Research On White Blood Cell Detection Algorithm Based On Improved Center Net

The number and type of white blood cells in the blood can be used to help diagnose various diseases of the body,such as infectious diseases,inflammation,leukemia and tumors.White blood cell detection can effectively detect the morphological changes of various white blood cells and the abnormal increase or decrease of the number,which helps clinicians to diagnose whether the patient has a certain disease and plays a vital role in guiding doctors to take effective measures.In recent years,the use of computer-aided methods for blood smear image detection is still a research hotspot in the medical community.However,due to multi-cell adhesion,different staining techniques and imaging conditions,and interference in large-scale staining background,leukocyte detection remains a challenging task.In view of the above problems,based on the Center Net based on center point prediction,this thesis combines attention mechanism and multi-scale feature fusion to construct a white blood cell detection network based on improved Center Net.The innovation of this work is mainly divided into the following three aspects:(1)In order to achieve more accurate fine-grained detection of white blood cell categories,this thesis proposes to combine the SE(Squeezed and Excitations)attention mechanism on the non-anchor center point prediction network Center Net to weight the effective feature channels of white blood cells and improve the accuracy of white blood cell detection.(2)For high-resolution large-scale blood smear images,the proportion of background negative samples and white blood cell foreground positive samples is extremely unbalanced,and the edge information of white blood cells is blurred.This thesis designs a FEM(Feature Enhancement Module)multi-scale feature fusion enhancement module,which combines the underlying detail information with the top-level semantic information to improve the recall rate of white blood cell detection.(3)In view of the large difference in the proportion of different types of white blood cells,the difficulty of sample training between white blood cells is unbalanced.This thesis uses the GHM(Gradient Harmonizing Mechanism)loss function.Through the gradient coordination mechanism,the training weights for different samples are dynamically adjusted to balance the learning of difficult and easy samples and reduce the risk of overfitting.This thesis makes effective experiments on the proposed white blood cell detection model.Ablation experiments were performed on the SE attention mechanism,multi-scale feature fusion module,as well as the improved loss function and activation function.In addition,compared with Faster R-CNN,SSD,YOLOv3,Center Net,the model proposed in this thesis has good results,and the maximum m AP reaches 95.5 %,which has reference value for practical clinical diagnosis.