Research On Semantic Segmentation Of Breast Ultrasound Images Based On Attention Mechanism And Deep Convolutional Network

Breast cancer is a type of cancer which is harmful to women.In clinical practice,there is an urgent need for an accurate,efficient and automatic breast segmentation method to detect breast lesion areas with ultrasound,in order to reduce the burden on doctors and reduce the cases of misdiagnosis and missed diagnosis.This article takes safe and highly sensitive ultrasound images as the research object.And uses the deep learning techniques,which is becoming more and more used in the field of medical imaging,to study the semantic segmentation of breast tumor lesions to realize the auxiliary diagnosis of breast cancer.The automatic segmentation method of breast tumors can effectively assist doctors in diagnosing the condition,carrying out surgery and treatment plans,and has great research significance.During the research process,it was discovered that the accuracy of breast lesion region segmentation is limited by the quality of ultrasound images and the number of typical samples.In the past,classic deep learning models has problems of serious feature information loss and poor feature fusion effects for breast tumor segmentation tasks.At the same time,convolution calculation has the problem of non-global receptive fields and inability to obtain long-range context information,which affects edge segmentation.Finally,the model network parameters are large,training and testing time is expensive,and there is a risk of overfitting.In response to the above problems,this thesis has carried out the following four aspects of research work:1.Aiming at the problem of poor ultrasound image quality and small data set,we use image preprocessing and data enhancement methods to improve image quality and expand the number of images,effectively avoiding the impact of poor image quality on segmentation accuracy and relatively small data sets,small data set brings risks such as overfitting.2.Aiming at the problem of large loss of feature information in Unet network and poor feature fusion effect,the DsUnet network,which can effectively reduce the loss of feature information and improve the effect of feature fusion is proposed.Experimental results show that this method can effectively improve the semantic segmentation of breast tumors.In view of the problem that the local receptive field of convolution operation is difficult to obtain global correlation information,thisartic article designed the attention mechanism module based on space and channel,constructed the attention matrix by obtaining the relevance of other pixels and channels to the current pixel and channel,and obtained the globally related context information.On the one hand,it can effectively reduce the difficulty of edge segmentation,and on the other hand,it can effectively improve the segmentation accuracy.3.Aiming at the problem that local receptive field of convolution operation is difficult to obtain global related information,this paper designs an attention mechanism module based on space and channel,and constructs an attention matrix to obtain the context information of the global correlation by analyzing the correlation between the remaining pixels and channels with the current pixel and channel.These mechanisms effectively reduce the difficulty of edge segmentation and improve the accuracy of segmentation.4.Aiming at the problem of the huge number of model parameters and long training and testing time,this paper improves the DsUnet network structure and designs a lightweight network structure names tiny-DsUnet,which reduces the number of model parameters and time usage of training and testing by half within acceptable accuracy loss,effectively avoiding the risk of overfitting.