Research On Bone Marrow Cells Detection And Recognition Methods Based On Deep Learning

Leukemia is a disease caused by the suppression of normal hematopoietic function.It is caused by the uncontrolled proliferation,differentiation and apoptosis of bone marrow hematopoietic stem cells,which accumulate in bone marrow and other tissues.At present,the diagnosis of leukemia is usually performed by doctors using manual counting methods based on the morphology of bone marrow cells,which has problems such as high work intensity,low efficiency,and strong subjectivity of results.Therefore,the research on automatic detection and identification of bone marrow cells based on deep learning technology can greatly improve the diagnostic efficiency of leukemia and other blood diseases.In response to the above problems,this paper uses deep learning technology to study the detection and identification of bone marrow cells.For the bone marrow cell detection task,this paper proposes a one-stage bone marrow cell detection model SSD-LA that improves the SSD model,and integrates the CBAM attention mechanism and the Io U Loss loss function to locate and segment bone marrow cells on the bone marrow global image.For the bone marrow cell recognition task,this paper proposes a bone marrow cell fine-grained recognition and detection model BM-Trans based on the fine-grained image classification model Trans FG,which is used to classify bone marrow cell images.The experimental results show that the model proposed in this paper is superior to the existing methods in the detection and identification of bone marrow cells.It can improve the diagnostic efficiency,automatically realize the detection and classification of bone marrow cell images,and assist doctors in the diagnosis of leukemia.The specific work is as follows:(1)In this paper,the FBMI large-scale bone marrow cell dataset was constructed with 1579 full-domain images of bone marrow cells collected from the Department of Hematology,Baiqiu’en Second Hospital of Jilin University;meanwhile,the SBMC-1 dataset consisting of 14,761 single-cell images of bone marrow was constructed based on the annotation information on the full-domain images.In order to address the imbalance phenomenon of bone marrow cell data and the problem of noise interference during medical image collection,this paper augments the SBMC-1 dataset with the publicly available dataset BM＿cytomorphology and constructs the SBMC-2 dataset consisting of 23,235 bone marrow single-cell images.The SBMC-1 and SBMC-2datasets were applied to Res Net50 separately,and the experimental results showed that the data augmentation improved the accuracy by 3.5% and the precision by 1.5% on the bone marrow cell classification task.In this paper,the 25 categories of cells in the morphological examination of bone marrow cells are divided into 17 categories for labeling for the characteristics of cells in the bone marrow with levels beyond the normal range that lead to common leukemias and for medical diagnostic purposes.(2)In this paper,we propose a bone marrow cell detection model,SSD-LA,which is an improvement on the first-stage target detection model,SSD,by using Io U Loss as the loss function of the SSD-LA model and adding a CBAM hybrid attention mechanism to the multi-scale feature maps used for prediction.To evaluate the performance of SSD-LA,the proposed SSD-LA model is compared with the classical target detection models YOLO-V3 and Retina Net in this paper.The experimental results demonstrate that the SSD-LA model achieves 0.570 in the cell detection evaluation index m AP,which is 2.3% improvement compared to the SSD model,and the SSD-LA model improves 0.7% in detecting targets on the full domain images of bone marrow cells compared to the SOTA model YOLO-V3.The experimental results demonstrate the effectiveness of the attention mechanism introduced by the SSD-LA model and the Io U Loss function employed on the bone marrow cell detection task.(3)To address the highly morphologically similar characteristics of bone marrow cells,this paper proposes the BM-Trans model for fine-grained classification of bone marrow cells.The BM-Trans model is an improvement on the Trans FG model applied by Transformer on fine-grained images by proposing a feature selection module(FSM)and combining the sum of global loss,local loss and contrast loss As the model loss function,the results of ablation experiments demonstrate the effectiveness of the feature selection module as well as the loss function on the fine-grained classification task of bone marrow cells.To evaluate the performance of the BM-Trans model on the bone marrow cell recognition task,the BM-Trans model is compared with VGG16,Res Net50,Bilinear＿CNN,Swin Transformer,and Trans FG network in this paper.The experimental results demonstrate that BM-Trans achieves 79.2% accuracy and 79.1% precision on the bone marrow cell classification task,which is a 1.1% improvement in accuracy and 2.2%improvement in precision compared with the SOTA model Trans FG.In this paper,the trained BM-Trans model was tested against the Trans FG model on the public dataset ACL,and the BM-Trans model improved 0.9% in accuracy and 1.4% in precision,verifying the generalization performance of the BM-Trans model.