Research On Microscopic Image Reconition Technology Of Leucorrhea Dry Slide Based On Computer Vision

The routine detection of leucorrhea(vaginal secretion)as a kind of routine examination item in medical insurance has a great market demand and quantity of detection.Because routine wet-film microscopy can only detect single infectious vaginitis and can not completely evaluate the severity of microecological imbalance in clinical vaginitis(36 or incompletely diagnosed),there are great limitations in the diagnosis of inflammation,the choice of treatment methods and the use of drugs.As the preferred method recommended by the National Practice Regulations for Clinical Laboratory,the 100-fold oil mirror microscopic examination of Gram-stained leucorrhea dry slide(hereinafter referred to as dry slices)has the advantages of high accuracy,comprehensive inspection,evaluation of micro-ecological environment and precise treatment with corresponding micro-ecological regulators.However,due to the complexity of operation and the large amount of film reading,it is difficult to popularize in high-end hospitals because of the high medical level of examiners.Computer vision technology is one of the research hotspots in the field of in-depth learning.With the improvement of computer computing power,it can recognize a large number of image information quickly and accurately.The application of computer vision in leucorrhea dry slide microscopic images can solve the visual fatigue caused by a large number of reading,and improve the consistency of batch reading speed and comprehensive diagnosis.Through consulting data and market research,due to industry and technical barriers,the current research focus is still focused on the fast detection of gray image of wet slide.There is no report on the micro-color image recognition of Gram-dyed leucorrhea under 100-fold oil mirror at home and abroad.This thesis mainly studies the automatic recognition technology of leucorrhea dry slide under 100-fold microscope based on computer vision technology.By using computer digital image processing technology,such as image segmentation,deep learning,feature statistics and other computer vision based methods,aiming at bacteria,fungi and pathogens,this thesis designs and implements a technology which can replace manual automatic identification of dry slide under 100-fold microscope.Some achievements have been achieved.The main contents of this thesis are as follows:1.In order to carry out this research,a large number of labour and material have been consumed to collect and classify a large number of color microscopic images of standard staining.A Learning Library of mycelia,spores,spores,spores,white blood cells,bacteria,trichomonas and more than 20,000 data sets of Mycobacterium in color images of 100-fold magnifying oil mirror of Gram-stained leucorrhoea dry slide have been established for model training.2.By studying the overall characteristics of the actual collected color microscopic images,we use digital image processing technology to sample and normalize the collected color images,filter the abnormal images and segment the color images.Through RGB channel separation and threshold segmentation,negative and positive targets are divided into small negative,large negative,mixed and independent positive targets according to area and region information.3.Different kinds of recognition methods are designed for target species after segmentation,including fungus classification and recognition technology based on Cafe framework,Trichomonas classification and recognition technology based on Tamura-SVM,fungus classification and recognition technology based on feature statistics and comprehensive identification method based on deep learning and morphological features,aiming at Pathogenic Fungus Identification and dominance estimation of fungus species number statistics.In-depth research and design have been carried out and considerable results have been achieved.Combining with the theory of vaginal microecology evaluation system,theoretical analysis of the test results were analyzed theoretically,and the evaluation report of female lower genital tract microecology was generated.The actual validity test was carried out in benchmarking hospitals.The comprehensive accuracy rate was 76%.The results showed that the detection ability of female lower genital tract microecology reached the average level of general examiners(the comprehensive accuracy rate was more than 50%)and could reduce the workload of examiners by more than 90%.