Research On Neck Ultrasound Image Segmentation Based On Improved U-Net Network

The prevalence index of global persistent disease-neck disease published by WHO is increasing year by year at a terrible speed.Ultrasound image is one of the commonly used medical diagnosis methods for neck diseases at present.The review and judgment of ultrasonic image in artificial clinical treatment basically depends on the knowledge reserve and experience of doctors,which is time-consuming and laborintensive,and is difficult to meet the needs of rapid and batch clinical diagnosis.In addition,a large number of medical image images are a huge workload for doctors.Aiming at the above problems,in order to improve the segmentation accuracy of neck ultrasound images,this topic proposes an improved U-Net network structure.Firstly,the ultrasonic image of neck nodules was preprocessed.Through comparative experiments,the best denoising and enhancement methods are selected to process the ultrasonic images of neck nodules.Secondly,expand the dataset through elastic transformation.In order to solve the problems of segmentation difficulty caused by obvious morphological differences,blurry borders and low contrast of ultrasonic images of neck nodules,as well as the problem of gradient disappearance of network models,a BDR UNet network based on normalized residuals is proposed,that is,to build a residual network ResNet with BN layer and Dropout layer structures.The residual network is introduced to enhance the segmentation effect of the model,suppress the gradient disappearance,and more comprehensively obtain the context information in the image.At the same time,the Dropout layer structure is introduced and the BN layer is normalized to prevent over fitting and improve the segmentation accuracy and accuracy of the model.In order to expand the receptive field of the convolutional neural network and enhance the extraction of effective features of nodules,blood vessels,nerves,etc.in the ultrasonic image of neck nodules,this paper proposes a RSED-UNet network structure that integrates residual attention mechanism and deformable convolution.At the same time of feature extraction,it increases the depth and stability of the network,and improves the segmentation accuracy and accuracy of the network.Through experimental comparison,BDR-UNet network solves the problem of difficult segmentation of ultrasonic images of neck nodules and the disappearance of gradient of network model.RSED-UNet expands the receptive field of convolutional neural network and strengthens the extraction of effective features of nodules,blood vessels,nerves,etc.in ultrasonic images of neck nodules.In the comparison of Dice,MIoU,F1,Precision evaluation indexes and visual segmentation effects,the network structure proposed in this paper is superior to other networks,and the target segmentation accuracy of normal receptive field area and small receptive field area is significantly improved.