Research And Implementation Of Cell Analysis Algorithm Based Onrandom Forest

In recent years,the widespread application of computer image processing technology in the field of biomedicine has greatly promoted the development of biomedicine.The realization of automated analysis of cytological smears is of great significance for clinical applications and academic research.In this paper,a dynamic cell analysis system was designed to achieve cell counting and cell classification.For cell counting,since high-density cell images are difficult to count by detection,the definition of density function is introduced,and a nonlinear model of density function is derived.The density counting algorithm based on random forest regression model is studied.For the case where cells with different sizes and shapes appear in the cell image,a method for identifying the cell center by extracting the local extreme values of the predicted image is proposed.The Gaussian density function is replaced by distance function.The distance function is defined as the distance between a pixel and its nearest cell center,and the value of the distance function of the cell center was doubled to make it easier to detect the cell center of the overlapping cell population.Experiments proved that the method had good robustness to cell counting with different morphology and size.For the training of the non-linear model of cell counting,we extracted a variety of local pixel features,including color features,edge features,and texture features for training.Due to the large number of original feature sets and high redundancy between features,we use MRMR and random forest feature selection algorithms to obtain feature subsets with low dimensions and good effects,reducing the complexity of the model and improving the accuracy and efficiency of the algorithm.For cell classification,a method based on the random forest algorithm is applied to achieve rough segmentation of the foreground and background of the cell image.And a watershed segmentation algorithm based on mathematical morphology is proposed to separate overlapping cells.The circularity of the area is used as a criterion and the overlapping areas are eroded layer by layer to obtain the seed points of the watershed.Thus,the separation of a single cell is achieved.Besides,the ellipse fitting method is used to obtain the optimized cell edge.For the training of cell classifiers,we extracted a number of important features describing cell objects,including morphological features,color features,optical density features,and texture features.These features are used as input to train a random forest classifier.This method can be applied to many types of cell classification problems.According to the application requirements,a dynamic cell analysis software was developed under the Windows platform,so that users can train the model through simple interaction to achieve cell counting and cell classification.Finally,we designed several experiments to verify that the cell counting algorithm proposed in this paper can effectively count high-density cell images accurately,and also has good robustness in the case of inconsistent cell shapes and sizes.It was also verified that the cell classification algorithm proposed in this paper can effectively implement cell segmentation and cell classification.