Breast Ultrasound Lesions Classification And Breast Mass Segmentation Based On Deep Learning

Breast cancer has become one of the most common cancers in women,and its incidence ranks first among female malignant tumors.For the treatment of breast can-cer,early diagnosis is crucial.Computer aided diagnosis system can provide doctors with effective diagnostic advice to reduce the possibility of missing and false detection,which provides great convenience for the diagnosis process.Based on the deep learn-ing,this paper studies two key technologies in the computer-aided diagnosis system of breast ultrasound images,including:1.Breast ultrasound lesions classification.Based on convolutional neural net-works and popular attention mechanisms,this paper proposes a multi-scale contextual attention network(MCA-Net)for breast ultrasound lesions classification.Based on the Res Net-50 architecture,MCA-Net includes an embedded feature ensemble module(EFE-module)and a spatial attention module(SA-module)for capturing multi-scale contextual information.EFE-module combines multi-scale dilated convolutions and channel attention mechanism,which plays a key role in enhancing network receptive field,capturing global context information and correlation features between channels.SA-module generates spatial attention map by using spatial relationships of feature map to improve the network ability for capturing location information of target regions.The proposed MCA-Net has achieved competitive results on two publicly available breast ultrasound datasets,BUSI and UDIAT,with the accuracy of 94.40%and 91.71%,re-spectively.2.Breast mass segmentation.In this paper,an axial Transformer and feature enhancement-based CNN(ATFE-Net)is proposed for breast mass segmentation by combining the traditional convolutional neural network with Transformer structure.Based on the Res Net-34 architecture,ATFE-Net includes an axial Transformer(Axial-Trans)module and a Transformer-based feature enhancement(Trans-FE)module to capture long-range dependencies.Axial-Trans module uses an axial architecture,which is achieved by calculating the feature associations between horizontal and vertical pixel elements in the feature map respectively.Compared to the original Transformer struc-ture,the computational complexity of Axial-Trans is greatly reduced from O(n~2)to O(n).Trans-FE module adopts a feature enhancement structure based on Transformer to enhance the current feature map by capturing dependencies between different feature layers.The proposed ATFE-Net has achieved better results than several state-of-the-art methods on two publicly available breast ultrasound datasets,BUSI and UDIAT,with Dice coefficients of 82.46%and 86.78%,respectively.