Research On The Recognition Method Of Urine Formed Components Based On Deep Learning

In the past decade,the incidence and prevalence of kidney disease have continued to rise,which has seriously affected human health.Early diagnosis and treatment are very important for kidney disease and can minimize the damage caused by the disease.Urine formation is one of the three routine examinations that accurately reflect the physical condition of the human body,especially kidney disease.There are two methods for automated urine formation: microscopic imaging and flow cytometry,where microscopic imaging is the standard method for forming sub-inspections.The urine microscopic examination pictures are extremely complicated.The common clinically significant urine has more than a dozen forms.The traditional identification method is difficult to segment and the recognition accuracy is low,which cannot satisfy the modern development.This topic proposes a method for identifying the formation of urine based on deep learning,which eliminates the segmentation process of traditional methods and greatly improves the recognition accuracy and speed.The research content of the method mainly includes the following four points:1.Establish a database of urine images.The urine samples of the subject were taken from 384 patients in several hospitals in Shenzhen in different time periods.8565 pictures were selected from the microscopic examination pictures of urine samples,and were deposited in the database after three calibration medical experts conducted calibration and review.2.To propose a five-category identification method for the formation of urine.The five categories include red blood cells,white blood cells,epithelial cells,crystal and cast.This method extracts the network with Resnet101 as its feature,enhances the feature extraction ability,and draws on the deconvolution module and prediction module of the DSSD network.The deconvolution module can continuously improve the resolution of the feature layer to add more semantic information.The prediction module can enhance the target prediction on the feature layer.The recognition accuracy of the five major categories of the formation was 85.5%,88.6%,87.9%,86.4%,and 80.1%,which met the clinical requirements.The test results are discussed by changing the model network configuration and training aspect ratio.3.A method for identifying the formation of fine urine(18 types)is proposed.The 18 types of formation include red blood cells,white blood cells,epithelial cells,small round epithelial cells,casts(hyaline cast,cellular cast,granular cast,waxy cast,fatty cast,broad cast),crystal(calcium phosphate crystal,calcium oxalate crystal,uric acid crystal,magnesium ammonium phosphate crystal,amorphous crystal),Candida,lactobacillus,clue cells.The method model uses Resnet101 as the basic network,the Feature Pyramid Network extracts the image features,classifies the components based on the classification sub-network and the regression subnetwork,and optimizes the classification results again using the improved loss function.The average recognition accuracy of the above 18 categories reached 89.2%,and the recognition accuracy of the five categories was 88.9%,87.8%,88.6%,88.7%,88.2%,and 91.3%.The test results are discussed by changing the model base network,the weight initialization method,and the loss function initialization parameters.4.The design of the software system for the identification of urine.This platform can be divided into six modules according to functions: communication module,image acquisition and processing module,total scheduling timing module,data module,information export module(LIS upload,print)and view interface module(control skin loading,language switching).This software system integrates the subject's urine forming identification model algorithm and can be connected to automated equipment for urine image acquisition.The average processing time of each image(1536*1024)CPU(Intel i5 quad-core CPU)is 5.54 seconds,and the average processing time of GPU(Nvidia 1060 GPU)is 0.21 seconds,which lays the foundation for the application of the algorithm.