A Research On Automatic Cell Counting Method In Fluorescence Microimaging Based On Deep Learning

In biology and medicine,cell counting is used to determine the number of cells in a sample of a given volume.In other words,it is a measure of the concentration of cells in the sample.Cell counting is a very important process in medical diagnosis and treatment as well as biological research,which is used by scientists and medical practitioners all over the world.Clinicians and hospitals use cell counting to diagnose a patient’s health,such as the full blood count(FBC),which involves counting cells in a patient’s blood samples.This test will determine the concentration of these cells via the number of certain cell types,and cell concentration information can reflect the patient’s physical condition,and cell counting also plays an important role in finding connections between brain regions.The correlation between one brain area and the other of the injection site can be explained by the number of fluorescent cells obtained by injection of fluorescent agent.In a word,cell counting can reflect human health information and the correlation among brain regions,but manual cell counting is time-consuming and labor-consuming,an automatic method is needed to help human beings to count cells.To facilitate researches based on cell counting,this paper proposes a model framework of automatic cell counting,which can count cells in varieties of situations.The framework of the model is divided into two stages: recall stage and sorting stage.The main purpose of the recall stage is to filter out the background image to remove the interference of the background noise,and retain the real cells,and meanwhile,update the tag of cells.In this stage,the split network is used,and the directional feature module extracted by RNN is added to the local feature extraction network to improve the classification performance.The sorting stage is a more precise cell count of the recalled output.In this stage,the total number of cells in the field of view is obtained by fitting the cell density map.The model consists of encoding,decoding and pixel-level fusion stages.In the encoding and decoding stage,in order to better obtain semantic and spatial detail information,semantic enhancement module and feature selection module for cascading information are introduced to better maintain semantic and detail information.In the final density map fusion stage,we designed a pixel-level feature map fusion network,and learn a fusion weight for each pixel via the attention mechanism to get the final fusion density map.Through our proposed model framework,the number of cells can be predicted quickly and accurately,which brings great convenience for researchers and the application of cell counting.