【摘要】 目的：运用超声组学技术提取三阴性乳腺癌（triple-negative breast cancer,TNBC）瘤内及瘤周4 mm区域的特征，结合多种机器学习分类器构建预测模型，以提高TNBC的诊断准确度。方法：回顾并收集2017年4月—2023年7月南京医科大学第一附属医院经病理学检查确诊的乳腺癌患者的超声图像，将其分为训练集和测试集。对瘤内和瘤周4 mm环形区域进行勾画，提取形态学、纹理及一阶特征。采用最小冗余最大相关算法（minimal redundancy maximal relevance,mRMR）筛选特征后，基于随机森林（random forest）、极端随机树（extra-trees）、极端梯度提升（XGBoost）和轻量级梯度提升机（LightGBM） 4类机器学习分类器，以及4种预测特征来源组合（包括仅来自瘤内区域的特征、仅来自瘤周区域的特征、瘤内和瘤周区域的图像融合特征以及瘤内区域特征与瘤周区域特征相结合的特征融合），构建16个超声组学模型，筛选最优预测模型，采用沙普利加性解释（SHapley Additive exPlanation,SHAP）揭示模型的重要特征。结果：共纳入563例乳腺癌患者，其中TNBC 107例，非TNBC 456例。按7∶3随机分为训练集（394例）和测试集（169例）。分别从瘤内区域、瘤周区域以及瘤内与瘤周图像融合区域各提取1 561个超声组学特征。基于extra-trees的特征融合模型表现最佳，训练集、测试集的曲线下面积（area under curve,AUC）达0.952、0.857，而瘤内模型AUC为0.912、0.820，瘤周模型AUC为0.909、0.806，图像融合模型AUC为0.924、0.848，表明特征融合模型在预测TNBC方面优于单独瘤内或瘤周特征的模型。SHAP进一步揭示了瘤内区域的3个超声组学特征和瘤周区域的2个超声组学特征为模型中最具影响力的5个决定性因素。结论：本研究强调了利用瘤内和瘤周的超声组学特征构建的extra-trees模型，在预测TNBC方面的潜力。更多还原

【Abstract】 Objective: To leverage radiomics based on ultrasonography to extract quantitative features from both the intratumoral volume and a 4 mm peritumoral rim of triple-negative breast cancer(TNBC), and to integrate these features with multiple machinelearning classifiers to construct a high-accuracy diagnostic model. Methods: A retrospective cohort of pathologically confirmed breast-cancer patients who underwent ultrasonography between April 2017 and July 2023 was collected and randomly split into training and testing sets. After manual segmentation of the intratumoral region and a 4 mm peritumoral annulus, radiomic descriptors(morphological, first-order, and texture features) were extracted. Following minimum redundancy maximum relevance(mRMR) feature selection, sixteen radiomics models were built by combining four machine-learning classifiers—random forest, extra-trees, XGBoost, and LightGBM—with four distinct feature-source strategies: intratumoral features only, peritumoral features only, fused images of both compartments, and concatenated intratumoral + peritumoral features. Model performance was evaluated by area under the receiver-operating-characteristic curve(AUC). SHapley Additive exPlanations(SHAP) were used to interpret the optimal model. Results: A total of 563 breast cancer patients were included, consisting of 107 TNBC cases and 456 non-TNBC cases. Patients were randomly divided into a training set(n=394) and a test set(n=169) at a ratio of 7∶3. A total of 1 561 ultrasound radiomics features were extracted from each of the intratumoral, peritumoral, and image fusion regions, respectively. The extra-trees model trained on concatenated intratumoral and peritumoral features achieved the highest AUC(training set 0.952, test set 0.857), outperforming intratumoral-only(0.912, 0.820), peritumoral-only(0.909, 0.806), and fused-image models(0.924, 0.848). SHAP analysis identified three intratumoral and two peritumoral radiomic features as the five most influential determinants of TNBC prediction. Conclusion: Integrating intratumoral and peritumoral ultrasound radiomics via an extra-trees classifier significantly enhances the non-invasive identification of TNBC, underscoring its translational potential in precision oncology.更多还原