| White blood cells(WBC)are of great significance to the defense function of the human body.When a disease occurs in the human body,it is often accompanied by changes in the number and shape of WBC in the blood.By detecting and identifying the five types of WBC in peripheral blood cells,it is helpful to quickly determine the type of disease of the patient,take effective treatment measures.At present,the detection and identification of WBC are mainly carried out in clinical practice by means of manual microscopy and blood cell analyzers.Manual microscopy inevitably results in errors due to the inspectors’ subjective judgments,differences in experience,and fatigue detection.The blood cell analyzer has shortcomings such as high research and development costs,inability to obtain images,inability to perform morphological evaluation,and manual re-examination when abnormal detection occurs.The rapid development of deep learning technology,computer hardware technology and imaging technology provides a new direction for the detection and identification of WBC.This paper discusses the detection and identification method of WBC based on convolutional neural network,divides the detection and identification of WBC into two steps,and conducts relevant experimental verification.(1)For the task of WBC detection,a selective attention center network(SA-Center Net)is proposed based on Center Net to detect WBC.The network includes two selection attention modules,which are independently calculated in the two selection attention modules and then fused.Each selection attention module is divided into two channels for channel attention calculation and spatial attention calculation respectively.The two attention channels are fused additively.Then,according to the characteristic of "near-circle" appearance of WBC,a circular regression method is proposed to reduce background interference,and a distributed loss function based on circular regression is extracted to train SA-Center Net.The experimental results show that SA-Center Net is comparable to the existing mainstream algorithms in detection accuracy,with a m AP of 96.4%,which is 2.3% higher than the original Center Net,and its detection speed is much higher than the existing mainstream detection algorithms.It is an effective WBC detection method.(2)For the task of WBC recognition,a superimposed network(SUPNet)is proposed based on Dense Net121 to recognize WBC,SUPNet mainly consists of four Superimposed block(SUP Block)and three transition layers.The SUP Block mainly improves the information representation power of the feature map by superimposing the feature map.The transition layer mainly connects the two SUP Block by adjusting the channel and size of the feature map,and finally recognizes it through the fully connected layer.Aiming at the problem of unbalanced number of five types of WBC,the class-balanced cross-entropy is proposed as the loss function to train SUPNet.The experimental results show that SUPNet leads the existing mainstream algorithms in recognition accuracy,reaching 99.36%,which is1.64% higher than the original Dense Net121.It has a medium speed in recognition speed and is an effective WBC recognition method.(3)A visual platform for SUP Block detection and recognition was built.Finally,the detection task and the recognition task are combined together for a comprehensive experiment.The experimental results show that the method in this paper is ahead of other mainstream detection algorithms in accuracy,lower than the general single-stage detection network in speed,but still better than the general two-stage detection network.Based on the convolutional neural network,this paper realizes the detection of WBC and the recognition of five categories through two steps,and builds the corresponding hardware platform and interactive software,which greatly improves the speed and accuracy of WBC detection and recognition,which is of great significance to the development of clinical blood routine testing and smart medical care. |
PDF Full Text Request