Research On Cell Counting Based On Deep Target Recognition

Cell counting is an important research field of medical image analysis.It plays an irreplaceable role in both the field of biomedical experiments and clinical medical diagnosis.By detecting the white blood cells,red blood cells and other indicators in the patient's blood or urine sediment microscopic image,the patient's disease type and development degree can be accurately judged.Currently,different types of cells are still counted and diagnosed by manual inspection in clinical practice.However,the manual inspection need a lot of labors and the counting accuracy is heavily affected by the status and experience of the staff,which hinders medical staff's judgment and affects the treatment progress.With the rapid development of deep object detection technology,researchers have also made breakthrough progress in the medical image analysis.Compared with traditional image processing methods,object detection algorithms based on deep learning have greatly improved detection speed and accuracy.A deep object detection algorithm was used to identify and count cells in microscopic images of blood and urine sediments.To verify the proposed methods,two compresion experiments were performed on two data sets.To address the problems including complex small object detection in blood cells and the slow inference speed of ordinary deep learning models,this dissertation proposes a novel YOLO-Dense network model based on the one-stage deep detection YOLO framework.First,the K-means cluster algorithm is used to obtain three different sizes anchor boxes for potential targets,and a residual module and a multi-scale module of feature pyramids are introduced into the YOLO basic network to improve object detection accuracy;secondly,residual training through layer-hopping connections to effectively solve the problems of gradient dispersion and explosion in deep networks;finally,by adding dense connection modules to the network architecture,the proposed model can effectively improve the network inference speed.To address the problems including low coverage rate of urine sediment microscopic image and relatively single cell scale,this dissertation proposes a unique Fast-CornerNet network model based on the CornerNet algorithm for object keypoint detection.First,in order to Reduce the amount of forward propagation calculations,the residual block in the hourglass structure of the CornerNet network model is reconstructed,and the 1 * 1convolution kernel and 3 * 3 convolution kernel are combined and spliced to replace the 3 *3 convolution in the residual block.Secondly,the MobileNet is incorporated into the proposed method to increase the detection speed without reducing the detection accuracy.Finally,adjust the input resolution of the original image and increase the residual module with a step size of 2 in the preprocessing stage for downsampling,and modify the Hourglass feature extraction structure as a whole to further reduce the amount of parameter calculation.By using urine sediment cell dataset,many comparison experiments are conducted with Faster R-CNN,YOLO series and CornerNet algorithm,Fast-CornerNet algorithm can effectively achieve the best counting accuracy with the fastest inference speed.