Research On Image Recognition Method Of Urine Formed Elements Based On Semi-supervised Learning

Powered by deep learning technology and hardware facilities,relevant automated medical testing equipment plays a more and more significant role in clinical practice,which alleviates the shortage of medical personnel in China to a certain extent.At the same time,there are a large number of accurate and fast-automated medical testing equipment needed to fill the gap under China’s huge population base.Urine microscopic examination is a common clinical examination subject,which can judge the health status of the examinee by analyzing the category and quantity of urine formed elements.Due to the complexity of the actual urine detection environment,the accuracy and robustness of the urine automatic analyzer designed based on dry chemistry method,staining method and traditional image processing technology are not high enough,the urine automatic analyzer has not been fully popularized,and the manual microscopic examination method is still the mainstream of urine routine examination.Image recognition algorithm is a vital part of urine automatic analyzer.Aiming at the problems of high cost to label data,easy confusion of classification and model redundancy in the existing algorithms for the recognition of urine formed elements,this paper makes the following research combined with deep learning technology and the characteristics of low-resolution urine formed image:(1)The difficulty of labeling data is particularly common in the medical field.Aiming at the problem that the urine formed element image is difficult to label or the label is inaccurate,this paper proposes a recognition scheme based on semi-supervised learning.Based on a large number of image data,this paper designs a re-parameterization network,and applies unsupervised contrastive learning for feature learning,which only needs a small number of labeled images can complete the training of classification model.At the same time,for some training sets,corresponding image preprocessing methods are proposed to improve the image quality.(2)Aiming at the problems of long tail distribution of urine formed element images collected in reality,some categories are easy to be confused,multiple forms of categories and the image quality under the microscope is poor.This paper proposes a dual attention mechanism of channel and space and an improved classification loss function.The attention mechanism can make the model pay attention to the key points in the current image,and strengthen the distinction between easily confused elements.The improved loss function and related strategies effectively avoid the impact of data imbalance on classification tasks.(3)In view of the large resolution distribution range of urine formed element images,using the same resolution for recognition affects the performance of the model.In this paper,a double branch network is designed to recognize urine formed element image,multi-scale feature extraction is carried out for feature maps with different resolutions,improve the classification performance of the model,and the model is compressed by knowledge distillation.The experimental results show that the method in this paper has achieved an average classification accuracy of more than 94% for 16 kinds of urine formed element images,and the average recognition time per image is 21 ms.And the method in this paper has better classification performance compared with the urine formed element images classification methods in recent years.