis; AdaBoost: Adaptive boosting; LDA: Linear discriminant analysis; RF: Random forest; MDA: Mixture discriminant analysis; svmRadial: Support vector machine with a radial basis function kernel; GLM: Generalized linear model; kNN: K-nearest neighbor; rpart: Recursive partitioning and regression tree; AUC: Area under curve; PPV: Positive predictive value; NPV: Negative predictive value. 表 4 各模型在不同队列的预测效能比较
Table 4 Comparison of predictive performance of each model across different cohorts
Cohort and model AUC (95%CI) Sensitivity Specificity Accuracy PPV NPV DeLong P valuea DeLong Padj valuea HL P value Training set Fusion model 0.845 (0.818, 0.872) 0.782 0.749 0.770 0.842 0.668 0.855 Ultrasomics model 0.774 (0.739, 0.808) 0.707 0.719 0.711 0.811 0.589 0.018 0.036 0.018 Clinical model 0.745 (0.710, 0.781) 0.646 0.753 0.685 0.817 0.554 0.001 0.002 0.022 Internal validation set Fusion model 0.829 (0.779, 0.878) 0.851 0.714 0.797 0.819 0.759 0.921 Ultrasomics model 0.788 (0.735, 0.840) 0.669 0.807 0.723 0.840 0.615 0.032 0.064 <0.001 Clinical model 0.724 (0.665, 0.784) 0.696 0.706 0.700 0.783 0.604 0.004 0.008 <0.001 External validation set Fusion model 0.854 (0.810, 0.898) 0.798 0.830 0.810 0.888 0.710 0.365 Ultrasomics model 0.775 (0.717, 0.834) 0.761 0.723 0.747 0.822 0.643 0.021 0.042 <0.001 Clinical model 0.784 (0.730, 0.838) 0.840 0.625 0.760 0.790 0.700 0.006 0.012 <0.001 Cohort and model Brier score Calibration intercept Calibration slope Mean bias Eavg Spiegelhalter Z test NRIb IDIb Z value P value Training set Fusion model 0.014 0.01 0.99 0.012 0.019 0.72 0.472 Ultrasomics model 0.243 0.10 0.85 0.041 0.075 2.80 0.005 -0.25 -0.26 Clinical model 0.242 0.12 0.82 0.049 0.083 2.95 0.003 -0.22 -0.22 Internal validation set Fusion model 0.010 0.02 0.98 0.015 0.031 0.41 0.682 Ultrasomics model 0.235 0.10 0.86 0.045 0.079 3.35 0.001 -0.07 -0.14 Clinical model 0.251 0.14 0.80 0.060 0.088 3.60 <0.001 -0.08 -0.14 External validation set Fusion model 0.020 0.01 1.02 0.018 0.035 1.05 0.294 Ultrasomics model 0.246 0.09 0.88 0.043 0.073 3.05 0.002 -0.09 -0.11 Clinical model 0.239 0.11 0.86 0.047 0.078 2.70 0.007 -0.11 -0.13 a: The P value for the AUC value was obtained by the DeLong test. b: The NRI and IDI results were obtained using the fusion model as the reference. AUC: Area under curve; 95%CI: 95% confidence interval; PPV: Positive predictive value; NPV: Negative predictive value; Padj: P corrected by Bonferroni method; HL: Hosmer-Lemeshow test; Eavg: Average absolute error; NRI: Net reclassification improvement; IDI: Integrated discrimination improvement. 表 5 融合模型与SLNB在对照子集中的诊断效能比较
Table 5 Comparison of diagnostic performance between fusion model and SLNB in comparison subgroup
Diagnostic metric Fusion model SLNB P value Formula, n/N Result (95%CI) Formula, n/N Result (95%CI) Sensitivity 50/58 0.862 (0.746, 0.939) 53/58 0.914 (0.810, 0.971) 0.286 Specificity 69/84 0.821 (0.723, 0.896) 84/84 1.000 (0.957, 1.000) <0.001 Accuracy (50+69)/142 0.838 (0.767, 0.894) (53+84)/142 0.965 (0.920, 0.989) <0.001 PPV 50/65 0.769 (0.648, 0.865) 53/53 1.000 (0.933, 1.000) <0.001 NPV 69/77 0.896 (0.806, 0.954) 84/89 0.944 (0.875, 0.982) <0.001 SLNB: Sentinel lymph node biopsy; 95%CI: 95% confidence interval; PPV: Positive predictive value; NPV: Negative predictive value. 表 6 融合模型在外部验证集不同亚组中的判别效能
Table 6 Discriminative performance of fusion model across subgroups in external validation set
