Research On Automatic Detection And Recognition Technology Of Bone Marrow Cells

Bone marrow is one of the important organs of the human body,from which human blood cells come.Bone marrow lesions can lead to abnormal hematopoietic function,leading to related blood diseases.Morphological examination of bone marrow cells is one of the most commonly used basic methods for clinical diagnosis of blood diseases.It generally includes taking materials,smears and staining,and special inspectors observe the cell morphology in bone marrow smears and count the ratio between various cells under the light microscope.For example,leukemia,related pathological cells have obvious pathological characteristics in morphology.Through the morphological examination of bone marrow cells,effective and rapid diagnosis can be preliminarily realized.However,the traditional manual inspection method is not only time-consuming and labor-consuming,but also has factors such as personal professional level and subjective differences.Therefore,using artificial intelligence technology to study and realize objective,efficient and automatic automatic detection and recognition technology of bone marrow cells is of great significance for digital auxiliary diagnosis of hematopoietic system diseases.At present,there are the following problems in realizing the automatic detection and recognition technology of bone marrow cells: 1)there are many kinds of bone marrow cells with complex and close characteristics,which increases the difficulty of recognition.2)Affected by uncontrollable factors during data acquisition,the color style of bone marrow cell images collected in different batches is inconsistent,which may affect the recognition effect of this technology.Based on the above contents,this thesis realizes the automatic detection and recognition technology of bone marrow cells based on deep learning technology.The specific work is as follows:1.This thesis constructs the bone marrow cell medical image data set bmsec.The image data is provided by the Hematology Department of the second hospital of Jilin University,and experienced experts are responsible for identifying relevant bone marrow cells.After the steps of collection,sorting,analysis,screening and labeling,the construction of the data set is completed.The data set bmsec consists of two sub data sets bmsec-1 and bmsec-2.There are 1166 bone marrow smears in the data set bmsec-1,with resolutions of 2592 * 1944,2448 * 2048 and 1024 * 768.The relevant information of bone marrow cell image is saved by XML file.There are 10538 single bone marrow cell images in data set bmsec-2,which are clipped from the bone marrow smear microscopic image in data set bmsec-1.The bone marrow cell image data set bmsec constructed in this thesis has sufficient quantity and a wide variety,which lays a good foundation for the subsequent automatic detection and recognition technology of bone marrow cells.2.According to the distribution characteristics of bone marrow cells and the complexity of morphological characteristics,this thesis carried out the application research of one-stage target detection algorithm SSD（single shot multibox detector）on bone marrow cell detection and recognition.Compared with the peripheral blood cell image,the image information of bone marrow cell image is more complex.For example,the cells may overlap and squeeze each other,and the morphological characteristics of cells in adjacent growth stages are close,which increases the difficulty of detection and recognition.This thesis uses the one-stage target detection algorithm SSD based on multi frame detection to complete the task.At the same time,vgg16 and mobilenetv2 are selected as feature extraction networks respectively to study the impact of different feature extraction networks on SSD algorithm.The results show that the SSD algorithm with vgg16 as the feature extractor performs better in performance.In addition,on the basis of this work,attention mechanism se module and ECA module are introduced into each effective feature layer of vgg16 to study the impact of the two attention mechanisms on bone marrow cell detection and recognition technology.The results show that Se module has a negative impact on the task,while ECA module can significantly improve the effect.3.Aiming at the complex morphological characteristics of bone marrow cells,especially the high similarity of bone marrow cells in the adjacent growth stages of the same category,on the basis of completing the task of automatic detection and recognition of bone marrow cells,this thesis proposes a bone marrow cell classification algorithm doublee,which uses the feature learning ability of different network architectures to fully learn the potential features of bone marrow cells,and effectively improves the accuracy of classification tasks.In addition,the algorithm is applied to the open blood cell data set rabbit and the data set mam to prove that the algorithm double e proposed in this thesis performs better in the accuracy of the result index than the ordinary deep learning model.4.After completing the research on the automatic detection and recognition technology of bone marrow cells,the color style in the data set bmsec is not unified,which may be affected by objective factors such as bone marrow samples,dye concentration and dyeing time.This thesis carried out the research on the influence of unified color style on the classification task of bone marrow cells.This thesis decided to use the software Photoshop function to match the color,select four pictures with representative color style as the color template,complete the color style unification of the data set bmsec-1,build four color style unified data sets bmsec-2,and explore the impact of the unified color style on the task of bone marrow cell classification.The results show that the unified color style has a certain impact on the classification task.Only one unified color style can improve the classification effect,while the other unified color styles will reduce the recognition effect.