Research On Mass Classification And Segmentation Of Breast Mammogram Based On Deep Learning

Breast cancer,as a common female disease,has affected thousands of women around the world.Studies have shown that carrying out early breast screening can effectively reduce breast cancer mortality.As a common breast screening method,mammogram has been proven to be the simplest and most effective early screening method for breast cancer,which is widely applied in all over the world.However,in clinical practice,doctors mostly read the pictures manually,and since the level of the doctors are not uniform and human judgments are too subjective,misdiagnosis and missed diagnosis are inevitable.However,the computer-assisted diagnosis of breast cancer can be a good reference for doctors,which helps doctors make faster and more accurate judgments from an objective perspective and brings good news to patients.In mammograms,mass is a common pathological feature,which indicates a cancer.Because masses have various shapes and are very similar to surrounding glands,the classification and segmentation of mass is a major difficulty.Since the classification of mass and the contour of mass is significant,the automatic classification and segmentation of breast mass is of great research value.This article mainly takes breast mass in mammograms as the research object.According to the characteristics of breast mass and the deep learning algorithms,this article aims at the classification and segmentation of breast mass in mammogram and the exploration is carried out in the following two aspects:1.Mass classification of breast mammograms based on deep learning:For the purpose of achieving the classification of benign and malignant mass automatically,the network model Res Net50 is applied.And a new classification model was used,in which CBAM block is combined with Res Net50 network model,aiming to improve the ability of the model for discriminating the feature of mass.According to the characteristics of the mass,a training method of secondary transfer learning is designed.A local mass patch dataset is sampled from the mammogram firstly,and the patch dataset are subjected to the first transfer learning on Image Net,and then global breast region dataset performs a second transfer learning on the local patch dataset.According to the data,the proposed model achieves an AUC value of 0.8607 on the local mass patch dataset,and the proposed model trained by the secondary transfer learning method achieves an AUC value of 0.8081 on the global breast region dataset,which demonstrates that the proposed method is efficacious.2.Mass segmentation of breast mammograms based on deep learning:Automated segmentation of masses in mammograms is a challenging problem.Aiming at the problem that some masses are similar to the surrounding glandular tissues and the edges of the masses are blurred and the contrast is not strong,a multi-attention mechanism is combined with an improved Unet network to create a mass segmentation network.In the encoder and decoder blocks of the Unet,a variety of unit structures are designed,and the combination of different structures is compared.By comparison the residual unit is used as the encoder and the basic unit is used as decoder to improved Unet.To further improve the segmentation performance,in the encoder block,the dual attention module from DANet is applied,which makes the encoder pay more attention to the important feature parts when extracting features,so that the edge information obtained during restoration is more accurate.And by applying the squeeze and excitation module from SENet,the decoder module pays more attention to the restoration of details when restoring the original features,which improves the restoration accuracy of the decoder.Experiments show that on the breast mass ROI dataset,the designed network structure obtains a Dice value of 0.9033 and an accuracy of 0.9722.The experimental data objectively verify the effectiveness of the proposed method.