基于人工智能术前预测乳腺导管内癌微浸润的价值

Preoperatively Predicting Breast Ductal Carcinoma in situ with Microinvasion Based on Artificial Intelligence

目的基于超声图像的影像组学特征构建术前预测乳腺导管内癌微浸润(DCIS-MI)的鉴别模型,探索人工智能术前预测DCIS-MI的价值。资料与方法回顾性分析经病理证实的80例乳腺导管内癌(DCIS)和23例DCIS-MI的超声图像。在ITKSNAP软件进行图像分割,通过Intelligence Foundry软件进行影像组学特征提取并整理成数据集。数据集以7∶3分为训练组和验证组。训练组用于构建预测模型,验证组用于评估预测模型的可靠性。训练组通过Spearman相关系数以0.95为阈值去除高相关性特征,再使用统计检验联合随机森林的方法(1.25倍重要性均值为阈值)进一步选择重要特征。最后通过决策树机器学习算法构建DCIS-MI的预测模型。使用受试者工作特征曲线及曲线下面积(AUC)评估或验证模型效能。结果训练组纳入72例患者,其中56例DCIS、16例DCIS-MI;验证组纳入31例患者,其中24例DCIS、7例DCIS-MI。最终选择27项特征构建DCIS-MI的预测模型。训练组和验证组预测模型的AUC分别为0.90和0.73,准确度、敏感度及特异度分别为0.79、0.94、0.75和0.74、0.71、0.75。影像组学评分在不同临床病理参数亚组的AUC均>0.77。结论基于超声图像构建的影像组学预测模型具有良好的预测效能,可在一定程度上达到术前辅助诊断DCIS-MI的作用,为临床快速有效决策提供依据。

Purpose Based on the radiomics features of ultrasound images,a differential prediction model of preoperative ductal carcinoma in situ with microinvasion(DCIS-MI)is constructed to explore the value of artificial intelligence to predict DCIS-MI before surgery.Materials and Methods Ultrasound images of 80 patients with breast ductal carcinoma in situ(DCIS)and 23 patients with DCIS-MI confirmed by pathology were studied retrospectively.Image segmentation was carried out in ITKSNAP software.Radiomics feature extraction were carried out by Intelligence Foundry software,and data set was generated.The data set was divided into the training set and the test set in a ratio of 7∶3.The training set was used to construct predictive models.The test set was used to evaluate the reliability of the model.In the training set,Spearman correlation coefficient was used to remove the high correlation characteristics with 0.95 as the threshold.Then,the important features were selected by the statistical test combined with the random forest method(1.25 times the mean value of importance was the threshold).Finally,the decision tree machine learning algorithm was applied to construct the prediction model of DCIS-MI.The area under the receive operating characteristic curve(AUC)was used to evaluate the effectiveness of the model.Results There were 72 cases in the training set,including 56 cases of DCIS and 16 cases of DCIS-MI;31 cases in the test set,including 24 cases of DCIS and 7 cases of DCIS-MI.27 features were eventually selected to construct a prediction model for DCIS-MI.The AUC of training set and test set were 0.90 and 0.73,respectively;the accuracy,sensitivity and specificity were 0.79,0.94,0.75;0.74,0.71,0.75,respectively.The AUC of radiomics scores in different clinicopathological parameters subgroups were more than 0.77.Conclusion The radiomics prediction model of DCIS-MI based on ultrasound images before operation has moderate predictive efficiency by the test set,which can assist in preoperative diagnosis of DCIS-MI to a certain extent,and provide a basis for rapid and effective clinical decision-making.