Research On Cell Counting Method Based On Density Regression Estimation

In recent years,medical image processing has become increasingly important in medical diagnostic tasks.Many tasks in biology and medicine require the calculation of cell numbers through microscopic cell images to aid in the diagnosis of diseases and the selection of appropriate treatment methods in medical processes.With the development of computer vision,more and more research in medical diagnosis is being applied to imaging devices to count cells in images.However,accurately counting the number of cells in an image is a challenging task,as cell size and shape can change significantly over time,and the distribution of cells in the image can be dense,causing issues such as occlusion and adhesion.Accurately calculating the number of cells in a cell image is therefore very difficult.To address these issues,many existing methods use density regression estimation to calculate the number of cells in cell images.By converting the cell image input into a predicted density map using techniques such as deep learning,the method can then perform regression counting on the predicted density map to estimate the number of cells in the image.Density regression estimation has higher accuracy than detection methods for calculating the number of cells in an image.Therefore,this thesis focuses on researching methods for calculating the number of cells in an image based on density regression estimation.Considering the factors that can cause errors in traditional density regression estimation methods,this thesis proposes two methods to solve the above problems.The main contributions and innovations of this thesis are as follows:(1)A method is proposed to count cells in an image by directly learning the mapping from predicted counting maps to annotation maps.This approach avoids the influence of the process of generating ground truth predicted density maps using Gaussian kernels.To optimize the loss function,a Bayesian algorithm is utilized,and a density contribution probability model is constructed from the annotations in the dataset.The expected count for each annotation point is calculated by summing the product of the probability of each pixel contributing to the count and the estimated density.(2)Using a loss function based on optimal transport theory,the algorithm treats cell counting as a probability distribution problem,mapping density values and binary point annotations to probability density functions,and measuring the differences between them.In this way,the algorithm reliably supervises the generation of predicted count maps,and adds an entropy regularization term to the loss function to improve the stability of the training process and make the predicted count maps closer to the annotated maps.To demonstrate the effectiveness of the proposed method,this thesis evaluated it on three publicly available cell counting benchmarks: the synthetic fluorescence microscope(VGG)dataset,the modified bone marrow(MBM)dataset,and the human subcutaneous adipose tissue(ADI)dataset.The experimental results show that the proposed method exhibits comparable or better performance than state-of-the-art methods on all three datasets.