Research And Implementation Of Cell Segmentation Based On Convolutional Neural Network For Intelligent Diagnosis Of Leukemia

White blood cell(WBC)image segmentation and recognition is an important part of medical image processing.The WBCs in the blood are very important for the immune function,it can help the immune system repair the injured part,eliminate pathogens,bacteria and microorganisms.The number of different types of WBCs reflects the level of the human immune system.In clinical routine blood examinations,physicians diagnose hematopoietic diseases by observing the number,proportion and morphological changes of different types of WBCs.Therefore,preventing the abnormal increase or decrease of the number of WBCs can prevent the invasion of the bacteria and viruses from infectious diseases,and accurate segmentation of WBCs is of the utmost importance for counting and classification.The automatic identification system of WBC morphology is an automated microscopy system,that is,the segmentation and classification of WBCs are realized through image processing and pattern recognition algorithms.Due to the large number of WBC types and their diverse morphology in blood smears,WBC segmentation and classification is a challenging problem.However,the deep learning algorithms based on convolutional neural network(CNN)are adept at extracting feature information from high-dimensional data,without the need for feature extraction and reconstruction process.This paper focuses on the precise segmentation and classification of WBCs in conjunction with CNN.The blood smear images mainly include WBCs,red blood cells,and platelets.Based on deep learning algorithms,this paper introduces a method for precisely segmenting adhesive WBCs,the pixels that belong to WBCs are set to the foreground and other pixels are the background.Due to the difference in color between WBCs and red blood cells,this paper proposed a target detection method based on color space transformation to extract WBCs.That is,based on the color characteristics of WBC,the color space of blood smear images from patients with acute lymphoblastic leukemia(ALL)is transformed from RGB to HSV,filtering out red blood cells.Because the size of platelets is much smaller than WBCs,the smaller targets are filtered out,thereby retaining the WBCs.Then,for those adhesive WBCs in the extraction results,this paper proposed a deep learning algorithm with multi-class and weighted loss function to precisely segment WBCs.The cell border is set to the third class,in addition to the inner region of WBCs and the background region,so a corresponding weight map for each sample is generated based on class weight and distance transformation weight,and it is applied to the cross-entropy loss function to improve the weight of cell border,thereby enhancing the CNN to learn the features of cell border during the process of fitting training set.In the process of WBC classification,due to the lack of sample amount,this paper introduced three methods of to avoid overfitting: L2 regularization,batch normalization(BN)and dropout,improving the generalization ability of the CNN model.The precise WBC segmentation method proposed in this paper has been fully tested on the dataset ALL_IDB1,achieving an accuracy of 97.92% on test set,which is superior to other algorithms.In addition,based on the fusion of Kaggle_Dataset and BCISC dataset for WBC classification,it proves that the mechanism introduced for avoiding overfitting can improve the robustness of the model.