| Cells are the basic units of living organisms,and changes in the number level of cells have important clinical significance for the diagnosis of diseases.For example,changes in the number levels of red blood cells(RBCs)and white blood cells(WBCs)are directly related to diseases such as germ infections,and changes in the number levels of circulating tumor cells are directly related to cancer identification and metastatic progression in tumor patients.When performing tests of cell count levels,it is important to classify and count cells because of the complexity of the human environment,where heterogeneous samples containing multiple cells,such as whole blood and ascites,are often collected.The bright-field images of cells contain rich information such as size,shape,and texture,and cell detection can be achieved by extracting and analyzing these features,which has the advantages of simple operation and low damage to cells compared with traditional detection methods based on fluorescent markers,or light and electrical signal measurements.However,most of the current image detection methods are static image detection,which can measure a limited number of cells.In addition,this static detection method can only capture the information of a single angle of the cell,which also limits the accuracy of the detection.Therefore,this study considers the combination of image method and microfluidic platform to improve the accuracy and test throughput of existing image detection methods from the perspective of dynamic imaging,with the following main research elements:(1)In this paper,a cell classification and counting method based on bright-field imaging with multi-frame association is proposed by combining microfluidics,target detection technology and target tracking technology.After the video images of the cells are acquired in the designed microfluidic chip,the single-frame recognition of the images is realized by YOLO-V4 target detection network to obtain the class and coordinate information of the cells.After that,the designed topology matching-based cell tracking algorithm is used to complete the cell tracking as well as the duplicate free counting.(2)Based on this bright-field imaging and multi-frame correlation detection idea,this study achieved the detection of erythrocytes,leukocytes,and circulating tumor cells in a built microscope and high-speed camera-based detection system.Based on the produced cell data set,this study obtained a mean average precision(mAP)of up to 98.76%on the target detection model,and correlated the detection data of multiple frames with the designed cell tracking algorithm to correct some of the wrong or missing classification results caused by the difference of imaging angles or accidental misidentification of the model.Taking CTC as an example,this multi-frame correlation of detection further improved the mAP of CTC to 99.40%.In addition,the detection limit experiment of this study showed that this method could easily detect as low as 10 CTCs from 105 WBCs and was not affected by the epithelial mesenchymal transition(EMT)process of tumor cells.(3)Based on this method and the detection of blood cells by this method,a portable whole blood instant detection(POCT)system was also designed and built in this study.The system achieves microscopic imaging of whole blood samples with a single lens combined with a smartphone camera instead of a conventional microscope,and quantification of blood cell count levels is accomplished by a target detection network deployed in the cloud and a cell tracking algorithm.The algorithm achieved 99.25%mAP for blood cell identification.In addition,the results of clinical experiments based on the system showed that the measured hematocrit distribution range(0.456×103～1.092×103)was basically consistent with the ratio distribution range(0.400×103～1.375×103)in the routine blood report,indicating the practical value of the POCT system.This study combined deep learning methods with microfluidic to identify and track cells in the bright field,which can enable cell counting of different kinds of cells as well as POCT detection. |
