Deep Convolutional Neural Network In Breast Mass Segmentation

Mammography is one of the most commonly used methods for screening breast cancer,which has the advantages of ease of operation,high imaging quality and noninvasive.Breast Mass(Mass)is the most common pathological feature of breast cancer.Combined with computer-aided diagnosis technology to segment the abnormal Mass in molybdenum target image,relevant clinical measures can be taken before the cancer of the Mass to achieve the purpose of reducing the mortality rate.Due to the large differences in the size,shape and distribution of breast masses,it is still a challenging task to achieve automatic segmentation of breast mass region.The existing segmentation methods have some problems such as low sensitivity of mass boundary recognition and loss of feature information during network training.Around how to efficiently realize the automatic segmentation of breast masses,this paper carried out the following work:1.Preprocessing of mammography images based on data enhancement.Aiming at the problem of lack of mammary gland molybdenum target data set,in addition to respectively from two foreign public CBIS-DDSM and mammary gland molybdenum target dataset INbreast from 668 and 101 breast image,in order to make network model can better apply to domestic conditions,based on the images provided by the mianyang central hospital building SMD-My local data sets,and the selected local image breast annotation,image normalization preprocessing operations,to support the follow-up training work.2.Research on breast mass segmentation method based on improved U-Net and compound weight loss function.Aiming at the problem that the current deep learning segmentation methods are not sensitive to the boundaries of breast masses and are difficult to effectively segment breast masses,a U-shaped symmetric residual semantic segmentation model based on a compound weighted loss function,SRes-UNet,is proposed.Embed the residual module into the overall U-Net architecture to improve the model’s global feature extraction capabilities.A composite weighted loss function w BCE_Dice is used to limit the optimization direction of the network during model training,so that the model will focus more on training Allocate to the mass area,thereby reducing the interference of the invalid background area on the training.Experimental results show that the segmentation model proposed in this paper has achieved better segmentation results on mammography target tumor segmentation tasks,reaching 82%and 86%on DSC and MIo U indicators,respectively;compared with typical U-Net,the results The proposed model has increased by 2%and 4%respectively on the two indicators of DSC and MIo U.3.Attention Guided Residual U-Net Method for Breast Mass Segmentation.The feature details are easy to be lost in the U-NET network structure,which leads to the low precision of features extracted from the network model.Based on the attention module of channel excitation and spatial attention mechanism,and embedded into sesa-Res-UNet,this paper proposed the segmentation model of molybda-target breast mass,sesa-Res-UNet.The experimental results showed that sesa-Res-UNet achieved the best performance,with the DSC and MIOU values reaching 84.2% and 86.8%,respectively,which were significantly better than the Baseline U-Net Dense_UNet,Res_UNet and SRes-UNet in each index.To realize the automatic segmentation of breast masses and improve the detection rate of breast cancer can not only detect asymptomatic patients early,reduce the mortality rate,but also play a positive role in promoting the development of medical science.Both of the two automatic breast mass segmentation methods proposed in this paper can outline the shape of the lump fine and recover the edge of the lump more accurately,which confirms the effectiveness of the improved segmentation network based on U-Net and has the potential to play a positive role in the future work of breast mass assisted diagnosis.