Research On The Method Of Urine Formed Element Recognition Based On BP Neural Network

Analysis of urine formed elements helps to assess the state of the patient's kidneys and urinary system.Focuses on the characteristics of unstained and unmarked urine sediment image with various elements,blurred edges,and severe defocusing,traditional methods mostly use morphological segmentation and support vector machine to segment and identify.When the amount of data is large,the accuracy is low and the recognition time is long.Based on the characteristics of urine elements,we designed a method for classification of urine formed elements with improved BP neural network,which can efficiently and accurately realize the identification and classification of urine formed elements such as red blood cells,white blood cells,epithelial cells,casts and crystals.The main work is described as follows:1.A fusion filter image preprocessing method was designed.The algorithm first used mean filtering to remove interference caused by uneven illumination in the original urine image,and then used Gaussian filtering and nonlinear wavelet transform threshold method to eliminate additive noise and multiplicative noise generated by the imaging system,so as to improve the image quality.2.An improved BP neural network based on particle swarm optimization was proposed to segment the urine formed elements.This method used particle swarm optimization to find the optimal weights and thresholds as the initial parameters of the network for training and prediction,which avoids the problem of traditional BP neural network falling into local extremum and improves the network structure.The improved BP neural network was used to train 50 manually segmented images to determine the parameters between each node,and then the trained network model to segment the newly input images,and good results were obtained.3.A urine formed elements classifier based on genetic algorithm improved BP neural network was proposed.The classifier uses genetic algorithms to optimize the initial parameters,while using the momentum gradient descent method to accelerate the network convergence and improve the recognition speed.The experiment identified 100 unstained and unmarked urine formed microscopic examination original images,and selected 15 features as the differences between the shape,texture and statistical characteristics of various urine formed elements.The feature vector group is input to the classifier for recognition.The experimental results show that the recognition rate of the classifier for red blood cells,white blood cells,epithelial cells and casts all reaches more than 90%,and the effect of crystals recognition is better than the control experiment.In this study,a comparative experiment was designed and compared with traditional BP neural networks,SVM classifier and convolutional neural networks classifier.The experimental results show that this method achieved a better classification effect and faster speed on urine formed elements images with complex background and a lot of noise interference,and the recognition results were more stable.