Automatic Classification And Recognition Of Peripheral Blood Leukocytes Based On Semi-supervised Learning

The automatic classification and recognition of peripheral blood leukocytes play an important role in the diagnosis of blood diseases,wound infections,immune diseases and the efficacy evaluation of infectious diseases.At present,the hematology analyzer uses physics,chemistry techniques for classification and counting but cannot use the cell morphology information under the microscope to provide further diagnostic assistance.The gold standard of the leukocyte classification is microscopy which is time-consuming and laborious.Therefore,it is of great significance to improve the automation level of blood smear microscopy based on computer image analysis.In recent years,many works have introduced semi-supervised learning into classification to reduce the dependence on labeled data and achieved good performance.Based on semisupervised learning and deep learning,the thesis studies the semi-supervised convolutional neural network based on the pseudo label and the semi-supervised cascaded convolutional neural network to solve the problems of lack of annotations and poor performance on the classification of some categories.Moreover,an automatic white blood cell analysis scheme based on semisupervised detection under microscopic images is designed in the thesis.The main research work is as follows:1.There is insufficient cell annotation therefore the teacher-student model and pseudolabel generation method are combined to achieve semi-supervised classification.The pseudolabel generation method combines the entropy threshold filter based on model prediction and the confidence threshold filter based on label propagation.In addition,the sampling strategy is also improved.Moreover,an attention layer is added to the feature extraction network to strengthen the model’s attention to cell region.The experimental results show that the semi-supervised convolutional neural network based on the pseudo label can improve the classification accuracy with a small amount of labeled data and much unlabeled data which reduces the dependence on labeled data.Additionally,the classification accuracy surpasses the supervision model with the same amount of labeled data.2.The small inter-class difference of some cells leads to poor classification performance.The semi-supervised cascaded convolutional neural network is studied to further improve classification accuracy.The cascade structure based on convolutional neural network is added on the teacher-student model and the cascade loss is added to the loss function to further improve accuracy in some categories.It is found that the classification accuracy is improved after experiment,where the F1 value of confusing categories such as basophils and lymph increases more.3.For the microscopic image,an automatic white blood cell identification scheme based on semi-supervised detection is proposed.Single white blood cells are located by region proposal network.The semi-supervised cascaded convolutional neural network is used to classify cells.Compared with the end-to-end detection models,the recall and the detection speed are improved.Therefore,the method can realize white blood cell automatic recognition more effectively.