Urinary Sediment Image Analysis System Based On Deep Learning

As one of the three items of routine physical examination,Urinalysis,also known as "in vitro kidney biopsy",has high clinical value.Generally,there are usually two types of urine tests: dry chemical analysis and manual microscopic analysis.Methods of dry chemical analysis draw a conclusion reaction with a solution by chemical reagents.This method is inefficient and it can only do qualitative analysis.This method has high efficiency but low accuracy and can only be tested qualitatively.It is difficult for doctors to make accurate judgments on the condition,and further testing is still required.Manual microscopic analysis directly collect the patient’s urine samples to make glass slides for microscopic examination,which can accurately obtain the quantity and shape of various formed components in a microscopic field of view.Detectors can make a diagnosis by analyzing the number and morphology of formed components in the microscopic field,such as red blood cells,white blood cells,epithelial cells and casts.However,the efficiency of manual microscopy is too low to meet clinical needs.Therefore,various types of automatic urine sediment detection instruments have appeared,such as UF in Japan and IQ in the United States,and Wison in China.These companies have developed many products within a decade.This shows that it has great market demand and research value.However,such instruments are often expensive,and the installation and maintenance of instruments and equipment also have strict standards,generally requiring additional hardware equipment and regular maintenance.In addition,such instruments require a large amount of reagents to pretreat urine samples for a single test,which leads to a high cost for a single test of the instrument.Besides,such instruments also have a high rate of missed detection,which will lead doctors to make erroneous diagnoses and delay the disease.Therefore,a urine sediment detection system with low single detection cost,favorable popularization and high accuracy has important practical value.In this thesis,taking the hospital clinical urine sediment detection project as the background,combined with the low detection accuracy and high single detection cost of the current urine sediment detection system,a deep learning-based method was developed based on computer vision,digital image processing,deep learning and other related technologies.And this thesis uses this algorithm to build and deploy a urine sediment detection website.The website only needs the user to upload the urine sediment image to obtain the test result,without redundant hardware equipment.The main work of this thesis is as follows:(1)Making a Urine Sediment Image Dataset.First,label the unlabeled urine sediment images provided by the hospital with the assistance of the doctor and create a dataset of urine sediment images to train deep learning algorithms.Secondly,according to the characteristics of urine sediment images,a data enhancement method based on original data is proposed to generate new data from original images.(2)A Urinary Sediment Image Segmentation Algorithm for Complex Detection Situations is Proposed.The segmentation algorithm is based on Smoothed Dilated Convolutions and Gated Channel Transformation urine sediment image segmentation algorithm and it can segment images in complex scenes.Compared with other medical image segmentation algorithms,this algorithm replaces ordinary convolution with smooth hole convolution,expands the receptive field and can utilize multi-scale feature information,and also uses gated fusion attention mechanism to multi-level features.The algorithm is applied to the urine sediment detection system,which can improve the detection accuracy of the system.(3)Built a urine sediment image detection system.The existing urine sediment detection system is usually a urine sediment detection analyzer.Such instruments generally rely on additional hardware and are expensive and difficult to popularize widely.In this thesis,a urine sediment detection system is built.The system is a B/S architecture website system based on Python-Django and Vue.js frameworks.The system does not require additional hardware equipment,just upload the urine sediment detection image to the system through the web page,and the system can automatically identify and analyze it.