机器学习基于动态对比增强MRI鉴别乳腺良性与恶性病变的价值

Value of Machine Learning Based on DCEMRI for Predicting Diagnosis of Benign and Malignant Breast Lesions

目的本研究旨在探讨不同机器学习方法在乳腺良性与恶性病变预测中的应用。方法回顾性分析厦门医学院附属第二医院2018年8月至2022年5月经病理证实的271个患者临床和影像资料。采用分层抽样方法以7:3的比例划分训练组和验证组,提取影像组学特征,训练组使用冗余性分析、最小绝对收缩和选择算子交叉验证算法进行特征筛选,采用逻辑回归、支持向量机、自适应增强算法及决策树4种不同具有监督学习的机器学习方法来预测乳腺良性与恶性病变的能力。使用受试者工作特征曲线(ROC曲线)、F1度量值和准确率对四种机器学习算法进行评估,并通过验证组进行验证。绘制校准曲线用于评价预测概率和实际概率之间的偏差。结果基于17个影像组学特征,逻辑回归算法鉴别乳腺良性与恶性病变的预测效果最好,验证组有最高的曲线下面积(AUC值)为0.832(0.744-0.919),准确率为78%,F1度量值为0.790。最终以逻辑回归机器学习算法建立预测模型,逻辑回归算法模型校准曲线具有良好的重叠性。结论基于乳腺动态对比增强MRI逻辑回归机器学习算法有助于鉴别乳腺良性与恶性病变,为临床医师的决策提供指导。

Objective To explore the value of imaging models with different machine learning methods in predicting benign and malignant breast lesions.Methods The clinical and imaging data of 271 patients confirmed by histopathology in the second affiliated Hospital of Xiamen Medical College from August 2018 to May 2022 were analyzed retrospectively.The stratified sampling method was used to divide the training group and the verification group at the proportion of 7:3.The third phase of dynamic contrast enhanced MRI(DCE-MRI)was used to extract imaging features.The training group uses redundancy analysis,minimum absolute contraction and selection operator cross-validation algorithm for feature screening.Four different machine learning methods with supervised learning,logical regression,support vector machine,adaptive enhancement algorithm and decision tree,are used to predict benign and malignant breast lesions.The receiver operating characteristic curve(ROC),accuracy and F1 measure were used to evaluate the advantages and disadvantages of the four machine algorithms,and verified by the verification group.The calibration curve is used to evaluate the deviation between the predicted probability and the actual probability.Results Based on 17 imaging features,the prediction effect of logical regression algorithm is the best in distinguishing benign and malignant breast lesions.The highest area under the curve(AUC value)is 0.832(0.744-0.919)in the verification group,and the accuracy is 78%.The F1 measure is 0.790.Finally,the prediction model is established by logical regression machine learning algorithm,and the calibration curve of logical regression algorithm model has good overlap.Conclusion Logical regression machine learning based on breast dynamic contrast enhanced MR imaging model is helpful to distinguish benign and malignant breast lesions and provide guidance for clinicians to make decisions.