Research On Detection Of Masses In Mammography Based On Convolutional Neural Network

With the progress of society and the continuous development of the economy,people's living standards are constantly improving.The fast-paced,high-stress working environment also brings hidden dangers to people's health.Among women,the incidence of breast cancer is increasing year by year,which causes serious threat to women's physical and mental health.However,early screening can effectively detect early lesions,so early diagnosis and early treatment can effectively reduce the incidence of breast cancer.Mammography X-ray images are widely used in early screening of breast diseases as a high-resolution,well-defined breast imaging method.Clinically,as a sign of breast cancer,the doctor will judge the degree of breast lesions according to the condition of the mass.Due to the large amount of reading by the doctor and the complexity of the structure of the mammary gland,it is easy to miss the diagnosis and misdiagnosis due to fatigue.In the case,the computer aided detection system for the breast mass as the "second pair of eyes" can assist the physician in making decisions and improve the detection efficiency of the lesion.This article combines the process of radiologists to detect lesions and the development of techniques in machine learning,artificial intelligence,etc.,to detect suspicious masses in breast X-ray images.The main work results are summarized as follows:Firstly,we introduce a two-stage method for detecting breast masses,which is different from previous methods that only detect masses based on mammography with masses.After different preprocessing for database of DDSM and INbreast,mammography with masses is screened by deep multi-instance learning method.Then,a region-based convolutional neural network is used to detect benign masses and malignant masses in the images.This two-stage approach is more aligned with radiologist's workflow and improves the utility of computer-aided detection algorithms for breast masses.Secondly,for low accuracy of classification based on convolution neural network,we apply a deep multi-example learning method for rough detection of breast mass.Dividing mammography into instances,the net of Inception-v3 is used to extract the features,and then the classification probability of each instance is obtained based on the linear logic function and the sigmoid function.At last,the probability of the instance with the highest probability will be see as the probability of the mammography that contains masses.Considering mass is sparse and the positive instanse is centered in the region of masses,a sparse low rank loss function is designed to adjust the parameters in the work,which improves the accuracy of coarse detection of mammogram further.Finally,considering speciality of database of DDSM and INbreast,Fast RCNN and SSD are used to detect masses in database of DDSM and INbreast respectively.In the method of Fast RCNN,we optimize selective search by generating initial regions by the super-pixel method of linear spectral clustering algorithm,which is better than the method based on graph in shape retention and contour compactness and improves the quality of generated bounding boxes and the efficiency of the algorithm after the multi-level merge algorithm,which further improves the performance of the breast mass detection algorithm.Experiments shows that the two-stage computer-aided detection algorithm can accurately screen out mammography containing masses and detect benign masses and malignant masses in the image,which improves the detection accuracy and reduced false positive rate.