Study On Classification Of Mammography Breast Mass Based On Improved DSOD Neural Networks

Breast cancer is one of the main murders of cancer death among women in worldwide.Digital mammography can effectively diagnose the early stages of breast cancer and greatly increase the chance of survival.The computer-assisted diagnosis system can achieve the automatic detection of breast X-ray images and improve diagnosis efficiency.Due to the burred edge of the breast mass and the low contrast with normal glandular tissue,the automatic detection of the mass still has a high rate of missed diagnosis and misdiagnosis.Traditional computer-aided diagnosis systems rely on empirical knowledge to manually extract limited feature information,which is subjective and uncertain.This paper classifies and locate the breast mass based on deep learning method,which utilizes the powerful feature extraction capabilities of neural networks to achieve more objective and comprehensive diagnosis result.It aims to provide second recommendations for diagnosis.The main research work of this paper is as follows:(1)Based on the original DSOD(Deeply Supervised Object Detector)network model,this paper improves the feature extraction backbone network and feature layer prediction sub-network of the DSOD model based on the characteristics of the mammography image.Firstly,in the feature extraction backbone network,retaining its densely connected network structure,we introduce a dense convolution module based on depthwise separable convolution to replace the original dense block.This improved method can not only reduce the demand for computing resources but also reduce the model parameters and the amount of calculation.Secondly,we add the SE-Block module which based on the channel domain attention mechanism between adjacent dense convolution modules to establish a feature correlation between channels and re-calibrate the channel features.The SE-Block can improve the favorable features of the current task and suppress irrelevant features,Thereby enhancing the ability of the network to extract image features.(2)In view of the small size of the mammography images,this paper redesigns the network structure parameters based on the above improvement ideas.,reselecting the network growth rate k and reduces the number of convolution units in the dense convolution module to make the network model more compact and streamlined.The new proposed model has fewer parameters and prevents overfitting effectively in the training process.(3)Due to the different sizes of breast masses,in order to improve the ability of DSOD model to classify small scale breast mass,a dilated convolution Inception module is added to the feature prediction sub-network.This module combines the advantages of dilated convolution and Inception network structure,which combines the multi-scale and multi-branch information to improve the detection effect on small breast mass.The receptive field is expanded effectively by using the dilated convolution.Considering the uneven distribution of positive and negative samples in the dataset,this paper improves the loss function and introduces the focal loss function to improve the model classification performance.(4)Through the ablation experiment,the effectiveness of the various improved modules proposed in this paper can be verified,which can improve the accuracy of the model in classifying breast masses and reduce the model parameters.Through comparison experiments,compared with other target detection models,the accuracy of breast masses classification is 82.9%,80.6%,and 80.3% respectively based on the improved DSOD model.As a result,the improved DSOD model which is proposed in this paper can effectively classify benign and malignant breast mass,and provide effective assistance to improve the doctor's clinical diagnosis.There are 34 figures,6 tables and 56 references in this paper.