Research On Breast Ultrasound Image Classification And Segmentation Method Based On Deep Learning

Breast cancer is one of the leading causes of cancer death in women,early detection and treatment can effectively reduce mortality.Ultrasound imaging has been used in clinical imaging screening of early breast cancer due to its advantages of low cost,no radiation,and high real-time performance.At present,clinical ultrasonography and diagnosis of breast cancer mainly rely on manual identification,which puts forward high requirements on the professional level and clinical experience of doctors.Deep learning technology can automatically learn features from raw images,and identify and analyze disease lesions through these features.Therefore,this paper uses deep learning technology to construct classification models and segmentation models,aiming to realize the automatic detection of breast tumors to assist doctors in diagnosis and reduce the burden on doctors.The main research content of this paper is as follows:(1)An attention-based breast ultrasound image classification model AS-Xception is proposed.Firstly,the Xception network was improved by implementing skip connections across multiple depthwise separable convolutional modules to fuse shallow and deep features,thereby enabling the network to fully utilize texture information extracted at the shallow layers.At the same time,under the premise of maintaining network performance,the amount of network parameters and calculations are reduced by reducing the network width.Secondly,a spatial and channel mixed attention mechanism SCA is proposed,and it is applied to the improved Xception,so as to enhance the network’s focus on salient information in both spatial and channel domains.Experimental results demonstrate the effectiveness of the model for breast tumor classification.(2)A multi-scale attention-based breast ultrasound image segmentation model MSAR-UNet is proposed.Aiming at the problem that the tumor size in breast ultrasound images varies greatly,which is not conducive to the network fully extracting tumor features,a multi-scale attention module MSA is proposed and applied to the downsampling operation in Attention U-Net.This module enables the network to capture multi-scale features and focus on useful information during downsampling,thereby fully extracting features of tumors of different sizes.Furthermore,residual connections are added around the convolutional blocks of the network to facilitate the propagation of information and prevent model performance degradation.Experimental results show that the model can achieve good segmentation results.(3)On the basis of the above research,this paper constructs a breast ultrasound image-assisted diagnosis system,which can automatically classify and segment breast tumors by uploading breast ultrasound images.The system can assist doctors in diagnosing breast cancer,to a certain extent reducing misdiagnosis rates and improving diagnostic efficiency.The breast ultrasound image classification model and segmentation model proposed in this paper can classify and segment tumors in breast ultrasound images,and provide a method reference for breast ultrasound image to assist diagnosis.