Research On Classification Algorithm Of Bone Marrow Smear Cells Based On Neural Network

With the continuous development of image processing algorithms and medical digital imaging technology,neural network-based analysis of medical digital image has become the current research mainstream.Bone marrow cell morphology is an important branch of laboratory medicine,which is mainly based on the classification of cell morphology in bone marrow smears.This work needs to be completed by experts with at least three to five years of work experience.The diagnosis for one patient requires the participation of multiple experts,and it takes at least three working days for a bone marrow smear from staining to issuing a test report,which is time-consuming and labor-intensive.Automated examination of bone marrow cytology can effectively relieve the pressure on medical professionals and reduce diagnostic costs.In this study,we propose a cell recognition algorithm that combines neural networks with traditional image processing techniques to address the issue of uneven sample size of bone marrow cells and similar morphological characteristics of nucleated red blood cells at different growth stages.It can the recognition of bone marrow cells and the classification of nucleated red blood cells at different growth stages:1.The improved YOLOV7 network is used to classify bone marrow cells.This study utilizes an image dataset of bone marrow cells annotated by professional doctors as the object,and considers nucleated red blood cells with different growth stages with a small sample size as one type.A variety of target detection networks are verified,in which YOLOV7 is selected as the basic network for improvement.The improvement measures include:adding attention mechanism in the backbone network,and improving the loss function of bounding box,which achieve a better performance.2.The method combining neural networks and traditional image processing algorithms is used to classify the different growth stages of nucleated red blood cells.Firstly,the U-Net network is utilized to segment the cell bodies and nucleus of nucleated red blood cells.To effectively remove the interference of high pigmentation of mature red blood cells when their hemoglobin content is high,contour search algorithms are applied to extract the maximum area contour and fill it as a cell body mask.Secondly,traditional image processing algorithms are used to detect the roundness of the nucleus and identify the nucleated red blood cells during division.Afterward,the image feature of radiomics in the region of interest of rest nucleated red blood cells is extracted,and a multi-layer perceptron is constructed to filter the feature values of different combinations.The results indicate that the first order statistical features(FOSF)can achieve an accuracy of 94.37%for the classification of nucleated red blood cells at different growth stages.