| Objective:To explore the value of deep learning combined with transfer learning in the study of two-dimensional ultrasound images of breast nodules in the differential diagnosis of benign and malignant breast nodules and diagnosis of axillary lymph node invasion in breast cancer.Methods:Ultrasound images and diagnostic results of breast nodules and axillary lymph nodes from March 2019 to December 2020 were collected prospectively.Taking the pathological diagnosis as the gold standard,the two-dimensional ultrasound images of breast nodules were analyzed by the method of artificial intelligence deep learning network and transfer learning,and the neural network was constructed to differentiate the benign from the malignant breast nodules,the gray-scale ultrasound images of breast cancer were used to construct neural network,for predict axillary lymph node invasion.Using ResNet50 as the initial network model,the network was pre-trained on Image Net before network construction.The region of interest(ROI),I.E.Image Segmentation,is manually delineated on gray-scale ultrasound images of breast nodules.The images were pre-processed by the methods of horizontal,vertical,image transposition and rotation transformation,so as to increase the sample size,in order to avoid the over-fitting phenomenon in the learning process.The data was normalized and the features of different dimensions were compared in numerical value.All the 509 cases of benign and malignant breast nodules(1716 images)collected in our hospital were used as the data set for constructing the benign and malignant breast nodules diagnosis model Resnet50-N,in which training set N(n = 356,1183 images)and test set N(n = 190,533 images).The ultrasound images of 240 cases of breast cancer collected in our hospital(1180 images)were used as the data set M of Resnet50-L,including training set n=168(training set M,882 images)and test set n=72(test set M,298 images).Start fine tune from the last layer of ResNet50 during the training process,and gradually include more layers during the updating process until the positive fitting is achieved.The learning rate is set to 0.001,the maximum iteration step is set to 5000,and the learning rate decays by 1/2 at 2000 and 4000 steps.The data set used to evaluate the performance of Resnet50-N was all 454 cases of benign and malignant nodules(validation set N,1294 images)collected in other hospital.The data set used to evaluate ResNet50-L was collected from 130 breast cancer cases(validation set L1,649 images).In the meantime,physician A(2 years clinical experience)and physician B(7 years clinical experience)diagnosed validation set N and validation set L2.The validation set L2(260 images)was lymph nodes images of the cases of validation set L1.The area under the curve(AUC),sensitivity(SEN),specificity(SPE)and accuracy(ACC)were used to evaluate the generalization performance and physician diagnostic efficiency of Resnet50-N.Results:(1)ResNet50-N classification performance: The AUC,SEN,SPE,ACC in training set N and test set N were 0.970 and 0.933,94.3% and 89.6%,97.6% and 85.7%,96.0% and 88.4%.The AUC,SEN,SPE and ACC in validation set N were 0.882,81.8%,87.6%,84.7%.(2)ResNet50-L classification performance: The AUC,SEN,SPE,ACC in training set M and test set M were 0.936 and 0.907,89.0% and 84.5%,92.6% and 89.6%,90.0% and 88.6%.The AUC,SEN,SPE and ACC in validation set L1 were 0.882,81.8%,87.6%,84.7%.(3)Diagnosis of two physician:The AUC,SEN,SPE,ACC of physician A in validation set N and validation set L2 were 0.710 and 0.705、62.6% and 65.1%、79.4% and 75.9%、70.9% and 72.3%.The AUC,SEN,SPE,ACC of physician A in validation set N and validation set L2 were 0.833 and 0.775、80.3% and 83.7%、86.4% and 71.3%、83.3% and 75.4%.(4)In validation set N,the diagnostic efficiency of Resnet50-N was higher than that of two physicians(P<0.001).The AUC value of ResNet50-L for axillary lymph nodes was higher than that of two Physicians for direct diagnosis of axillary lymph node involvement.Conclusion:(1)AI technology which combined deep learning with transfer learning was reliable as attending doctors in the diagnosis of benign and malignant breast nodules at least,and can provide additional diagnostic assistance to residents.(2)AI can detect and differentiate breast cancer accurately,which is helpful for early detection,diagnosis and treatment of breast cancer.(3)The AI technology has a good application prospect in indirect prediction of axillary lymph node invasion,which is helpful to establish a more reasonable management scheme of axillary lymph nodes. |
PDF Full Text Request