Abdominal Organ Recognition Of Medical Ultrasound Image Based On Convolutional Neural Network

In the field of medical imaging,ultrasound imaging technology is widely used in clinical patients' physical evaluation and disease examination because of its non-invasive,non-invasive,non radiation,low cost,real-time imaging and other characteristics.In the field of medical ultrasound image,image segmentation can be used to locate human organs and calculate the volume of lesions,which will help doctors to speed up the diagnosis of patients' physical conditions.Secondly,combined with the physiological position of organs in the abdomen,the recognition of abdominal organs in ultrasound images can also provide the position and posture information of the probe,which lays the foundation for the intelligent guidance and scanning of ultrasonic probe based on ultrasonic images.This paper mainly carries out the following three aspects of work:Firstly,the annotated medical ultrasound abdominal organ images are divided into 868 training sets and 103 test sets,and the pixel statistics of gallbladder,liver,right kidney,left kidney,spleen,pancreas and six kinds of organs and background in the data are carried out to ensure that the distribution of all kinds of organ data in training set and test set is similar,and it is realized that the recognition of six kinds of organs is a difficult small target segmentation task.Through the image preprocessing of histogram equalization,the better organ display and detail representation in ultrasound image can be realized.The layers of FCN and U-Net are changed,the upper sampling mode is adjusted,and the BN layer is added.The evaluation indexes were designed to test and compare the converged network.The results show that the U-Net is 6.12%,18.05%,26.06%,5.98%,43.5% higher than FCN in pixel accuracy,mean pixel accuracy,mean intersection over union,frequency weighted intersection over union and kappa coefficient,respectively.Furthermore,the reasons why u-net performance is better than FCN are analyzed in detail by combining the visualization of middle layer feature map,prediction segmentation results and various organ confusion matrix,as well as the improvement ideas.Secondly,the graph network structure is introduced to improve the problem of pixel error classification in image segmentation.Aiming at the problem that FCN can't make full use of the spatial structure of image,this paper proposes to use the graph convolution neural network to transfer information efficiently,and builds the graph network reasoning structure that can be embedded.Through the information exchange between the nodes in the interactive space,we can help to establish the dependency relationship between the distant pixels in the image,so as to improve the performance of semantic segmentation.Compared with u-net,the convolution neural network has a significant performance improvement.The pixel accuracy of kidney and spleen is 34.2% and 21.3% respectively.Finally,the structure of encoder decoder is designed based on attention mechanism,and the segmentation algorithm of neural network based on self-attention mechanism is realized.The whole convolutional neural network consists of input module,down sampling module,bottom module,up sampling module and output module.In the bottom module and the upper sampling module,the residual connection and self-attention layer are introduced.The self-attention layer can effectively capture the reasoning information of the long-distance relationship between input sequences.Compared with u-net,it is 1.66%,5.11%,2.58%,2.96%,14.84% higher in pixel accuracy,mean pixel accuracy,mean inetersection over union,frequency weighted inetersection over union and kappa coefficient,respectively.At the same time,it solves the problem of kidney and spleen segmentation confusion.Especially in the gallbladder,kidney and other organs with small pixel proportion,it can also achieve excellent performance through the distribution of self-attention layer weight coefficient.