Subgroup N ALNM+, n AUC (95%CI) Molecular subtype Luminal 55 19 0.838 (0.731, 0.933) HER2+ 18 9 0.914 (0.740, 1.000) TNBC 18 9 0.864 (0.642, 1.000) Clinical T stage T1 36 9 0.848 (0.691, 0.959) T2 46 24 0.873 (0.758, 0.970) T3 9 4 0.700 (0.250, 1.000) ALNM+: Positive axillary lymph node metastasis; AUC: Area under curve; 95%CI: 95% confidence interval; HER2+: Positive human epidermal growth factor receptor 2; TNBC: Triple-negative breast cancer. -
[1] Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022:GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries[J]. CA Cancer J Clin, 2024, 74(3):229-263. DOI: 10.3322/caac.21834. [2] Giuliano A E, Ballman K V, McCall L, et al. Effect of axillary dissection vs no axillary dissection on 10-year overall survival among women with invasive breast cancer and sentinel node metastasis: the ACOSOG Z0011(Alliance) randomized clinical trial[J]. JAMA, 2017, 318(10): 918-926. DOI: 10.1001/jama.2017.11470. [3] Zhao D, Xu M, Yang S, et al. Specific diagnosis of lymph node micrometastasis in breast cancer by targeting activatable near-infrared fluorescence imaging[J]. Biomaterials, 2022, 282: 121388. DOI: 10.1016/j.biomaterials.2022.121388. [4] Zhang D, Zhou W, Lu W W, et al. Ultrasound-based deep learning radiomics for enhanced axillary lymph node metastasis assessment: a multicenter study[J]. Oncologist, 2025, 30(5): oyaf090. DOI: 10.1093/oncolo/oyaf090. [5] Jiang M, Li C L, Luo X M, et al. Radiomics model based on shear-wave elastography in the assessment of axillary lymph node status in early-stage breast cancer[J]. Eur Radiol, 2022, 32(4): 2313-2325. DOI: 10.1007/s00330-021-08330-w. [6] Liu H, Zou L, Xu N, et al. Deep learning radiomics based prediction of axillary lymph node metastasis in breast cancer[J]. NPJ Breast Cancer, 2024, 10: 22. DOI: 10.1038/s41523-024-00628-4. [7] Yang Z Q, Zhang Y, Lu F, et al. Integrating multimodal ultrasound imaging for improved radiomics sentinel lymph node assessment in breast cancer[J]. Gland Surg, 2025, 14(7): 1348-1365. DOI: 10.21037/gs-2025-223. [8] Zhang W, Wang S, Wang Y, et al. Ultrasound-based radiomics nomogram for predicting axillary lymph node metastasis in early-stage breast cancer[J]. Radiol Med, 2024, 129(2): 211-221. DOI: 10.1007/s11547-024-01768-0. [9] Hu Z, Tian Z, Wei X, et al. Machine learning model for predicting axillary lymph node metastasis in clinically node positive breast cancer based on peritumoral ultrasound radiomics and SHAP feature analysis[J]. J Ultrasound Med, 2024, 43(10): 2007-2008. DOI: 10.1002/jum.16520. [10] Chen Y, Li J, Zhang J, et al. Radiomic nomogram for predicting axillary lymph node metastasis in patients with breast cancer[J]. Acad Radiol, 2024, 31(3): 788-799. DOI: 10.1016/j.acra.2023.10.026. [11] Qiu X, Jiang Y, Zhao Q, et al. Could ultrasound-based radiomics noninvasively predict axillary lymph node metastasis in breast cancer?