Breast Mass Detection Based On Deep Learning

From the incidence of female malignant tumors and related statistical data,the highest proportion of female malignant tumors is breast cancer.The core of prevention and treatment of breast cancer lies in early detection and treatment.Mammography is one of the most effective breast cancer screening techniques because of its advantages such as fast acquisition,high cost-effectiveness and little harm to the body.It projects the breast with oblique and axial position as the breakthrough point,so as to obtain the mammography target image,and then doctors observe whether there is any sign of canceration according to the image.However,radiologists mainly rely on their own professional quality and long-term practical experience in the process of image discrimination,which is subjective to a certain extent.Moreover,due to the long-term work of doctors may be distracted,stressed and tired,it is easy to cause missed diagnosis and misdiagnosis.In order to alleviate the shortage of medical resources,the researchers developed a computer-aided diagnosis system(CAD)to help radiologists read mammography images.CAD can automatically analyze mammography images and mark suspected lesions through some algorithm.The development of CAD system based on the traditional algorithm depends on the features extracted manually,which is not good in the complex situation.With the emergence of deep learning method,the CAD system using deep learning training strategy can automatically learn the features that are helpful for diagnosis from the training set,so as to improve the accuracy of diagnosis in an all-round way.In this paper,the breast image mass detection model based on deep learning algorithm can effectively avoid the defects of one sidedness and subjectivity in the process of traditional algorithm relying on prior knowledge and feature extraction,so the intelligent and objective of this full-automatic technology is more prominent.The main work of this paper is as follows:Firstly,the training set is preprocessed,and the transfer learning training strategy from natural image to mammography image is used to prevent over fitting;secondly,in order to explore different convolutional neural networks(CNNs),the training set is preprocessed,We use Res Net-101 and VGG16 as the backbone network to construct Mask R-CNN,and make a comparative experiment After getting the preliminary results of Mask R-CNN network,if there are two kinds of prediction frames for benign and malignant tumors in the prediction of images,a voting decision system composed of three convolutional neural networks(Res Net-152,Inception-v4,Dense Net-201)is used to determine whether the images are classified into benign and malignant by using the images containing masses in INbreast dataset In addition,Mask R-CNN has a certain segmentation ability.Through its unique mask branch,it can segment the target in the prediction frame,making the predicted contour closer to the label of professional doctors,but the segmentation is only mask As a sideline of Mask R-CNN,the accuracy is not high.In this paper,a Mask IOU branch is added to enhance the segmentation performance of Mask R-CNN.Finally,we use AUC to evaluate the classification performance of the auxiliary discriminant system,evaluate the model detection performance based on FROC,and use Di to evaluate the segmentation ability of the model.The results show that the performance of the improved Mask R-CNN breast tumor detection system based on improved Mask R-CNN plus voting decision system is excellent.The best result of the model is that when the number of false positives per graph is 0.09,the true positive rate reaches79.6%,the di value reaches 0.8677,and the AUC of the auxiliary discrimination system reaches 0.8895.