Research On Blood Cell Detection Based On Deep Learning

As one of the most routine items in medical laboratory,blood routine plays a vital role in diagnosing diseases.When doctors actually check,most of them use manual microscopy and instrument counting,but this method is inefficient.With the development of image processing technology,more and more scholars study blood cell detection methods based on image processing,but the feature extraction process of such methods is cumbersome and complex,and can no longer meet the actual needs.In recent years,deep learning has gradually become a hot research topic among scholars,and Internet companies with big data are scrambling to invest a lot of resources in research and development of deep learning technology.In this thesis,an in-depth study of the blood cell detection problem based on deep learning is carried out.The main research contents are as follows:（1）Research on traditional blood cell counting methods,such as manual counting method,instrument counting method,and blood cell counting method based on image processing.The blood cell database was constructed.This involves manually taking and staining blood samples,taking microscopic images of cells and labeling them.（2）Research on blood cell detection algorithm based on deep learning.The basic theory of deep learning is studied,including the structure of convolutional neural network,training methods and related parameter settings.Apply different neural network models to blood cell detection,namely,the blood cell detection algorithm based on SSD（Single Shot Multi Box Detector）,the blood cell detection algorithm based on Reina Net,and the blood cell detection algorithm based on YOLO＿v3（You Only Look Once＿v3）.Finally,the accuracy rate of the red blood cell detection algorithm based on SSD reaches 87%,and the detection time is 3.23 s.The accuracy of white blood cell detection based on Reina Net can reach 83%,and the detection time is2.73 s.The average accuracy of red blood cell detection based on Retina Net is 93%,and the time is 2.95 s.The accuracy rate algorithm of platelet detection based on YOLO＿v3can reach 84%,and the detection time is 2.85 s.The accuracy of red blood cell detection is 93% and the time is 3.56 s,and the accuracy of white blood cell detection is 89% and the time is 2.84 s.According to the usefulness of different models in blood cell detection,the type of detection cells suitable for each model is obtained.It can be seen from the experimental results that SSD can be used to detect red blood cells,and Retina Net can be used to detect red blood cells and white blood cells.YOLO＿v3 can be used to detect three types of blood cells.（3）On the basis of（2）,a blood cell detection algorithm based on improved SSD is proposed,which improves the accuracy of white blood cell detection by 1.5%,and the accuracy of platelet detection by 19.5%.A blood cell detection algorithm based on improved YOLO＿v3 is proposed to improve the detection effect.There are three kinds of ideas for improvement,the scheme to enhance the prior box improves the accuracy of platelet detection by 2.38%.The improved network model increases the accuracy of platelet detection by 1.8%,the improved matching parameters increases the accuracy of platelet detection by 2.05%,and the accuracy of red blood cell detection by 0.39%.The other two schemes of the superposition of the three ideas also improved the accuracy of red blood cell and platelet detection respectively.Finally,the best detection model for three kinds of blood cells is obtained.That is,the red blood cell detection method based on the improved matching parameter method of YOLO＿v3 has an average accuracy of93.5% and a detection time of 2.72 s.The white blood cell detection method based on YOLO＿v3 has an average accuracy of 88.69% and a detection time of 2.85 s.The platelet detection method based on improved priori box and matching parameter method of YOLO＿v3 has an average accuracy of 87.31% and a detection time of 4.07 s.It can be seen from the experimental results that the deep learning method studied in this thesis can accurately detect blood cells in real time.Compared with the instrumental notation method and the manual counting method,the method adopted in this thesis simplifies the operation steps,saves time and the cost of experimental equipment.