Segmentation Algorithm For Breast Mass Based On Mammography Images

One of the most common cancers among women is breast cancer,and early screening and diagnosis are very meaningful.Breast mass is one of the main signs of early breast cancer.The most effective means of breast cancer screening is mammography.Studying the algorithm of breast mass segmentation,by accurately segmenting the breast mass,assisting doctors to judge the benign and malignant of the breast mass,help to improve the survival probability of patients,and have important clinical significance.Breast masses are usually characterized by diverse shapes and irregular edges.Aiming at the problem of diverse breast morphology and size,the MU-Net model based on the U-Net model was designed and implemented.The characteristics of the residual network structure and the U-Net model structure were merged.Six down-sampling was set to obtain more advanced Semantic information and a larger receptive field.At the same time,the strategy of convolution followed by pooling is adopted in the down-sampling module to reduce the loss of position information,which makes the model’s performance for multi-scale breast mass segmentation improved.In view of the problem that the MU-Net model segmentation results are not detailed enough,the characteristics of the cascaded network structure are studied,and the MDCas MU-Net model formed by cascading six MU-Net models with different sampling depths is designed and implemented.The corresponding sampling times are One to six,through the simple network coarse segmentation and complex network fine segmentation,the segmentation results of breast masses are more refined.Aiming at the problem that the edge of irregular breast mass is difficult to segment,the effect of multi-scale breast mass image on the model is studied,and a two-stage cascade training method is designed and implemented.The breast mass image of different sizes is used as the input of the two-stage model.The background area around the lump makes the model more familiar with the background.At the same time,the prediction results of the first-stage model are added to the input of the second-stage model to reduce the interference caused by the background area.The boundary of the resulting breast lump mask is more complete and smooth.Aiming at the problem of fewer training samples,a histogram matching data enhancement method is designed and implemented.By changing the pixel value distribution of the breast mass image,the model’s dependence on high pixel value points is reduced,thereby significantly improving the model’s robustness.In order to verify the performance of the breast mass segmentation algorithm,performance evaluation was performed on the INbreast data set,and the experimental results of DSC,IOU,REC,SPC,Haus were 0.9419,0.8917,0.9417,0.9963,9.4200,respectively.Compared with the current best algorithm results,DSC increased by 0.19 percentage points and IOU increased by 2 percentage points,achieving the most superior breast mass segmentation performance.