Segmentation Of Calcifications In Mammograms Based On Deep Learning

Among female cancer patients,breast cancer is the most common cancer and one of the important factors that threaten women’s health.Studies have shown that if breast cancer can be detected early,it can effectively increase the cure rate,and the five-year survival rate will also be greatly improved.Therefore,the early screening of breast cancer has extremely important significance for the follow-up treatment of patients.Medical imaging is an important method for early breast cancer screening.Among them,mammography is the most commonly used screening method because it has the advantages of clear imaging,low cost and high sensitivity.However,doctors usually need to rely on a large amount of clinical experience to make a diagnosis.Especially for calcifications,because of its small size,it also brings great difficulties to the diagnosis of doctors.In recent years,in order to help doctors make more accurate diagnoses,various computer-aided diagnosis algorithms have been continuously proposed.Especially with the improvement of deep learning,computer-aided diagnosis has shown great research prospects.Therefore,in this study,a calcification segmentation method based on deep learning is proposed.In this method,the segmentation of calcifications is achieved in two stages.In the first stage,the convolutional neural network classification model is used to achieve the screening of suspected calcification areas,and the second stage is to achieve accurate segmentation of calcifications through the U-Net,and through a multi-scale image feature fusion model to reduce the number of false positives.The main contributions of this thesis can be summarized as follows.(1)Screening of suspicious calcification areasIn the first stage,a sliding window method was used to intercept patches from the whole mammograms.And then,a multi-scale Dense Net based on attention was constructed to determine whether these patches contain calcifications.In this thesis,in order to effectively aggregate the features of different scales,a new multi-scale convolution module is constructed and integrated into the Dense Net.This module uses the residual structure to increase the range of the receptive field layer by layer.At the same time,this thesis proposes a new channel space attention module and integrates it into the multi-scale convolution module.Finally,the model achieved good classification results in five-fold cross-validation experiment.(2)Segmentation of calcificationsIn the second stage,this thesis uses an improved U-Net model to segment calcifications in suspicious areas.In this segmentation model,firstly,the decoder can aggregate features of all scales through full-scale connections.Secondly,for the encoder network of the model,the classification model proposed in the first stage is used,and the parameters are initialized by transferring the weights learned in the first stage.Finally,in order to reduce the number of false positives generated during the segmentation process,a multi-scale feature fusion model is used to restore the non-calcifications.Finally,a new test set was used to test the overall performance of the method,and the values of Dice,accuracy and positive prediction rate were 85.06%,85.37%,and 87.55%,respectively.