Research On Breast Cancer Ultrasound Image Segmentation Method Based On Mask R-CNN And Attention Mechanism

Breast cancer is one of the most prevalent cancer-like diseases in the female population today.As the second leading cause of cancer deaths in women,breast cancer is threatening women’s health.How to efficiently screen and detect breast cancer is a current research hotspot.Ultrasound imaging is the preferred modality for breast imaging because of its high accuracy,high real-time performance and low cost.Since the boundaries of breast cancer foci are not clear,it is difficult to ensure the efficient reading of ultrasound images,so it is especially important to use computers to assist in the diagnosis.The problems of breast cancer ultrasound image segmentation at the present stage are:(1)the collective volume of existing breast cancer ultrasound image data is small,which makes it difficult to train a network model with strong segmentation ability;(2)the size and shape of breast cancer ultrasound image lesion regions vary,and the boundary between the target region and normal tissue is blurred,so the existing network model has poor effect on its segmentation.To solve the above problems,the main research of this article is to improve the Mask R-CNN network to enhance the segmentation accuracy of breast cancer ultrasound image lesion regions.The main work is as follows:(1)Replace the Visual Geometry Group 16(VGG16)network with Residual Network 50(ResNet-50)in feature extraction to improve the accuracy of feature extraction.The segmentation of the lesion region of breast cancer ultrasound images using mask branching yields a segmentation result map with higher accuracy.(2)Adding the Convolutional Block Attention Module(CBAM)to the output of Feature Pyramid Networks(FPN)to enable the network to focus on the focal region of breast cancer ultrasound images from both channel and space adaptively,effectively improving the The network segmentation effect is effectively improved.(3)The training set consisting of breast cancer ultrasound images from the public dataset BUSI is augmented with the fused UDIAT dataset and BUSI dataset to increase the richness of the dataset.The weights are obtained by pre-training the Microsoft Common Objects in Context(COCO)dataset and migrated to the improved Mask R-CNN network,so that the network model has the ability to recognize the generalized features and can quickly learn the high-latitude features of the breast cancer ultrasound image dataset.The experimental results show that the segmentation evaluation index of the improved Mask R-CNN network is improved by data enhancement and adding attention mechanism,and the Dice coefficient is increased to 0.93,which is 12.1%higher compared with the U-Net network,and the proposed model can improve the segmentation accuracy of breast cancer ultrasound images.