[J]. J Ultrasound Med, 2020, 39(10): 1897-1905. DOI: 10.1002/jum.15294. [12] Guo Q, Dong Z, Zhang L, et al. Ultrasound features of breast cancer for predicting axillary lymph node metastasis[J]. J Ultrasound Med, 2018, 37(6): 1345-1353. DOI: 10.1002/jum.14469. [13] Dong Y, Feng Q, Yang W, et al. Preoperative prediction of sentinel lymph node metastasis in breast cancer based on radiomics of T2-weighted fat-suppression and diffusion-weighted MRI[J]. Eur Radiol, 2018, 28(2): 582-591. DOI: 10.1007/s00330-017-5005-7. [14] Han L, Zhu Y, Liu Z, et al. Radiomic nomogram for prediction of axillary lymph node metastasis in breast cancer[J]. Eur Radiol, 2019, 29(7): 3820-3829. DOI: 10.1007/s00330-018-5981-2. [15] Yu F H, Wang J X, Ye X H, et al. Ultrasound-based radiomics nomogram: a potential biomarker to predict axillary lymph node metastasis in early-stage invasive breast cancer[J]. Eur J Radiol, 2019, 119: 108658. DOI: 10.1016/j.ejrad.2019.108658. [16] Li X, Yang L, Jiao X. Comparison of traditional radiomics, deep learning radiomics and fusion methods for axillary lymph node metastasis prediction in breast cancer[J]. Acad Radiol, 2023, 30(7): 1281-1287. DOI: 10.1016/j.acra.2022.10.015. [17] Leong S P, Naxerova K, Keller L, et al. Molecular mechanisms of cancer metastasis via the lymphatic versus the blood vessels[J]. Clin Exp Metastasis, 2022, 39(1): 159-179. DOI: 10.1007/s10585-021-10120-z. [18] Ferroni G, Sabeti S, Abdus-Shakur T, et al. Noninvasive prediction of axillary lymph node breast cancer metastasis using morphometric analysis of nodal tumor microvessels in a contrast-free ultrasound approach[J]. Breast Cancer Res, 2023, 25(1): 65. DOI: 10.1186/s13058-023-01670-z. [19] Gao Y, Luo Y, Zhao C, et al. Nomogram based on radiomics analysis of primary breast cancer ultrasound images: prediction of axillary lymph node tumor burden in patients[J]. Eur Radiol, 2021, 31(2): 928-937. DOI: 10.1007/s00330-020-07181-1. [20] Wang L, Li J, Qiao J, et al. Establishment of a model for predicting sentinel lymph node metastasis in early breast cancer based on contrast-enhanced ultrasound and clinicopathological features[J]. Gland Surg, 2021, 10(5): 1701-1712. DOI: 10.21037/gs-21-245. [21] Yajima R, Fujii T, Yanagita Y, et al. Prognostic value of extracapsular invasion of axillary lymph nodes combined with peritumoral vascular invasion in patients with breast cancer[J]. Ann Surg Oncol, 2015, 22(1): 52-58. DOI: 10.1245/s10434-014-3941-x. [22] Sun R, Limkin E J, Vakalopoulou M, et al. A radiomics approach to assess tumour-infiltrating CD8 cells and response to anti-PD-1 or anti-PD-L1 immunotherapy: an imaging biomarker, retrospective multicohort study[J]. Lancet Oncol, 2018, 19(9): 1180-1191. DOI: 10.1016/S1470-2045(18)30413-3. [23] Yu Y, He Z, Ouyang J, et al. Magnetic resonance imaging radiomics predicts preoperative axillary lymph node metastasis to support surgical decisions and is associated with tumor microenvironment in invasive breast cancer: a machine learning, multicenter study[J]. EBioMedicine, 2021, 69: 103460. DOI: 10.1016/j.ebiom.2021.103460. [24] Saha A, Harowicz M R, Grimm L J, et al. A machine learning approach to radiogenomics of breast cancer: a study of 922 subjects and 529 DCE-MRI features[J]. Br J Cancer, 2018, 119(4): 508-516. DOI: 10.1038/s41416-018-0185-8. [25] Zha H L, Zong M, Liu X P, et al. Preoperative ultrasound-based radiomics score can improve the accuracy of the Memorial Sloan Kettering Cancer Center nomogram for predicting sentinel lymph node metastasis in breast cancer[J]. Eur J Radiol, 2021, 135: 109512. DOI: 10.1016/j.ejrad.2020.109512. [26] Gu J, Tong T, Xu D, et al. Deep learning radiomics of ultrasonography for comprehensively predicting tumor and axillary lymph node status after neoadjuvant chemotherapy in breast cancer patients: a multicenter study[J]. Cancer, 2023, 129(3): 356-366. DOI: 10.1002/cncr.34540.
下载: