Improved Cell Detection Model Based On Deep Learning

Cell recognition is an important field of medical image analysis.In blood cell analysis,viral infection,leukemia and other diseases can be preliminarily screened by checking the type of white blood cells and calculate the ratio of various cells.In blood routine examination,the quantity and quality of peripheral blood cell components such as red blood cells,white blood cells and platelets can be checked to accurately judge the development degree of the disease and the effect of drug treatment,which provides important help for doctors’ diagnosis and treatment.The current clinical cell counting and abnormal cell detection are still dominated by artificial form,which is not only time costly but also susceptible to the experience and status of staff.With the development of convolutional neural network in computer vision,target detection technology based on deep learning is gradually applied to medical image analysis.A cell detection model based on YOLOv5-5.0 is proposed in this thesis,the main research contents and innovations are as follows:1.The idea of anchor frame is referred in this thesis,by taking the center point as the center of the circle and the maximum width and height of the anchor frame as the diameter to constructs a circular anchor frame。Based on this,a new IoU calculation method(ZIoU)is proposed to improve bounding box loss function.It is proved theoretically that the new bounding box loss function can not only retain the advantages of the original calculation,but also accelerate the convergence rate of the model.After modifying the bounding box loss function,the overall mAP is increased by 1.6%,and the convergence rate of the loss function is significantly accelerated compared with the original loss function.2.A new interaction ratio calculation method(ZIoU)is proposed based on the circular anchor frame.Change the Euclidean distance to IoU distance based on ZIoU when K-means clustering is performed on Anchor.The training results show that the overall accuracy rate is increased by 0.7%after the improvement of K-means clustering method.3.The attention mechanism is introduced into the model.After comprehensive comparison between SE and CBAM,the CBAM module is selected to be embedded in the backbone network extraction layer and combine with the original network layer to form new structural blocks,so as to extract the channel weights of different feature maps,make up for the deficiency of convolutional neural network in the extraction of overall information and improve the detection accuracy of small targets.Experimental results show that after adding CBAM module,the overall accuracy of the model is increased by 1.8%,and the overall mAP is increased by 1.4%.