Two-stage Detector For Automatic Classification And Recognition Of Leukocyte

Identification of white blood cell types and morphology is an essential basis for diagnosing leukaemia.On the one hand,manual identification method has complicated process and high cost.On the other hand,it is limited by expert experience and the quality of blood smear with low accuracy in white blood cell detection.Deep learning algorithms,especially convolutional neural networks,are becoming more and more mature.It is of far-reaching significance to apply them in the automatic detection process of medical images.This article designs a two-stage object detector for the automatic classification and detection of white blood cells in microscope images.Our research methods are as follows:(1)The global feature layer is gained by using residual network,and the two-stage detector is connected in a feature sharing manner.In the first-level detector,the sliding window module is introduced to train the convolutional neural network based on the global feature layer to extract features of different scales for primary target detection.The region suggestion network(RPN)is used to generate a large number of rough region suggestions.(2)The rough suggestion box and the global feature layer are combined into the second detector to generate the local feature layer and further extract the pixel-level fine features.The boundary regression network is continuously trained.The prediction box is fine-tuned,using the non-maximum suppression to filter the target prediction box that is closest to the actual value after decoding.(3)A classifier network is available to recognise the type of leukocyte in the prediction box.Multi-task loss is used to evaluate the final detection loss value,including classification loss and regression loss,and realize the auto-detection of white blood cells.This thesis proposes an automatic classification and location technology for white blood cells with high precision.This technology combines multi-scale target screening and fine-tuning to learn the distinguishing characteristics of different leukocyte types thoroughly.It overcomes the cumbersome and costly defect of artificial detection in white blood cells.The proposed method achieved 97.12% mean of average precision(m AP)on 11580 clinical images of six types of leukocyte.Among them,the average precision(AP)of eosinophils(EO)is 99.37%,the AP of basophilic granulocyte(BA)is 100%,the AP of band neutrophil(BNE)is 93.69%,the AP of segmented neutrophils(SNE)is 90.92%,the AP of monocyte(MO)is 99.09%,the AP of lymphocyte(LY)is99.64%.This thesis introduces a high-accuracy algorithm of leukocytes in microscopic images.It will be helpful for hospitals to automatically classify white blood cells utilizing computer-aided blood image analysis.The method provides technical support for automatic recognition of microscope blood images and also opens up a new computer path for early detection of leukaemia.