Research On Cell Detection Algorithm For Dense Multi-targets

Bone marrow is an important hematopoietic organ and immune organ of the human body.Classification and statistics of bone marrow hemolytic based on cell morphology is the basis of clinical diagnosis of many major blood diseases such as leukemia,anemia and malignant histiocytic disease,and is the decisive factor in the selection of disease treatment.At present,the classification and counting of bone marrow blood cells still uses traditional manual microscopy,which has problems such as large workload,labor intensive,professional level of doctors and subjective judgments that have too much influence on the discrimination results.Therefore,the use of artificial intelligence algorithm to mine different myeloid granulocyte characteristic information from medical image data for classification and recognition,to provide a more sufficient basis for clinical medical diagnosis and treatment and scientific research,is a hot spot of medical big data research,but also a crucial link in the intelligent medical system.Existing research on blood cell classification and identification methods mainly focus on the classification and identification of peripheral blood nucleated cells,and does not involve bone marrow blood nucleated cells and adjacent growth cycle cells.Nucleated cells are denser and contain more cell types than those in peripheral blood.The morphological similarity of adjacent growth cycle cells is high,and the classification is much more difficult than that of peripheral blood cells.Therefore,in the diagnosis of acute leukemia,there is an urgent need for a classifier that targets the whole category of bone marrow cells.This thesis focuses on bone marrow granulocytes in bone marrow blood nucleated cells.The main research contents are as follows:(1)In order to solve the difficulty in identifying adjacent growth cycles of myeloid granulocytes,this paper integrated four target Detection networks based on Voting strategies and proposed VOD(Voting Objects Detection)model to accurately locate and classify six types of myeloid granulocytes.Using the bone marrow granulocyte image data collected by the cooperative hospital,the anchor-based target detection network YOLO v4 and SSD and the key point-based target detection network CornerNet and Center Net are selected to train and experiment on the bone marrow granulocyte multitarget detection data set.Combined with the analysis of experimental results,a VOD model was designed.This model integrates the advantages of each network in identifying different bone marrow granulocytes and has a better recognition accuracy.(2)In order to improve the classification accuracy of the model and reduce the complexity of the model algorithm,this paper further proposes the Center NetTBCRes Net(C-TR)cascade model,which includes two parts: the positioning module and the classification module.The positioning module selects the Center Net network for cell positioning,and obtains the frame information of the location of bone marrow granulocytes;the classification module adopts the idea of bundling block convolution,and by sharing the convolution kernel parameters,it improves the Res Net network,reduces the amount of network parameters,and improves the accuracy of network classification rate.Experiments show that the VOD model increases the average accuracy of cell recogntion from 76% to 78%,and achieves good classification and recognition performance in dense and similar bone marrow granulocyte data sets.The C-TR cascade model was finally tested with 79% accuracy on the bone marrow granulocyte data set.With the same amount of data,the test image time is reduced by 27% compared with the VOD model.The research in this article can effectively assist doctors in completing the task of classification of bone marrow granulocytes,reduce the workload of professional doctors,increase the diagnosis rate of patients,and reduce the occurrence of misjudgment by human factors.It has certain application value in the field of smart medicine.