Research On Urine Sediment Images Recognition Based On Deep Learning

With the continuous improvement of people’s living standards,people pay more and more attention to health problems.As a developing country with a large population,China has uneven distribution of medical resources,and the burden of doctors is increasing year by year.Especially in the laboratory,the automatic and intelligent medical inspection equipment has a broad market prospect.Urine sediment detection,also known as urine microscopy,is an important part of in vitro medical detection.By analyzing the composition and frequency of urine sediment,we can judge the disease of the examinee correspondingly.Due to traditional image processing algorithm and machine learning algorithm of the automatic detector accuracy and robustness is not high enough,can not adapt to the complex urine detection environment.In clinical operation,manual examination is usually used as the main method and automatic examination as the auxiliary method for diagnosis,with low efficiency.In view of this situation,this paper applies the advanced learning technology which has made a breakthrough in recent years to the detection of urine sediment image.In view of some characteristics of urine sediment image,as well as the problems that the traditional convolution neural network will encounter when dealing with the classification of urine sediment image,such as confused in recognition process,greatly affected by light and impurities,this paper has done the following research:(1)In view of the characteristics of some categories of urine sediment often confused in the recognition process,as well as the poor image quality,low resolution and less carrying characteristics of the urine sediment under the microscope,this paper proposes a scheme of combination network.Through the cascade combination of multiple convolutional neural networks,the confused categories will be into the secondary recognition process targeted,and the results of the first recognition are corrected,at the same time,it is more It makes full use of the features in the image.To solve the problem of data imbalance of urine sediment,the methods of transfer learning and data augmentation are used.(2)Because the convolution neural network is not sensitive to the area features of sediment objects in the image,it can not be effectively classified by the area features.For many categories of urine sediment,the feature of area is one of the important features that distinguish them from other categories.Therefore,this paper proposes an area feature extraction algorithm,which can approximate the area features in cells,and through the corresponding designed probability conversion function It can reduce the influence of convolution neural network on the attenuation of area feature by weighting the output probability of convolution neural network with the output probability of convolution neural network(3)In view of the disadvantage of convolutional neural network which is greatly influenced by light,this paper proposes a method based on subtracting image mean to optimize.At the same time,in order to reduce the space occupied by the model,so as to improve the loading speed of the model and reduce the difficulty of model fitting,in this paper,combined with the advantages and disadvantages of Alexnet and VGG models and the characteristics of urine sediment image,Urinet is designed to identify urine sediment image.In terms of ROC curve,macro average,micro average,global accuracy rate,recall rate,recognition time and other indicators,the combination network method and its series of optimization algorithms proposed in this paper show good recognition performance through test analysis and experiment comparison.