Renal Ultrasound Image Segmentation Based On Deep Learnin

Kidney is one of the most important organs in the human body,which can ensure the stability of the internal environment of the human body and enable the metabolism to proceed normally.In recent years,ultrasonography has been widely used in the diagnosis of chronic kidney diseases.Studies have shown that in renal ultrasound images,the crosssectional areas of the kidney,renal parenchyma and renal sinus,the length of the long axis and the short axis of the kidney,and the echo values of each area are closely related to the health of the kidney.In addition,the thickness of kidney and renal parenchyma is also a common indicator for the diagnosis of chronic renal failure.Therefore,the automatic segmentation of kidney,renal parenchyma and renal sinus by computer can provide premise support for further measurement of kidney-related indicators,and has important medical research value.However,the shape of kidney,renal parenchyma and renal sinus in renal ultrasound images are complex and diverse.There are many artifacts and noises in the images.In addition,in many ultrasound images,the edge of the segment to be segmented is fused with the background,which leads to the fact that the existing methods cannot segment the kidney,renal parenchyma and renal sinus well,and the edge and interior of the part to be segmented are inaccurate.At the same time,there are few existing datasets of kidney,renal parenchyma and renal sinus,which also brings great challenges to the experiment.Therefore,this thesis creates three datasets and proposes two deep learning methods suitable for the segmentation of renal ultrasound images for these three-part segmentation challenges.The specific work is as follows:(1)A method of renal ultrasound image segmentation based on global-local UNet11(GL-UNet11)is proposed.Aiming at the problem of inaccurate interior and edge segmentation in the automatic segmentation of kidney,renal parenchyma and renal sinus in ultrasound images: Firstly,we deepen the UNet network to improve the expression ability of the network;secondly,a new channel attention module—global-local(GL)module is proposed,which comprehensively considers the influence of global channels and local channels on predicting the importance of each channel so that the network can better focus on important information;thirdly,the proposed global-local network is added to the convolutional blocks in the subsampled part of the deepened UNet,which effectively enhances the features of important channels;finally,this method uses Dice＿BCEloss that is more suitable for kidney ultrasound image segmentation as the loss function.The experimental results show that the segmentation effect of this model on these three parts is significantly better than other several deep learning methods in the four indicators of Dice coefficient,Io U,specificity and sensitivity.(2)A method of renal sinus ultrasound image segmentation based on multi-scale UNet11(M-UNet11)is proposed.Since renal sinus has no obvious characteristics and is a small area,the segmentation accuracy in the previous method is lower than that of kidney and renal parenchyma.Therefore,further research on the segmentation of the renal sinus is based on the GL-UNet11 method: Firstly,the dataset is center-cropped to allow the network to better focus on the renal sinus and shorten the training time;secondly,the atrous spatial pyramid pooling is added at the bottom of the GL-UNet11 model to fully integrate the semantic information and detail information;finally,the traditional convolution in the upsampling part of the model is replaced with the depthwise separable convolution,which greatly reduces the number of parameters of the model and improves the segmentation performance of the algorithm.The experimental results show that the segmentation accuracy of the renal sinus is greatly improved in the four indicators of Dice coefficient,Io U,specificity and sensitivity.