Research On Detection And Recognition Method Of Breast Lesions Based On Mammograms

Breast cancer is the most common cancer among women and the second leading cause of death for women.Mammography screening can early detect breast cancer,so that doctors can treat it as early as possible.Mammography screening mainly depends on the observation of the doctor,but the reading of the image requires higher doctors’ clinical experience,and the diagnosis result is often affected by subjective factors.Computer-aided diagnosis systems can improve the effectiveness of breast cancer screening by helping radiologists.The specificity of traditional systems is low,and radiologists have not improved their screening performance when using computer-aided systems.the latest developments in deep neural networks machine learning algorithms have greatly improved the performance of computer vision and medical imaging models.Breast masses are one of the most important signs of breast cancer,and their automatic detection and identification are very important for predicting cancer.Therefore,the main research content of this article is the mammography classification,breast mass detection and mass segmentation.This article classifies whether mammography is benign or malignant and assists doctors in screening and judgment.Due to the high resolution of mammography and the small proportion of the mass in the image,this paper proposes a method of merging patch features and global feature classification using the mutual attention mechanism.Cut image patch to train a classification network,which is used to generate patch features.Extend the patch feature extraction network,and input the entire image to extract global features.Use the mutual attention mechanism to merge global features and patch features for classification.Use global features as a guide to learn the importance of different channels of patch features.According to the distribution of patch features,learn the location of lesions or tissues that are conducive to classification.Give different spatial weights to global features.Experiments show that this method effectively captures the characteristics of lesions and better classifies mammographies.In this study,the mass lesions in the mammography are detected and classified to help the doctor confirm the location and type of the lesion.Because mammographies have multiple views and there is structural correlation between the mammary glands on both sides,this article uses the image features on both sides to detect the mass.Because the detection of commonly used smooth L1 loss calculates the distance of the transformation parameters,it cannot directly represent the gap between the detection frame and the real label,so the generalized intersection overunion loss is used as the regression loss.In this study,a region mapping and feature fusion structure is combined with a generalized intersection over union loss function for mass detection.The algorithm first uses the RPN structure of Faster R-CNN to generate mass candidate regions,and uses the region mapping method to map the candidate regions to the opposite image to obtain the corresponding position image blocks,and then merges the similarity and asymmetry information of the features on both sides to further regression and classification.The training process of fusion feature regression combined with the generalized intersection over union loss function to improve the positioning accuracy.Experiments prove the effectiveness of this method.This article studies the method of breast mass segmentation and provides help for the determination of subsequent treatment plan.In the related research of mass segmentation,the conditional random field is integrated into the fully convolutional network to make the segmentation result more structural.Compared with the obvious edges in natural scenes,the boundary between the lesion and the background is gradual.Detecting edges in semantic segmentation is helpful for segmentation of the scene,while due to the characteristics of the edge gradient,if the marked edge is not accurate enough,the wrongly marked pixels around the edge will affect the network training.Therefore,this paper designs the erosion loss,uses the morphological erosion operation to generate a mask for calculating the loss,so that the part around the edge with a high possibility of error does not participate in the loss calculation.This method improves the segmentation performance of the baseline method.The methods in this paper have obtained good results.But,the classification network cannot be trained end-to-end,which affects the feature encoding effect of the mammography.Model compression can be further studied.The false positives for mass detection can be further reduced.The mass segmentation edge is not accurate enough,and further improvement methods can be explored.