深度学习模型在超声影像特征诊断乳腺病变性质中的应用  

Application of Deep Learning Model in the Diagnosis of Breast Lesions with Ultrasound Image Features

作　　者：谭垚 霍苓[2] 欧阳涛[2] 姚晨[1] Tan Yao;Huo Ling;Ouyang Tao(Peking University First Hospital,100034,Beijing)

机构地区：[1]北京大学第一医院,100034 [2]北京大学肿瘤医院暨北京市肿瘤防治研究所乳腺癌预防治疗中心,恶性肿瘤发病机制及转化研究教育部重点实验室

出　　处：《中国卫生统计》2021年第4期514-518,共5页Chinese Journal of Health Statistics

基　　金：北京市科委重点项目课题(D161100000816006)。

摘　　要：目的探究并比较深度神经网络模型与传统学习浅层模型在基于超声影像特征诊断乳腺病变性质的应用价值。方法将建模数据集以75%:25%比例拆分为训练集和测试集,同时构建6种浅层学习模型和深度神经网络模型,比较其在测试集和验证集的性能,以ROC曲线下面积(AUC)作为模型主要评估指标。结果在浅层学习模型中,logistic回归的AUC最大,除多层感知器外,其他差异均有统计学意义;深度神经网络模型的ROC曲线下面积(AUC)高于logistic回归,差异具有统计学意义。结论深度神经网络模型相比于传统浅层学习模型在基于超声影像特征诊断乳腺病变性质中有更大的诊断价值,但需要进一步探索并优化DNN模型,从而最终使临床医师能从深度学习模型的辅助诊断中获益。Objective To compare the application value of deep neural network model and traditional learning shallow model in the diagnosis of breast lesions based on ultrasound image characteristics.Methods Split the modeling data set into a training set and a test set at a ratio of 75%:25%,and construct 6 shallow learning models and deep neural network models at the same time to compare their performance on the test set and verification set,the ROC curve area under(AUC)is used as the main evaluation index of the models.Results In the shallow learning models,logistic regression has the largest AUC.Except for MLP(multilayer perceptron),other differences are statistically significant.The AUC of the deep neural network model is higher than logistic regression,and the difference is statistically significant.Conclusion Compared with the traditional shallow learning models,the DNN model has greater diagnostic value in diagnosing the nature of breast lesions based on ultrasound image features.However,it is necessary to further explore and optimize the DNN model,so that clinicians can finally benefit from the auxiliary diagnosis of the deep learning model.

关 键 词：乳腺超声 诊断模型 深度神经网络 浅层学习 LOGISTIC回归 

分 类 号：R445[医药卫生—影像医学与核医学]