Research And Implementation Of Deep Neural Network Classification Method Of Urinary Sediment Cell

Urinalysis is a very important medical detection mode,detection of the formed elements in the urine has extremely important value for clinical diagnosis and judging curative effect.The automatic classification of urinary sediment cell image is the emphasis and difficulty in the present study,so enhancing the accuracy of automatic classification urinary sediment cell image has exceedingly important significance for the development of the detection of the formed elements in the urine.This thesis summarizes the different ways and theoretical basis for detection of urinary sediment cells,proposes a classification method based on the deep neural network classification method,and develops a urine sediment cell classification platform based on the MFC framework.The main tasks are as follows:(1)For the problem of tiny differences among cell types,the characteristics of urinary sediment cell images were analyzed,and a classification method based on pre-classification of cell cross-sectional area for dual-model mixed recognition was designed.The urinary sediment cell image has a small difference in the shape of some cell types but a large difference in volume,this kind of difference was indicated in cell cross-sectional area.The cell image is divided into two groups to be trained separately,and the dimensions are predetermined before classification of the recognized image.Then the large-cell CNN model and the small-cell CNN model were used for classification,respectively,and the accuracy rate was improved by 2.3%.(2)Aimed at the problem that the image input size in the Alexnet network classification process does not match the actual size of the cell and affects the classification accuracy,an improved Alexnet network classification method was designed.This method solves the problem that excessive scaling of the image will generate a large amount of redundant information,and the existing network operation efficiency cannot meet the requirements.The accuracy rate is improved by 1%,and the recall rate of various types is increased by 3%.Meanwhile,the amount of calculation is reduced,so the operating efficiency increased by 9%.(3)For the problem that the difference between abnormal red blood cells and abnormal white blood cells is small,a classification method based on the fusion of gray features and FFT circularity features is designed.The grayscale feature is only for the morphological characteristics of the image and cannot distinguish between abnormal red blood cells and abnormal white blood cells.However,in the high frequency spectrum of FFT,there is a great difference in the circularity of the two,which can distinguish the two very well.The original image and FFT-transformed high-spectrum image were divided into two models for training,and the learned features were merged.After improvement,the accuracy of abnormal red blood cells and abnormal white blood cells is increased by 1.63% and 2.5% respectively.(4)A urine sediment cell image classification software based on MFC framework was designed and implemented.In the Visual Studio 2013 environment,using the caffe framework and CUDA language,we developed the image preprocessing module,the image classification module,the result display module,the data transmission and the file processing module,and achieved rapid classification of large-scale urine sediment cell images.Compared to the CPU version,the speedup ratio is 180 times.