Image Segmentation Of Breast Masses Based On RAU-Net

Breast cancer has always been the leading cause of female cancer and has become an important factor that endangers female health.Research have shown that early detection of cancer can improve the survival rate and the quality of life of patients.In the way of breast examination,doctors can obtain effective inf ormation through X-ray,ultrasound,magnetic resonance and other imaging data.With the continuous progress of artificial intelligence technology,using deep learning technology to achieve better segmentation results of breast masses,that have the great significance for the diagnosis and treatment of breast cancer.In order to improve the segmentation accuracy of breast masses,this thesis uses MIAS dataset to study and improve the RAU-Net model for breast X-ray image mass segmentation based on residual attention and context relation.In view of the complexity of breast image information,the RAU-Net structure is improved based on the idea of fusing different dimensional features and optimizing network parameters:1.Considering the large number of parameters and long training time of RAU-Net network model,this thesis studied the attention module.By optimizing the module structure,the network parameters are reduced by about 29.54 %,and the purpose of reducing training time and improving segmentation efficiency is realized.2.This thesis studies and designs the context relation module.By calculating the information deviation between the upper feature and the target feature,the context relation feature is obtained,and the context feature connection is enhanced,which forces the model to focus on the shape of breast masses.By connecting contextual features and lower features,the se module calculates the feature weight of breast masses,strengthens the feature information of breast masses,and improves the segmentation accuracy.By using the context relation module,the aggregation of shallow features and deep features is strengthened,and the transmission error of feature map in the encoding – decoding process is reduced.In addition,this thesis also optimized the parameter design of the improved RAU-Net model and the selection of convolution layer and activation layer.The segmentation accuracy is further improved by data preprocessing and data enhancement.In order to verify the effectiveness of the improved breast mass segmentation model,a comparative experiment was carried out on MIAs breast X-ray images.The performance of the segmentation model was evaluated by cross entropy function,F1 score and ROC curve.The improved RAU-Net model proposed in this thesis is better than FCN and U-Net network in the segmentation experiment of complex breast mass images,and the F1 score can reach 88.36 %.This proves that the improved RAU-Net model proposed in this thesis can provide effective segmentation results for subsequent computer-aided diagnosis and treatment of breast cancer.