Mammographic Mass Recognition Based On Convolutional Neural Network

According to the World Health Organization’s International Agency for Research on Cancer,female breast cancer has become the most commonly diagnosed cancer worldwide.Breast mass recognition is an important step in Computer-Aided Diagnosis(CAD)system,but it is still a challenging task due to the influence of varying in mass size,low contrast,and similarity with surrounding tissues.Traditional breast mass recognition algorithms rely on manual feature extraction,which is laborious and lack of robustness.In this dissertation,convolutional neural network is used to automatically extract the features of breast masses,and accurately recognize whether the mass is present in the Region of Interest(ROI).The experimental results verify the accuracy and validity of the proposed network.In this dissertation,the detected ROI containing suspicious masses is recognized to identify whether the ROI contains breast mass.Since the number of ROIs is relatively limited,and convolutional neural network needs to use a large amount of image data for training.In order to solve the demand of network for training data,two methods of rotation and flip are adopted to augment the data set by 24 times,so that the model can learn various representational information of mass and improve generalization ability.In convolutional neural networks,stacking a large number of convolutional layers will significantly increase the training complexity of the model.Using more convolution kernels at the convolutional layer can extract more features from the image,but excessive features will also have some redundancy,thus making the network model slow and bloated.Based on the above analysis,a Simple VGG16(SVGG16)network is proposed in this dissertation based on improving the VGG16 model.The proposed SVGG16 network contain five convolutional layers and two fully-connected layers,by reducing the number of convolutional layer stacking and using fewer convolution kernels,the complexity of the proposal is degraded and the architecture is simplified,so as to avoid the network to extract redundant features and improve the recognition accuracy of the model for masses.In medical image processing,the size of data set is always too small and the semantic information in the image is too little,so it is very important to make good use of the feature information in the image.Therefore,in this study,a new convolutional neural network model SE-FRNet18 was proposed based on the idea of feature reuse and channel attention mechanism.The network model is composed of two convolutional layers,three integrated modules and tow fully-connected layers.In the integrated module,the information flow of shallow layers is improved and feature reuse is implemented through dense connection,then the information between channels is modeled,so the automatically extracted feature information can be fully used to further improve the recognition performance of breast mass on mammography.The data set used in this dissertation was DDSM mammography data set,and the evaluation indexes were accuracy,precision,sensitivity,specificity,F1＿score and area under curve AUC.in the test set containing 176 mammographic ROI,the test results of SVGG16 and SE-FRNET18 were: accuracy rate of 90.34%/92.05%,precision rate of 89.87%/91.24%,sensitivity rate of 88.75%/91.25%,specificity rate of91.67%/92.71%,F1＿score rate of 0.89/0.91,and AUC rate of 0.95/0.96,respectively.The experimental results show that the two models presented in this dissertation have good performance in the recognition task of mass on mammography,which provides a theoretical basis for the application of computer-aided diagnosis system of breast mass on mammography.