Research On Ultrasound Image Segmentation Based On Interactive Two Stream Network

Ultrasound imaging is one of the most widely used medical imaging techniques to visualize human tissue due to its economical,convenient,practical,and safe advantages.The automatic segmentation of the region of interest(ROI)in ultrasound images is of great significance in improving the clinical efficiency of ultrasound images and the accuracy of disease diagnosis.However,due to its characteristics,ultrasound imaging brings some challenging segmentation difficulties,such as speckle noise,low contrast between the target area and the background,and blurred boundaries.To solve these problems,this paper designs two different interactive two-stream networks for ultrasound image segmentation.The main work is as follows:1)We propose a novel two-stream network based on feature separation and complementation(FSC-Net)for ultrasound image segmentation.The main feature is that a two-stream network is introduced into ultrasound image segmentation,and an interactive two-stream network is constructed by using Convolutional Neural Networks(CNN),which effectively realize the extraction of global semantic information and local detail information respectively.Specifically,for the feature separation,FSC-Net uses two branches,namely Top-To-Bottom(T2B)and Bottom-To-Top(B2T)streams,to extract global semantic information and local detailed information respectively.For the feature complementation,FSC-Net performs the interaction between the global semantic information and the local detailed information at each stage gradually,so it can complement the boundary feature of regions of interest(ROIs)in the T2 B stream and suppress the noise in the B2 T stream timely.Besides,we design Global Position Attention(GPA)to strengthen the context relationship between the inhomogeneously distributed regions to achieve better segmentation results.2)We propose a novel two-stream network based on detail screening and compensation(DSC-Net)for ultrasound image segmentation.Compared with the previous model,its main feature is the introduction of the Transformer network,which forms an interactive two-stream network with CNN,so as to achieve accurate acquisition of the main body and accurate details.Specifically,DSC-Net utilizes a Transformer stream to obtain multi-scale detailed features,and a CNN stream to extract body features with less noise.the body feature guide multi-scale detailed features to filter out noise through the Detail Screening Module(DSM),and then the filtered detailed feature is applied to Detail Compensation Module(DCM)to supplement rich details for the CNN stream.Through such interactions,DSC-Net ensures that more noise-free details are extracted.3)We construct a new lymphatic ultrasound image dataset(called LUSI),combined with two other breast ultrasound public datasets UDIAT and BUSIS,which can be used to validate the performance of segmentation methods in ultrasound images of different organs.Extensive experiments on the proposed method show that the proposed methods have better segmentation accuracy and stability than existing methods.