Breast Cancer Computer-Aided Diagnosis System Based On Multi-View Convolutional Neural Network

Breast cancer has become a serious threat to the health of women,and its incidence has been at the first place in female cancer.Mammography is considered to be an effective method of early detection and diagnosis,which is widely used in the examination of early breast lesions.Radiologists usually need to compare the images of multiple views for diagnosis of the lesions.In the face of huge workload,there may be missed diagnosis or misdiagnosis.Therefore,using multi-view-based deep learning method to assist diagnosis can not only reduce the workload of diagnosis,but also improve the accuracy and diagnostic efficiency of breast lesion recognition.In this paper,according to the multi-view characteristics of mammography and the different signs of benign and malignant breast masses,the deep learning technique was used to develop a computer-aided diagnosis system.The multi-view network built on the basis of the single-view mass recognition was employed to realize the multi-view mass recognition by feature fusion.At the same time,the advantages of the convolutional neural network were fully utilized to locate the target position through the visualization method.The work of this paper mainly includes the following aspects:First,a multi-view network model is designed.Generally,masses can be seen in both the cranial caudal position and the mediolateral oblique position.In order to make full use of the effective information of each view,a multi-view network is built on the basis of single-view branch network to accurately identify the masses.Under the Tensor Flow deep learning platform,the VGGNet model is fine-tuned to extract the feature of single-view network.And then the global average pooling layer is used to simplify a large number of parameters generated in the convolution and pooling process,so as to solve the problems of time consuming and memory consumption caused by the full connection layer.After feature fusion,the extracted features of each single view branch network are sent to the classifier for benign and malignant classification.Second,the model is optimized and visualized.In order to improve therecognition accuracy of the model and enhance the generalization ability of the model,rigid methods such as horizontal and vertical flipping are not only used to generate new images,non-rigid registration method is but also added to generate new training samples.Through the way of visual class activation map,the original classification performance of the network training can be maintained,and the target can be located in the original image.It helps to understand which key parts of the image enable the network to make the final classification decision.Finally,the system function interface is designed and implemented.Through the function modules of the computer-aided diagnosis system interface,the application of the multi-view network model is realized.The basic operations such as image preprocessing,image display and report writing of the system are completed,so as to ensure the integrity and practicability of the computer-aided diagnosis system.The multi-view network model built in this paper has been tested on the public date set.Its better performance can guarantee the accuracy,stability and reliability of the developed breast cancer computer-aided diagnosis system.