Research On The Effect Of Generative Adversarial Network On Improving The Accuracy Of Leucorrhea Microscopic Image Detection

Gynecological routine examination is one of the most widely used items in gynecological examination.It is a diagnostic method for different types of vaginitis.At present,most hospitals use artificial microscopy to carry out gynecological routine examination.The method is to apply normal saline soaked with leucorrhea secretion on slide,put the slide under the microscope and observe dozens of visual fields.The examination relies on the professional knowledge and work experience of the laboratory physician.There are unavoidable problems in artificial microscopy,that is,fatigue of repetitive work and halo caused by long-term microscopic observation.These problems will reduce the efficiency of examination and the accuracy of judgement.If the doctor’s own work experience is insufficient,the accuracy of routine gynecological examination cannot be guaranteed,and it is prone to miss examinations.In order to solve this problem,the laboratory conducted research on the automation of routine gynecological examination,including instruments,scanning procedures,software interface,focusing algorithm and image detection.This paper mainly studied the image detection algorithm.Firstly,this paper used optical microscopy to collect leucorrhea microscopic images from the hospital and established image dataset.The classification of objects on images was determined by discussing with doctors,and data annotation is carried out by experts to ensure the reliability of the annotation.Then,to the collected image dataset,this paper used Convolutional Neural Network(CNN)for detection.CNN has achieved enormous breakthroughs in computer vision tasks and has been applied more and more widely in various fields.CNN can perform various tasks such as object classification,image segmentation,target detection and so on by feature extraction,and proves its superiority over traditional image algorithms with high accuracy.However,when CNN is used in the medical field,the small amount of data and the unbalanced number of samples in medical image data set make the detection algorithm difficult to achieve satisfactory results.This paper proposed a method for generating multi-objective medical images by combining Generative Adversarial Network(GAN)and Region Proposal Network(RPN),and used the generated images for data enhancement to improve CNN performance.The method was to use Faster RCNN to detect leucorrhea microscopic images,use GAN to generate images to improve the data volume and balance the number of each sample,and then train Faster R-CNN again with the generated images.The experimental results showed that the accuracy of CNN detection can be effectively improved by using GAN to generate leucorrhea microscopic images and use them to train the detection network with the true images.To the three targets that doctors pay attention to when diagnosing,if only using the original database,the AP(Average Precision)of Candida was 0.4933,the AP of trichomonas was 0.8074,the recall of white blood cell(WBC)was 0.8589,and the mAP(mean Average Precision)of all 7 classes was 0.6174.After using the image generated by GAN to train the network,the AP of Candida increased to 0.5020,the AP of trichomonas increased to 0.8081,the recall of white blood cell(WBC)increased to 0.6628,and the mAP of all 7 classes increased to 0.6628.