Research On Breast Disease Classification And Detection Method Based On Deep Learning

Mammography is the first choice for breast pathological screening because of its simple,reliable and noninvasive advantages.However,due to the laborious manual screening is timeconsuming,it’s easy to lead to misdiagnosis,and the computer-aided diagnosis system can screen the mammography before doctors intervene,which can improve the efficiency of working and the quality of screening,and has positive clinical significance.Focusing on the task of mammography aided diagnosis,this paper has done some work on breast disease classification and breast mass detection,proposed a variety of training strategies such as domain adaptive pre-training and data fusion,and further improved the performance of the model by using attention mechanism.Specifically,the work and contribution of this paper are divided into the following parts:1.Classification of breast diseases based on dual-view convolution neural networkAccording to the dual view imaging characteristics of mammography,a dual view model based on convolution neural network is proposed to classify breast mass and calcification.In the process of model training,aiming at the problem of less available training data,this paper proposes two training strategies: domain adaptive pre-training and fusion of public dataset.Firstly,the local breast tissue dataset is made for the domain adaptability pre-training of the backbone network,and the dual view model is initialized by using the transfer learning method.Then,the positive samples in CBIS-DDSM are selected to expand the training data to improve the performance of the model.The experimental results show that in the classification of breast mass and calcification,the proposed method achieves area under curve(AUC)of 0.906 and0.863 respectively in the classification of breast mass and calcification,an average increase of10.7 percentage points of the AUC value.The visualization results of grad cam show that the areas of concern predicted by the model are highly coincident with the doctor’s annotation,which proves the feasibility of this method.2.Classification of breast diseases based on channel-level attention mechanismAiming at the problem of feature redundancy in the process of network forward propagation,this paper proposes a backbone network namely Dense SENet.Firstly,a small core input layer based on 3x3 convolution block is used to replace the traditional large core input layer based on 7x7 convolution block.Secondly,the channel-level attention module is constructed in the network,which to redistribute the importance of different channels.The results show that the method can further improve the performance of the model in the classification of breast masses and calcification clusters.Res SENet and Den SENet have increased AUC values of 4.4% and 2.4% respectively compared with Resnet50 and Densenet121,Dense SENet showed the best performance,with AUC values of 0.921 and 0.887,and the visualization results of model prediction also showed that this method can focus on the lesion area more accurately.3.CSABlock-based Cascade RCNN for Breast Mass Detection in MammogramAiming at the problem that the classification model can’t accurately locate the breast mass area,this paper proposes a cascade breast mass detection model,which can help doctors better locate the mass area through rectangular bounding box.In addition,this paper proposes a CSARes2 net backbone network based on attention mechanism,which establishes inter channel attention module and self-attention module.The experimental results show that the detection model based on csares2 net achieves the average precision(AP)value of 0.822,which is 4.4 and 0.9 percentage points higher than Resnet and Res2 net.Respectively,and the AP value of CSA is 0.7 and 1.6 percentage points higher than SE and CBAM,respectively.