Research On Automatic Identification And Analysis Of Nodules Based On 3D Breast Ultrasound Images

At present,breast cancer has become the most common cancer in the world,which is a great threat to the health of women.Ultrasound plays an important role in early screening of breast cancer.However,the quality of ultrasound image is not high,and it is easily interfered by artifacts,vascular texture and other factors,which brings great challenges to the identification and analysis of breast nodules.In this paper,the segmentation and classification of breast nodules in ultrasound images are improved based on model integration and capsule network respectively.In terms of breast nodule segmentation,this paper analyzes the characteristics of ultrasonic image,such as insufficient resolution,difficult to extract key information and so on,and proposes a network structure based on “encoder integration”,and uses shared decoder to reduce network parameters.Experiments show that this method can significantly improve the segmentation performance of breast nodule compared with the “shared encoder”.At the same time,this paper redesigns the skip-connection to improve the flow and fusion of information in different directions and further enhance the network segmentation performance.Finally,this paper proposes a new connection mode between encoder and decoder,which can flexibly adjust the encoder without affecting the connection of decoder,,so it can be used for model pruning,so that users can make a flexible compromise between network parameters and performance.In the aspect of breast nodule classification,the full connection layer of convolutional network is removed,and capsule network is used instead.Specifically,the feature information is extracted by convolution network,and the classification is completed by capsule network.At the same time,in order to further improve the network performance,this paper extends the existing attention mechanism and proposes the capsule attention mechanism.The capsule attention mechanism can adaptively adjust the importance of different capsules and the importance of each dimension in the capsule vector.Experimental results show that compared with convolutional neural network,this method can not only reduce the network parameters,but also obtain better classification performance.