| The classification and recognition of red blood cell（RBC）morphology is of great significance in evaluating the quality of stored RBCs and diagnosing blood-related diseases.At present,manual microscopy based on blood smear is the most basic method of RBC morphology classification and recognition.Since the optical absorption coefficient of RBC is small,the imaging contrast of traditional two-dimensional bright field image is poor,and there is not enough detail to distinguish different RBCs.In clinical trials,chemical staining operations are generally used to enhance the imaging contrast of RBCs.However,staining artifacts can blur the cell boundaries in fragile images,which makes the separation of individual cells between adherent cells particularly difficult.In addition,the detection of cell morphology in blood smears is accomplished through a process of repeated staining by professionals.This process is cumbersome,costly,and slow,depending on the skills and experiences of the examiner.Therefore,the development of a complete system will be of great significance to the development of RBC morphology classification and recognition.In addition,integrating quantitative phase imaging and deep learning technology is more conducive to simplify the operational complexity of inspectors,thus operation errors are reduced.In view of the above problems and corresponding needs,the influence of different optical imaging techniques on the segmentation and extraction of unstained RBCs has been studied in this thesis.Moreover,the cell classification algorithms and identification systems have been discussed.In this study,we found that this quantitative phase imaging method based on differential phase contrast could not only avoid the occurrence of imaging artifacts caused by chemical staining,but also improved the microscopic imaging contrast between cells and background.To solve the difficult problem of traditional optical experiments,we converted the bright-field images into the quantitative phase images with the help of deep learning（cycle-consistent adversarial network）,providing another phase recovery pathway for subsequent quantitative phase image analysis.Since echinocytes have the diverse and irregular morphologies,this characteristic leads to the difficulty of insufficient extraction of morphological features during the morphological analysis of RBCs.In our experimental,the boundary curvature could indeed increase the accuracy of the RBC classification,which improve the insufficiency of morphological characteristics.Finally,we applied the stacked sparse autoencoder to the training and learning of RBC morphology.Through the research of network structure and parameters,we realized RBC classification and recognition with high precision and high speed.Only a short processing time was required under the premise of a classification accuracy of 93.7%.Meanwhile,the misclassification rate was less than 10%,and the cell recognition time was about 20 cells/s.In addition,this study can be used for real-time recognition of RBC morphology,which can effectively improve image contrast while reducing costs.That is expected to provide new ideas for real-time high-resolution cell recognition.Based on quantitative phase imaging we had designed and implemented a complete set of RBC recognition and classification system,which provided all kinds of operation,including realize image filtering,image noise reduction,image enhancement,cell segmentation,feature extraction,cell recognition and classification of phase images.Finally,each part of the system was tested,and the accuracy and reliability of the system were verified through the marked red blood cell phase image.This system had the advantages of low cost,simple operation and high precision,and could perform fully automated image processing operations.This would greatly improve the RBC detection accuracy and the operating efficiency of inspectors. |
PDF Full Text Request