Research On Classification Method Of Mammogram Based On Deep Convolutional Neural Network

Breast cancer is the most common cancer in women,and its early symptoms can be detected by X-ray radiography.Therefore,mammography is the most reliable method for early screening and diagnosis of breast cancer,but the huge amount of reading will increase the the difficulty of diagnosis.This study is based on the computer-aided diagnosis(CAD)of mammography,which can help radiologists to provide decision-making opinions during reading,improve diagnostic accuracy and objectivity,and reduce the workload of diagnosis.In this paper,a 9-layer deep convolution neural network(DCNN)is designed for automatic classification of mammographic images.The database used was the MIAS data set.First,330 ROIs were extracted from 322 samples according to the physician’s mark and normalized.Then the data were expanded using downsampling and geometric changes.Dropout technology was also used in the network to prevent over-fitting,and finallyFinally,the three classifications of mammograms were achieved(ie,normal,benign and malignant).At the same time,the influence of different network layers on the network performance was discussed,and the best model was determined by experiments.The model achieved the classification accuracy of 90.18% and the average F1 score of 89.48%.In general,in the face of small data sets,especially medical images,it is difficult to train the network from scratch in depth learning.Transfer learning is an effective way to deal with this problem.The importance of transfer learning also were discussed in this study.We finetuned the three networks of VGG16,Inception V3,and ResNet50 pre-trained on ImageNet.With the random initialization network as the control,the classification accuracy of the three networks was 89.95%,86.37%,and 87.43%,respectively.After usingtransfer learning,the classification accuracy rate reached 90.36%,92.58%,and 91.10%.The results show that transfer learning is effective in medical image analysis.Compared with other classification methods on the same public dataset,the best model in this paper has greatly improved the classification accuracy,which can provide a certain reference value for the classification of breast masses in practical applications.