【摘要】 目的 探讨基于超声图像的迁移学习模型在乳腺肿块良恶性鉴别诊断中的应用价值。方法 收集我院经手术病理证实的300例乳腺肿块患者共计582张超声图像作为超声数据集（其中训练集482张，测试集100张）。采用迁移学习方法对经ImageNet数据集预训练的3种深度卷积神经网络（VGG-16、Inception-v3、ResNet-50）模型进行训练和测试。第1次迁移学习为3种模型分别对CBIS-DDSM数据集中良恶性乳腺肿块的X线图像进行识别学习，并对模型进行微调；第2次迁移学习为使用乳腺超声数据集中随机挑选的训练集超声图像对3种模型进行微调，在测试集中输出最终分类结果。绘制受试者工作特征曲线分析迁移学习后3种模型对乳腺良恶性肿块的鉴别诊断效能。结果 VGG-16、Inception-v3、ResNet-50 3种模型经过迁移学习后鉴别诊断乳腺肿块良恶性的准确率、敏感性、特异性、精准率、F1分数、约登指数均有所提高，其中基于ResNet-50建立的模型具有最优的诊断效能，准确率88.0%，敏感性82.7%，特异性93.8%，曲线下面积0.915，均高于其他两种模型，差异均有统计学意义（均P<0.05）。结论 基于超声图像的迁移学习模型在乳腺肿块良恶性鉴别诊断中具有较高的应用价值，其中基于ResNet-50构建的模型效能最佳。更多还原

【Abstract】 Objective To explore the value of transfer learning model based on ultrasound images in the differential diagnosis of benign and malignant breast masses.Methods A total of 582 ultrasound images from 300 patients with breast masses confirmed by biopsy in our hospital were collected as an ultrasound data set(482 for training set and 100 for test set).Three deep convolutional neural network models(VGG-16,Inception-v3,ResNet-50)pre-trained on the ImageNet dataset were trained and tested by transfer learning method.For the first transfer learning,three models were used to identify and learn the X-ray images of benign and malignant breast masses in the CBIS-DDSM dataset and fine-tune the models.For the second transfer learning,three models using a randomly selected training set of ultrasound images from the ultrasound dataset were finetuned,and the final classification results were output in the test set.Receiver operating characteristic(ROC)curves were drawn to compare the diagnostic efficacy of three models after transfer learning for benign and malignant breast masses.Results The accuracy,sensitivity,specificity,precision,F1 score,and Jorden index of VGG-16,Inception-v3 and ResNet-50 models in the differential diagnosis of benign and malignant breast masses were all improved after transfer learning,and the model established based on ResNet-50 had the best efficacy with an accuracy of 88.0%,sensitivity of 82.7%,specificity of 93.8%,and AUC of0.915,which were higher than those of the other two models,and the differences were statistically significant(all P<0.05).Conclusion The transfer learning model based on ultrasound images has high diagnostic efficacy in the differential diagnosis of benign and malignant breast masses,among which the model established based on ResNet-50 has the best efficacy.更多还原