Research On The Classification Of Urine Formed Components

In vitro diagnostic technology obtains diagnostic information by detecting human samples,providing medical staff with more than 70% of information clinical diagnostic.In vitro diagnosis has the characteristics of non-invasiveness,ease of operation and economy.With the development of technology,the role in vitro diagnosis played in clinical diagnosis becomes more and more important.Urinalysis is of great significance in the diagnosis and prognosis of human urinary system diseases.Urine formed component analysis is a part of urinalysis,which mainly provides quantitative indicators for the analysis and judgment of diseases.At present,most of the analyzers used in the market to identify the formed components of urine use machine learning and simple neural networks,most of them require artificial auxiliary judgments.The recognition accuracy rate is also insufficient.Based on the deep learning method,this paper studies the detection and recognition of urine formed elements.First,according to the guidance of professionals and the "Practical Urine Formed Components Illustration",urine formed classification dataset is established by use the urine formed component pictures under the clinical microscope.The specific operations mainly include image selection,labeling,data amplification and data set division.There are 11 types of formed elements,including red blood cells,acanthocytes red blood cells,minor red blood cells,shadow red blood cells,other abnormal red blood cells,white blood cells,leukocyte mass,calcium oxalate crystals,uric acid crystals,phosphate crystals and other crystals.The urine formed classification data set contains 20842 formed urine images,which are randomly divided into training set,validation set and test set at 8:1:1.Regarding the detection and classification of urine formed elements as a target detection problem,the two-stage target detection representative network Faster RCNN and the one-stage target detection representative network YOLO series network are used to solve the problem.Two regional feature aggregation methods are adopted in the Faster RCNN network,namely ROI pooling and ROI align.When the feature extraction network is Resnet50,the highest mean average precision of the Faster RCNN model is 90.55%.When the feature extraction network is Resnet101,the highest mean average precision of the Faster RCNN model is 94.64%.The YOLO v3 network and YOLO v4 are selected to classify the formed elements,whose effects are compared on the urine formed classification data set.The experimental results show that the YOLO v4 neural network model has the best overall classification effect on the urine formed elements data set,whose mean average precision is 97.07% and which has a high recognition accuracy and strong generalization.The mean average precision of the YOLO v3 neural network on the urine formed classification dataset is 94.72%.In the paper,a deep learning-based target detection method is used to carry out the research on the automatic classification of urine formed elements.The results show that mean average precision of the YOLO v4 neural network model on urine formed classification dataset is97.07%,which has the possibility of practical application and promotion.