影像诊断与介入放射学

目的 探讨基于超声与动态对比增强MRI(DCE-MRI)结合机器学习算法构建的预测模型，在良恶性乳腺肿瘤鉴别中的价值。方法 回顾性分析我院2019年8月—2024年5月行乳腺超声及DCE-MRI检查的348例女性乳腺肿瘤患者临床及影像资料(经伦理批准，排除42例后入组)，根据病理结果分为良性乳腺肿瘤组(BBTG)(n=156)和恶性乳腺肿瘤组(MBTG)(n=192)，按7∶3分为训练集(n=243)和测试集(n=105)。提取超声与DCE-MRI的定性及定量特征，经最小绝对收缩和选择算子(LASSO)-Logistic回归筛选特征后，通过DeLong检验比较联合诊断与单一诊断效能，构建决策树、随机森林、极限梯度提升(XGB)等7种机器学习分类器。通过独立测试集评估模型性能，采用沙普利加和解释(SHAP)值解释最优模型。结果 训练集中，BBTG与MBTG在DCE-MRI药代动力学参数[容积转运常数(K^trans)、速率常数(K_ep)、血管外细胞外容积分数(V_e)]、病灶形态、钙化、纵横比、形状、边缘、内部强化、时间-信号强度曲线(TIC)类型及乳腺影像报告和数据系统分级等方面差异有统计学意义(P<0.05)。LASSO-Logistic回归筛选出纵横比、形状、K_ep、K^trans、边缘、TIC曲线6个独立预测因子(P<0.05)。DeLong检验结果显示，超声与DCE-MRI联合诊断的效能最佳。7种模型中，XGB模型性能最优，训练集与测试集曲线下面积分别为0.983和0.978，敏感度、特异度均优于其他模型，决策曲线显示其分别在0.20～1.00和0.22～0.92阈值内临床净收益最高。SHAP分析提示边缘清晰、形状圆形/类圆、高K_ep值、低K^trans降低恶性风险，纵横比大于1、TIC曲线Ⅲ型增加恶性风险。结论 联合超声与DCE-MRI特征的XGB模型可有效提升良恶性乳腺肿瘤的鉴别效能，具有临床应用潜力。

Objective To explore the value of a predictive model constructed by combining ultrasound( US) and dynamic contrast-enhanced magnetic resonance imaging(DCE-MRI) with machine learning algorithms in distinguishing benign from malignant breast tumors. Methods A retrospective analysis was conducted on the clinical and imaging data of 348 women with breast diseases, who underwent breast US and DCE-MRI in our hospital from August 2019 to May 2024. According to pathological results,the patients were divided into benign(156) and malignant(192) breast tumor groups(BTG). They were further divided into a training set(n=243) and a test set(n=105) at a ratio of 7∶3. Qualitative and quantitative features from US and DCE-MRI were extracted. After feature selection via least absolute shrinkage and selection operator(LASSO)-logistic regression, the diagnostic efficacy of combined diagnosis and single-modality diagnosis was compared using the DeLong test. Seven machine learning classifiers including decision tree, random forest, and extreme gradient boosting(XGB) were constructed. The performance of the models was evaluated using the independent test set, and the optimal model was interpreted using Shapley Additive Explanations(SHAP) values. Results In the training set, there were significant differences between benign and malignant BTG in DCE-MRI pharmacokinetic parameters including volume transfer constant(K^trans), rate constant(K_ep), and extravascular extracellular volume fraction(V_e), lesion morphology, calcification,aspect ratio, as well as shape, margin, internal enhancement, time-signal intensity curve( TIC) type, and Breast Imaging Reporting and Data System classification(all P<0.05). LASSO-logistic regression screened out 6 independent predictors including aspect ratio,shape, K_ep, K^trans, margin, and TIC curve(all P<0.05). The DeLong test showed that the combined diagnosis of US and DCE-MRI had the best efficacy. Among the 7 models, the XGB model performed optimally with area under receiver operating characteristic curves of 0.983 in the training set and 0.978 in the test set. Its diagnostic sensitivity and specificity were superior to other models. The decision curve showed that it had the highest clinical net benefit within the thresholds of 0.20-1.00 in the training set and0.22-0.92 in the test set. SHAP analysis suggests that clear margins, round/quasi-round shape, high K_epvalue, and low K^transreduce the likelihood of malignancy whereas an aspect ratio >1 and type Ⅲ TIC increase the likelihood of malignancy. Conclusion The XGB model integrating US and DCE-MRI features improves the efficacy of distinguishing benign from malignant breast tumors.