Research On The Detection Method Of Breast Mass Lesions Based On Deep Neural Network

Breast cancer is one of the main cancers that threaten human life and health.Developing countries account for nearly half of breast cancer patients,but the number of deaths accounts for 60% of the world.The main reasons for this phenomenon are hormonal differences and effective early detection.Discovering potential lesions from patients as early as possible is of great significance for improving the survival rate of breast cancer patients.Medical imaging represented by mammography plays an important role in the early diagnosis and treatment of breast cancer.Computer Aided Diagnosis(CAD)can assist doctors in interpreting a large number of medical images more quickly,objectively and accurately,so that potential patients can receive accurate treatment as soon as possible.At present,medical image analysis based on manual selection and design features is subjective,and cannot effectively extract high-dimensional features in the image,and cannot meet the requirements of complex function model modeling.The timely emergence of deep learning has effectively solved this problem and greatly improved the accuracy of computer-aided diagnosis based on traditional medical image analysis.The pathology of breast masses is complicated.There is a strong correlation between the masses in mammograms and adjacent pixels.The mammograms of different patients have high similarities,and different mammograms have certain structural properties.Distortion,etc.,adds a lot of difficulty to the detection of the mass.In addition,many mammography images are detected on a private data set,which is difficult to reproduce,and it is difficult to measure the true performance of the model.Currently,most of the publicly published mammography image data sets are digital mammography images,and the image quality is not high,and the number of publicly disclosed global digital mammography image data sets is scarce.All of these have caused obstacles to the research of breast mass detection based on deep learning models.In response to the above problems,this article uses all public data sets,starting with the fine-tuning of Faster R-CNN on a small data set,and verifying the feasibility of improving the model detection performance through small data set fine-tuning.Select part of the data in the OMI-DB database to carry out the experiment.First,initialize the network with the weights obtained by pre-training the network with a large natural data set,and then use the smaller-scale mammography data to fine-tune and test the network.After five-fold cross-validation The average accuracy rate(ACC)of the final mass detection is 85.36%,and the area under the ROC curve(AUC)is 0.84.After that,using the YOLOv3 network,combined with transfer learning,try to transfer the features learned by the model on the digital image to the global digital image to alleviate the lack of global digital mammography images.After learning the features of the images in the DDSM data set,the network is applied to the two data sets of DDSM and INbreast for the training and testing of bump detection.The final ACC is 81.34%;finally,on the basis of the first two experimental work,Applying the YOLOv4 network,supplementing and adjusting the data set with reference to the cases detected in the previous experimental results,training and testing the network,the final ACC is 85.64%,and the AUC is 0.85.In addition,by merging the BN layer to the convolutional layer,the problem of long training time for YOLOv4 is improved.