Research On Cell Counting Based On Density Estimation And Its Applications In Microscopic Image

With the rapid development of China’s economy and the continuous improvement of the medical system,life science research has been more and more in-depth.As the basic unit of all life activities,the spatial distribution and the number of cells can greatly reflect the individual’s health,which has important research value in life science and medical diagnosis,and so on.Traditional cell counting task is implemented by manual work,which requires high concentration for operators.The work intensity of operators is increased with the cell image reading,which may decrease the efficiency and accuracy,so it is very necessary to automate the cell counting task.In recent years,supervised learning has been made great progress in the image and video field,and many learning-based cell counting models that can infer how many cells a cell image contains have been put forward.Compared to the cell counting models based on traditional image processing,the learning-based models improve accuracy a lot.Density estimation can not only provide the cell number in an arbitrary region of interest,but also show the spatial distribution of cells.It is currently the mainstream counting method in object counting,including cell counting.Based on counting cells by density estimation,this thesis respectively analyses the limitation of random forest,attention mechanism,and model complexity,and proposes three cell counting models.The problems in applications such as cell overlap,uneven staining,background noise,out-of-focus blurring,cell deformation,and various sizes are solved.Based on the proposed models,this thesis designs and implements an interactive microscopic image cell counting software.Firstly,this thesis proposes a cell counting method based on the random forest and the density map,and it consists of two stages:training data preparation and cell detection framework In the first stage,a probability map that can reflect the cell distribution is defined.It alleviates the data imbalance and overcomes the statistical limitation of the random forest output,which boosts the model’s fitting capability.Meanwhile,the cell centers are highlighted to make cells detected better in the next stage.In the second stage,a cell detection framework based on multiple random forests is presented.Leveraging the properties of the defined probability map,it detects cell centers by Hessian matrix and reduces the influence of noise.Merging the various detection results generated by different random forests and converting them into a density map,which improves the accuracy and robustness further.Compared to the machine learning methods directly estimating density values,the proposed counting method estimates density value by detection,which greatly eases the model fitting and breaks through the limitation of random forest.The combination of multiple random forests effectively improves the counting performance.Secondly,this thesis proposes a cell counting method based on the attention mechanism and it takes the feature pyramid network as the backbone.While extracting multi-scale features,the proposed network filters features in various dimensions and scales by embedding channel attention blocks and spatial attention blocks and merges features with various scales and receptive fields,which enhance the feature representation.The pyramid structure solves the scale variations.The attention blocks highlight the helpful feature channels in channel dimension and increase the attention on cell regions in spatial dimension to suppress the background noise.Merging local information and global information in different scales makes the network more robust to multi-scale cells.Experiment results indicate that the proposed network achieves good counting results on different cell datasets and the embedding of two types of attention blocks efficiently helps improve the counting accuracy.Thirdly,this thesis proposes a low-complexity bilateral cell counting network based on a real-time semantic segmentation network.The proposed method respectively extracts detail informantion and semantic information by two branches.With the action of the detail guide block,the feature fusion block,and the feature refinement block,the proposed network increases the counting performance while keeping low complexity.The detail guide block filters the low-level features by high-level semantic information,which strengthens the cell location information.The feature fusion block merges detail features and semantic features at different scales,which bridges the detail path and context path adequately.The feature refinement block reuses low-level detail information repeatedly to refine the coarse-grained semantic features,and it improves the quality of the high-resolution density map.Training on the samples with different sizes,the experiment results indicate that the proposed model achieves low count error and avoids over-fitting when the training sample decreases.It overcomes the usual practical case that the cell images for training are often scarce.Finally,an interactive microscopic image cell counting software for the Windows platform is designed and developed,it integrates the three proposed cell counting models and makes the cell counting task automatic.Finally,different types of cell images are sampled and trained by the software to proves its praticality.