探讨基于新辅助化疗（NAC）前及2周期NAC后常规超声图像的瘤内、瘤周影像组学特征联合临床病理因素构建的列线图模型，对乳腺癌NAC后病理完全缓解（pCR）的预测价值。​

回顾性收集2019年7月1日至2023年12月1日南方医科大学第十附属医院332例经病理证实、并接受标准6周期NAC的女性乳腺癌患者资料。所有患者均在NAC治疗前及第2周期NAC治疗后进行了超声检查，用于本次疗效的早期预测研究。将患者按7∶3的比例随机分为训练集（233例）和验证集（99例）。在NAC前及2周期NAC后超声图像上勾画瘤内区域及瘤周1 mm、2 mm、3 mm区域，提取影像组学特征，通过组内相关系数（ICC）评估特征稳定性，经Student t检验、皮尔逊相关系数和LASSO回归筛选特征，采用分类梯度提升（CatBoost）、轻量级梯度提升机（LightGBM）、逻辑回归（LR）、随机森林（RF）、支持向量机（SVM）、极端梯度提升（XGBoost）6种机器学习方法分别构建NAC前影像组学模型（RS1）以及2周期NAC后影像组学模型（RS2）；通过多因素Logistic回归分析筛选临床独立预测因素并构建临床模型；将最佳影像组学模型与临床独立预测因素结合构建联合模型，进一步基于最优模型构建列线图。通过受试者工作特征（ROC）曲线、校准曲线和临床决策分析（DCA）曲线评估模型效能。​

雌激素受体（ER）、孕激素受体（PR）、人表皮生长因子受体2（Her-2）、Ki-67为pCR的临床独立预测因子（OR=0.461，95%CI：0.218~0.978，P=0.044；OR=0.414，95%CI：0.207~0.829，P=0.013；OR=1.893，95%CI：1.417~2.530，P＜0.001；OR=5.178，95%CI：1.488~18.022，P=0.010），以ER、PR、Her-2、Ki-67构建临床模型，该模型在验证集中预测pCR的ROC曲线下面积（AUC）为0.797。影像组学模型中，NAC前最佳模型为瘤内+瘤周2 mm区域的RS1_XGBoost（AUC=0.726），2周期NAC后最佳模型为瘤周2 mm区域的RS2_SVM（AUC=0.770）。将临床独立预测因子（ER、PR、Her-2、Ki-67）分别与RS1_XGBoost及RS2_SVM结合构建联合模型，联合模型预测性能显著优于单一模型，NAC前联合模型预测pCR的AUC为0.842，2周期NAC后联合模型预测pCR的AUC为0.864；DCA曲线表明，两模型在阈值概率范围为6.7%~90.4%、5.6%~94.1%内均能提供临床净获益，曲线下面积分别为0.092和0.116；校准曲线分析显示，预测概率与实际概率一致，模型校准度良好（P=0.48）。基于预测效能最优的2周期NAC后联合模型构建列线图，列线图C指数为0.866。

基于NAC后超声影像组学特征联合临床病理因素构建的列线图模型能有效预测乳腺癌NAC疗效，尤其2周期NAC后瘤周2 mm区域的影像组学特征结合临床指标构建的模型可提供更精准的预测效能，为个体化治疗决策提供重要参考。

To investigate the predictive value of a nomogram model that integrates intratumoral and peritumoral radiomic features derived from conventional ultrasound images obtained before and after two cycles of neoadjuvant chemotherapy (NAC), along with clinicopathological factors, for assessing pathological complete response (pCR) in breast cancer patients following NAC.

A total of 332 female patients with pathologically confirmed breast cancer who received a standard six-cycle neoadjuvant chemotherapy (NAC) regimen at the Tenth Affiliated Hospital of Southern Medical University between July 1, 2019 and December 1, 2023 were retrospectively enrolled. All patients underwent ultrasound examinations before NAC initiation and after the second cycle of NAC, and these imaging data were used for early treatment response prediction. Patients were randomly divided into a training cohort (n=233) and a validation cohort (n=99) at a 7∶3 ratio. Intratumoral regions and peritumoral areas at 1 mm, 2 mm, and 3 mm margins were delineated on ultrasound images obtained before NAC and after two cycles of NAC. Radiomic features were extracted, and their stability was assessed using the intraclass correlation coefficient (ICC). Stable features were selected via Student's t-test, Pearson correlation analysis, and least absolute shrinkage and selection operator (LASSO) regression. Six machine learning algorithms—CatBoost, LightGBM, logistic regression (LR), random forest (RF), support vector machine (SVM), and XGBoost—were used to construct radiomics models. Independent clinical predictors were identified by multivariate logistic regression to build a clinical model. The optimal radiomics model was then combined with clinical predictors to establish a combined model, which was visualized as a nomogram. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA).

In the clinical model, estrogen receptor (ER), progesterone receptor (PR), Her-2, and Ki-67 were identified as independent predictors of pCR, yielding an area under the curve (AUC) of 0.797 in the validation cohort. Among the radiomics models, the optimal pre-NAC model was the RS_XGBoost model based on intratumoral and peritumoral (2-mm margin) features (AUC = 0.726), while the optimal post–two-cycle NAC model was the RS_SVM model based on peritumoral (2-mm margin) features (AUC = 0.770). The combined models demonstrated superior predictive performance compared with the single models, achieving AUCs of 0.842 before NAC and 0.864 after two NAC cycles. DCA demonstrated that both combined models provided clinical net benefit across threshold probability ranges of 6.7%–90.4% and 5.6%–94.1%, respectively, with area under the decision curve (AUDC) values of 0.092 and 0.116, respectively. Calibration curves indicated good agreement between predicted and observed outcomes (P=0.48). The nomogram achieved a concordance index (C-index) of 0.866.

The nomogram model integrating post-NAC ultrasound radiomic features with clinicopathological factors effectively predicts the therapeutic response of breast cancer to neoadjuvant chemotherapy. Notably, radiomic features from the peritumoral 2-mm region after two NAC cycles, when combined with clinical indicators, provide enhanced predictive accuracy and offer valuable guidance for individualized treatment decision-making.