Assisted Diagnosis Algorithm Of Breast Cancer Based On Attention Mechanism

Breast cancer is the most common malignant tumor among women worldwide.It has the highest incidence and the second fatal rate.Breast cancer screening for female groups and treatment after early detection of disease can reduce breast cancer mortality.Mammography X-rays are the primary method for early screening of breast cancer due to their high sensitivity to microcalcifications and small masses.However,even if a professional X-rays doctor is performing mammography,he is often faced with problems such as large subjective diagnosis,high false positive rate,and neglect of small lesions.With the rapid development of deep learning technology,computer assisted diagnosis(CAD)based on deep learning has become a research focus.However,there are three common problems in existing deep learning-based CAD algorithms: First,the CAD algorithm based on the detection algorithm model has high precision and good interpretability,but it is necessary for a professional doctor to mark lesions in images.This process is very expensive and cannot be adapted to the rapidly increasing amount of data.Second,CAD models based on full-image classification only require image-level labels fortraining,and training data is easy to obtain,but the accuracy is low and the diagnostic results have poor interpretability.Third,most of the current models are based on single image training and cannot effectively utilize the information of multiple views in the mammography.In response to the above problems,I will conduct research from the following three aspects:In this paper,I suppose the main reason for the low accuracy of breast cancer classification is that the lesions are very small and very vague in mammography.The output of general classification model depends on the whole image,and it is not effective to learn the features of local lesions..In view of this problem,this paper proposes to apply the attention mechanism to the mammography classification of malignant cancer.So that reduce or enhance the influence different features to the classification results.The attention mechanism is applied to the feature maps of different scales in the network multiple times,which effectively improves the classification accuracy.And this paper proposes that the classification results of different scale feature maps are used to calculate the loss during the training process,then added to the total loss with different weights.When detection,only the head of network's feature map is used for classification.This method is beneficial to the learning of the underlying features.On the other hand,it solves the over-fitting problem and can improve the accuracy to some extent.Finally,the experimental results show that the proposed method can significantly improve the accuracy compared with the original classification diagnosis.And through the class activation mapverification,the attention mechanism proposed in this paper can better locate the lesion..In the process of breast cancer screening,usually take two image of different views for each breast.So there are 4 images in each case.It is foreseeable that effective use of different views during screening can bring about an improvement in accuracy.Inspired on this idea,this paper proposes a multi-view information fusion method based on attention mechanism.First proposed that there is a obviously spatial correspondence between the left and right images,second the lesions on the different views of one breast are same.Based on these relations,the left and right views are taken as input to the spatial information fusion branch to obtain the spatial attention weight,and the different perspective images of one breast are taken as input to the feature information fusion branch to obtain the attention weight in the channel dimension.Then,the two attention weights are multiplied to obtain the weight with same dimension as the input feature map.Multiply the weight with input feature map for changing the influence of each feature point for classification after.Then analyzing the defects of existing feature fusion methods,and compared with the method proposed in this paper.The experimental results show that the feature fusion method proposed in this paper achieves the highest accuracy.Finally,the activation region is displayed on the input image through the class activation map.Compared with the original classification algorithm,the feature fusion method proposed in this paper pays more attention to the lesion region and cansuppress the influence of the background region on the result to achieve more accurate prediction results.At last,for improving the diagnostic efficiency and diagnostic accuracy during screening,build an assisted diagnostic system based on the deep learning detection algorithm.For the detection algorithm,YOLO v3 is used in this paper.The one-step detection algorithm improves the diagnostic efficiency and reduces the equipment requirements,making the system convenient to different devices.Then for the problem of small quantity of lesions,the improvement of the loss function greatly reduces the impact of imbalance between different classes.The system is built using the Flask framework,which is also based on Python,so it is easy to embed the trained YOLO v3 detection model into the system.The breast cancer assisted diagnosis system can directly and effectively improve breast cancer screening efficiency and accuracy.