【摘要】 目的 基于栖息地超声影像组学特征构建机器学习模型，探讨其鉴别诊断乳腺结节良恶性的临床价值。方法 选取2019～2022年公共数据库波兰医疗中心的248个乳腺结节，依据7∶3的比例随机分为训练集174个和验证集74个。基于所有结节的超声图像进行栖息地聚类分析划分不同肿瘤亚区域，使用Pyradiomics软件包提取肿瘤整体区域和不同亚区域影像组学特征。采用线性判别分析机器学习算法构建不同区域特征的机器学习模型，绘制受试者工作特征（ROC）曲线分析各模型及波兰医疗中心专家医师、本中心具有3年乳腺诊断经验的青年医师评估的BI-RADS分类在训练集和验证集中鉴别乳腺结节良恶性的诊断效能，并进行比较。选择曲线下面积（AUC）最高者为最优模型，绘制校准曲线及决策曲线分别评估其校准度及临床适用性。结果 从栖息地聚类分析的轮廓系数变化可见，当聚类数为2时，聚类效果最好，划分为亚区域1和亚区域2。基于不同区域提取的特征构建模型，整体区域模型、亚区域1模型、亚区域2模型、专家及青年医师BI-RADS分类在训练集和验证集中鉴别乳腺结节良恶性的AUC分别为0.87、0.86、0.82、0.84、0.80和0.77、0.89、0.76、0.87、0.76。在验证集中，亚区域1模型鉴别乳腺结节良恶性的AUC高于整体区域模型、亚区域2模型、青年医师BI-RADS分类（均P<0.05），与专家BI-RADS分类的AUC比较差异无统计学意义；整体区域模型与亚区域2模型、整体区域模型与青年医师BI-RADS分类、亚区域2模型与青年医师BI-RADS分类的AUC比较，差异均无统计学意义。选择验证集中AUC最高的亚区域1模型为最优模型。校准曲线显示，亚区域1模型预测概率与实际概率之间的一致性良好；决策曲线显示，亚区域1模型在5%～93%概率阈值范围内可使患者获益。结论 相较于传统的整体区域特征，基于栖息地超声影像组学特征的机器学习模型能有效提升乳腺结节良恶性的鉴别诊断效能，为相关人工智能辅助诊断系统的临床转化与应用提供依据。更多还原

【Abstract】 Objective To construct a machine learning model based on habitat-based ultrasound radiomics features,and to explore its clinical value in differentiating benign from malignant breast nodules.Methods A total of 248 breast nodules from the public database of the Polish Medical Center(2019-2022)were selected and randomly divided into a training set(174 nodules)and a validation set(74 nodules)with a ratio of 7∶3.Habitat clustering analysis was performed on ultrasound images of all nodules to divide different tumor subregions.Radiomics features were extracted from the whole tumor region and different subregions by the Pyradiomics software package. Linear discriminant analysis machine learning algorithm was employed to construct machine learning models based on features from different regions.Receiver operating characteristic(ROC)curves were plotted to analyze and compare the diagnostic performance of the models,the BI-RADS classifications assessed by expert physicians from the Polish Medical Center,and the BI-RADS classifications assessed by junior physicians with three years of breast diagnostic experience at our center in differentiating benign from malignant breast nodules in both the training and validation sets. The model with the highest the area under the curves(AUCs)was selected as the optimal model. Calibration curves and decision curves were then plotted to evaluate its calibration and clinical applicability,respectively.Results The silhouette coefficient from habitat clustering analysis indicated optimal clustering performance when the cluster number was 2,dividing the nodules into Subregion 1 and Subregion 2. Models were constructed based on extracting features from different regions,and AUCs for the whole-region model,Subregion 1 model,Subregion 2 model,expert physician BI-RADS classification,and junior physician BI-RADS classification in the training set were 0.87,0.86,0.82,0.84,and 0.80,respectively,and in the validation set were 0.77,0.89,0.76,0.87,and 0.76,respectively.In the validation set,the AUC of the Subregion 1 model for differentiating benign and malignant breast nodules was higher than that of the whole-region model,Subregion 2 model,and junior physician BI-RADS classification(all P<0.05),while no statistically significant difference was observed compared with the expert physician BI-RADS classification.No significant differences were observed in comparisons of AUC between wholeregion model and Subregion 2 model,whole-region model and junior physician BI-RADS classification,Subregion 2 model and junior physician BI-RADS classification.The Subregion 1 model was selected as the optimal model.The calibration curve showed good agreement between the probability of the Subregion 1 model and actual probability.The decision curve indicated that the Subregion 1 model provided clinical benefit within a probability threshold range of 5%-93%.Conclusion Compared with traditional whole-region features,the machine learning model based on habitat-based ultrasound radiomics can effectively enhance the diagnostic performance in differentiating benign from malignant breast nodules,and provide a basis for the clinical transformation and application of related AI-assisted diagnostic systems.更多还原