The Study Of Full Mammography Images Segmentation Based On Fully Convolutional Network

Breast cancer poses a serious threat to women’s health and safety worldwide,which Let the whole society pay attention.But at present,for breast cancer,the medical community has not given a clear cause,early prevention screening is particularly important,and it can effectively reduce the risk of breast cancer.The main methods of breast cancer screening mammography X-ray images,its non-traumatic,intuitive imaging and strong reliability,so the performance of lumps in the target image can be used as one of the main symptoms for breast cancer diagnosis.Traditional diagnosis,mainly by the experience of doctors’ diagnosis,it is time-consuming.Breast lumps have various shapes,the edges are very blurred and irregular,this complex medical imaging structure can cause interference,and the accuracy and stability of diagnosis strongly reply on the physician experience.In view of the difficulties of breast cancer diagnosis,more and more scholars have introduced deep learning into this field,which can not only realize the automatic learning of the characteristics of mammography image,but also achieve the segmentation of mammography images more efficiently and accurately,which are great significance to breast cancer diagnosis.In this paper,focusing on the research on mass segmentation of mammography mammography targets.The main contents and innovations are as follows:a)We designed the asymmetric encoder-decoder structure with dense skip connection and multi-scale fusion.Dense jump connection multi-scale structure can more effectively retain the low-level and high-level features of the image;at the same time,this connection method enables the low-level network to more fully extract the high-resolution information of the image,effectively improving the breast with complex edges The accuracy of the mass segmentation.b)We designed an attention-guided fusion up-sampling module.The upsampling module combines the Pixel Shuffle and DUpsampling upsampling models and connects to the attention mechanism.The module has the following characteristics: First,the module compensates for the loss of image information caused by bilinear upsampling through Pixel Shuffle and DUpsampling upsampling operations;secondly,By fusing the two up-sampling features to avoid the loss of image features;finally,a channelattention module and spatial-attention module are introduced to better simulate the doctor’s visual attention mechanism,and to better target mammography images highlight the breast mass information while suppressing other interference factors in the image,so that the network focuses on the learning of breast mass features and achieves accurate segmentation.c)Based on the research of the full convolutional network and the foundation of the above model,proposing a multi-scale fusion and attention-guided up-sampling mammography mass segmentation network（Am UNet）.The segmentation experiment of breast mass on the image mammography target,Am UNet realizes the end-to-end segmentation of the mass on the whole X-ray image of the mammography target.d)The Am UNet segmentation model proposed in this paper has conducted a large number of experiments on three public data sets: CBIS-DDSM,Inbreast and lgg-mriSegmentation.In the experiment,based on the above public database,this article also compared the breast mass segmentation results of Attention-UNet,UNet++,Deep Lab V3+ and Fusion Net（four fully convolutional neural networks）.The following conclusions are drawn: Compared with the above four segmentation results of the product neural network,the Am UNet proposed in this paper achieves the highest segmentation Dice similarity coefficients（DSC）.The average Dice similarity coefficients of the mass segmentation on the CBIS-DDSM,DDSM ROI,INbreast and lgg-mri-Segmentation databases respectively are 81.4%,89.6%,78.9% and 92.1%.The experimental results also proves that the network has good universality.