Research On Automatic Leukocyte Classification And Detection Method Combined With Transformer And CNN

White blood cells are the guards of the body and an important part of the body’s defense system.Their classification plays an important role in assisting doctors in diagnosing major diseases.At present,in clinical practice in hospitals,the classification usually relies on experienced doctors to manually count with the hematology analyzer.This method is time-consuming,labor-intensive,and errorprone.It results in physicians misdiagnosing or overlooking diagnoses,which endangers patients’ lives.Therefore,in order to further assist physicians in diagnosing diseases,it is of great significance to realize the automatic classification of white blood cells.In recent years,with the continuous development of deep learning,many scholars have applied deep learning methods to the automatic classification and detection of white blood cells and achieved certain results.Based on the deep learning method,this thesis studies the white blood cell classification model combined with Transformer and CNN to solve the problems that the global features of leukocyte images cannot be learned,slight difference between the different leukocyte categories,and the lack of features in leukocyte images in previous studies.In the application of the white blood cell image classification model,it is usually necessary to cut out the white blood cell from the microscopic image.Therefore,in order to further realize the automatic detection of blood microscopic images,this thesis studies a white blood cell detection model which combined Transformer and CNN.The main work of this thesis is as follows:(1)For the automatic classification of white blood cells.This thesis proposes a white blood cell fine-grained classification model(WBC-GLAformer)based on global-local attention.Aiming at the problem of being unable to extract global features of leukocyte images and lack of features in leukocyte images in previous studies,a low-level feature extractor and a global-local attention encoder module are constructed to fuse local and global features of leukocyte images,enrich the features of leukocytes,and improve the classification accuracy of the model.In addition,for the problem of slight differences in different leukocyte categories,a discrimination region select module is constructed to select the regions with higher discrimination in leukocytes,which is beneficial for the classification of the model.(2)For the automatic detection of white blood cells.This thesis proposes a white blood cell detection model(MFDS-DETR)based on multi-level feature fusion and deformable self-attention.Aiming at the scale differences for different leukocytes,a high-level select-feature pyramid network(HS-FPN)is designed based on the characteristics of white blood cells to achieve multi-level fusion,which is using the channel attention module to filter the information of low-level features by using highlevel features as weights,and then fusing the filtered low-level features with highlevel features to improve the feature expression ability of the model.For the problem of fewer features in white blood cells,encoders are used to extract the global features of the feature map.Furthermore,in order to reduce the complexity of the model and improve the detection effect,a multi-level deformable self-attention mechanism is used in the encoder to extract features.Finally,decoders are used to convert the output feature map into the position and category information of the target box to achieve white blood cell detection.(3)Comparing the WBC-GLAformer classification model proposed in this thesis with existing classification models,it has the relatively high accuracy compared to the typical convolution model in the bone marrow leukocyte fine-grained classification dataset and improves the accuracy by 2.03% compared to the Transformer model(Vi T).In order to verify the effectiveness of the MFDS-DETR proposed in this thesis,the proposed model is compared with existing detection models on the white blood cell object detection dataset,and its effectiveness is relatively well.Compared with the Transformer model(Deformable DETR),its AP is increased by 4.8%.In addition,in order to verify the generalization of the two models,a comparison is also made on the public dataset,and their effect is relatively well.Further,the importance of key components in the two models is verified by ablation experiments.