评估并比较基于不同机器学习算法建立的乳腺癌超声影像组学预测模型的诊断性能。

回顾性收集2017年1月至2019年4月就诊皖南医学院第一附属医院、有明确病理结果的乳腺肿块病例828例，以2018年8月31日为节点将其分为训练集（526例）和验证集（302例），提取肿块的超声影像组学特征并进行特征筛选，运用k最近邻（kNN）、逻辑回归（LR）、朴素贝叶斯（NB）、随机森林（RF）和支持向量机（SVM）5种机器学习算法分别建立预测模型，使用重复交叉验证方法做内部验证，计算比较各模型的敏感度、特异度、阳性预测值（PPV）和阴性预测值（NPV），并实施外部验证，绘制ROC曲线并比较ROC曲线下面积（AUC）以评价模型的鉴别诊断性能，绘制校准曲线评价模型校准度。

从提取的109个影像组学特征中筛选出19个特征建立了5种机器学习算法模型。在内部验证中，5种模型的敏感度、特异度、PPV、NPV比较，总体差异均有统计学意义（P均＜0.001）。LR模型的特异度、PPV、NPV中位数分别为0.769、0.816、0.778，3项指标均高于其他4种模型；敏感度中位数为0.824，高于kNN、RF和SVM模型。此外，SVM模型的特异度、PPV、NPV中位数分别为0.706、0.774、0.759，虽均低于LR模型，但均高于其他3种模型。在外部验证中，LR、SVM、RF、kNN和NB的AUC依次为0.890、0.832、0.821、0.746和0.703，其中LR与SVM的AUC差异有统计学意义（P=0.012）；此外，各模型在校准性能上表现并不一致，LR和SVM模型的校准曲线显示乳腺癌实际概率与预测概率之间的一致性较好。

以超声影像组学特征为基础，运用不同机器学习算法建立的乳腺癌超声预测模型，均表现出较高的诊断性能，其中LR模型表现最为突出；选择合适的机器学习算法有助于进一步提高预测模型的诊断性能，提供更加准确的量化预测结果。

To assess and compare the diagnostic performance of ultrasound-based radiomics models constructed by different machine learning (ML) algorithms for identifying breast cancer.

Between January 2017 and April 2019, 828 consecutive patients with pathologically confirmed breast lesions at the First Affiliated Hospital of Wannan Medical College were included in this study. The patients were divided into a training set (n=526) and a validation set (n=302) according to the cutoff date of August 31, 2018. Radiomics features were extracted from the gray-scale ultrasound images. After features selection, five ML models based on five ML algorithms including k-nearest neighbor (kNN), logistics regression (LR), naive Bayes (NB), random forest (RF), and support vector machine (SVM) were constructed using the training set. Internal validation was performed using repeated k-fold cross-validation. Diagnostic metrics including sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated for comparison. Furthermore, model discrimination and calibration were assessed in the external validation via ROC and calibration curve analysis.

Nineteen from 109 radiomics features were selected as effective features, and five ML prediction models were established. All the diagnostic metrics were statistically different across the models in the internal validation (P＜0.001). The median specificity, PPV, and NPV of the LR model were 0.769, 0.816, and 0.778, respectively; the median sensitivity was 0.824, which was higher than that of the kNN, RF, and SVM models. The median specificity, PPV, and NPV of SVM model were 0.706, 0.774, and 0.759, respectively, which were lower than those of the LR model, but higher than those of the other three models. The LR, SVM, RF, kNN, and NB models demonstrated good discrimination in external validation, with area under the curve (AUC) values of 0.890, 0.832, 0.821, 0.746, and 0.703, respectively. Of these, a significant difference was observed in AUC values between the LR and SVM models (P=0.012). While all the models did not perform consistently, the calibration curves for LR and SVM models indicated that the actual probability and the predicted probability agreed well.

The five ultrasound-based radiomics models developed based on different ML algorithms all demonstrate high diagnostic performance, but the LR model stands out for its superior performance. By applying an appropriate ML algorithm, the diagnostic performance of the final model could be further enhanced.